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.0005177 | The Presence, Persistence and Functional Properties of Plasmodium vivax Duffy Binding Protein II Antibodies Are Influenced by HLA Class II Allelic Variants | The human malaria parasite Plasmodium vivax infects red blood cells through a key pathway that requires interaction between Duffy binding protein II (DBPII) and its receptor on reticulocytes, the Duffy antigen/receptor for chemokines (DARC). A high proportion of P. vivax-exposed individuals fail to develop antibodies that inhibit DBPII-DARC interaction, and genetic factors that modulate this humoral immune response are poorly characterized. Here, we investigate if DBPII responsiveness could be HLA class II-linked.
A community-based open cohort study was carried out in an agricultural settlement of the Brazilian Amazon, in which 336 unrelated volunteers were genotyped for HLA class II (DRB1, DQA1 and DQB1 loci), and their DBPII immune responses were monitored over time (baseline, 6 and 12 months) by conventional serology (DBPII IgG ELISA-detected) and functional assays (inhibition of DBPII–erythrocyte binding). The results demonstrated an increased susceptibility of the DRB1*13:01 carriers to develop and sustain an anti-DBPII IgG response, while individuals with the haplotype DRB1*14:02-DQA1*05:03-DQB1*03:01 were persistent non-responders. HLA class II gene polymorphisms also influenced the functional properties of DBPII antibodies (BIAbs, binding inhibitory antibodies), with three alleles (DRB1*07:01, DQA1*02:01 and DQB1*02:02) comprising a single haplotype linked with the presence and persistence of the BIAbs response. Modelling the structural effects of the HLA-DRB1 variants revealed a number of differences in the peptide-binding groove, which is likely to lead to altered antigen binding and presentation profiles, and hence may explain the differences in subject responses.
The current study confirms the heritability of the DBPII antibody response, with genetic variation in HLA class II genes influencing both the development and persistence of IgG antibody responses. Cellular studies to increase knowledge of the binding affinities of DBPII peptides for class II molecules linked with good or poor antibody responses might lead to the development of strategies for controlling the type of helper T cells activated in response to DBPII.
| Vaccines are a crucial component of the current efforts to eliminate malaria, and much of the vaccine-related research on P. vivax has been focused on the Duffy binding protein II (DBPII), a ligand for human blood stage infection. A high proportion of individuals who are naturally exposed to P. vivax fail to develop neutralizing antibodies, but the host genetic factors modulating this immune response are poorly characterized. We investigated whether DBPII responsiveness was dependent on the variability of human leucocyte antigen (HLA) class II cell surface proteins involved in the regulation of immune responses. To obtain a reliable estimate of DBPII antibodies, we carried out a longitudinal study, collecting serum from the same individuals over a period of 12-months. The results confirmed the heritability of the DBPII immune response, with genetic variation in HLA class II genes influencing both the development and persistence of the antibody response. HLA class II genotype also influenced the ability of DBPII antibodies to block the ligand-receptor interaction in vitro. Computational approaches identified structural specificity between HLA variants, which we propose as an explanation for differences between a good or poor antibody responder. These results may have implications for vaccine development, and might lead to strategies for controlling the type of immune response activated in response to DBPII.
| Plasmodium vivax infects human reticulocytes through a major pathway that requires interaction between an apical parasite protein, the Duffy binding protein (DBP), and its cognate receptor on reticulocytes, the Duffy antigen/receptor for chemokines (DARC) [1–3]. Although most individuals lacking DARC on their red blood cells (RBCs) are naturally resistant to P. vivax [1], some infections occur in DARC-negative individuals living in vivax malaria endemic areas [4–6, 70]. So far, no alternative ligand facilitating the binding of P. vivax to reticulocytes has been identified, which makes the DBP one of the most promising P. vivax vaccine targets [8].
The importance of the interaction between DBP (region II, DBPII) and DARC to P. vivax infection has stimulated a significant number of studies on DBP antibody responses (reviewed in [8]). The available data demonstrate that naturally occurring antibodies to DBP are prevalent amongst individuals living in P. vivax endemic areas, and that these antibodies can inhibit the DBPII-DARC interaction [7, 9–12]. Even though DBPII-specific binding inhibitory antibodies (DBPII BIAbs) seem to confer a degree of protection against blood stage infection [11], the majority of people naturally exposed to P. vivax do not develop a DBPII BIAbs response [8]. In the Amazon Basin, for example, this inhibitory activity was detected in only one third of malaria-exposed subjects [8, 13]. Similarly, less than 10% of children from Papua New Guinea (PNG) with immunity to malaria had acquired high levels of DBPII BIAbs [11]. Given the significant differences in epidemiology and parasite genetics between the Amazon Basin and PNG, the fact that the DBPII BIAbs response is relatively low but also remarkably stable over time is particularly intriguing.
The reasons for the low immunogenicity of DBPII are not clear, but may be linked to a complex immune response driven by genetic diversity in both the parasite and human populations. Several studies have demonstrated the existence of variant specificity in the natural immune response against DBPII, which has been attributed to allelic diversity [12, 14]. On the host side, recent evidence suggests that host genetic polymorphisms might also affect humoral immunity against DBP [15, 16], with DARC polymorphisms thought to affect the ability of DBP antibodies to block parasite invasion [16]. In a previous study, we demonstrated that the naturally acquired BIAbs response tended to be more frequent in heterozygous individuals carrying a DARC-silent allele (FY*BES), which suggested that gene-dosage effect occurred [7]. In this context, we were interested in determining if DBPII non-responsiveness could be associated with variation in the major histocompatibility complex.
While malaria infection represents a key selection pressure for the human leukocyte antigen (HLA), and has left clear evolutionary footprints on the alleles observed in different countries [17], the association between HLA gene expression and responsiveness (or non-responsiveness) to defined malaria antigens has produced contradictory results [18–21]. Beyond the extreme genetic diversity of HLA class II, which hinders interpretation of the role of HLA on antibody responses elicited during malaria, most studies rely on antibody prevalence data collected at a single time-point in cross-sectional analysis of a population. Since malaria transmission is intermittent and seasonal in many endemic areas, it is possible that antibody levels fluctuate over time such that individuals could appear to be non-responders on some occasions and responders on others [20]. In the current study, we present data of the first ongoing population-based study of the relationship between HLA class II genes and DBPII immune response. The methodological approach included a community-based open cohort study in an agricultural settlement of the Brazilian Amazon, in which 336 unrelated volunteers were genotyped for HLA class II (DRB1, DQA1 and DQB1 loci), and their DBP immune responses were monitored over time by conventional serology (DBPII IgG ELISA-detected) and functional assays (DBPII BIAbs).
The study was carried out in the agricultural settlement of Rio Pardo (1°46’S—1°54’S, 60°22’W—60°10’W), in the Presidente Figueiredo municipality, located in the Northeast of Amazonas State in the Brazilian Amazon. The Rio Pardo settlement is located approximately 160 km from Manaus, the capital of Amazonas, along the main access to a paved road (BR-174) that connects Amazonas to Roraima (Fig 1). The settlement was officially created in 1996 by the National Institute of Colonization and Agrarian Reform (INCRA) as part of a large-scale colonization project focused on agriculture and wide-ranging human settlement in the Amazon area [22]. The mean annual temperature is 31°C with a humid climate and an average annual rainfall of 2,000 mm per year. The rainy season extends from November to May, and the dry season from June to October. According to a census conducted between September and October 2008, Rio Pardo has 701 inhabitants, most of whom live on subsistence farming and fishing along the tributaries of the Rio Pardo River. The study population was quite stable. Most residents were native to the Amazon region, and their average age of 28 years roughly corresponded to the time of malaria exposure in the Amazon area [7]. In the study area, migration rates were relatively low, as only 28 (8%) of 336 individuals moved out of the village during the follow-up period.
Based on the spleen size of the local children and parasite infection rates, the study area was classified as hypo- to mesoendemic, which is consistent with the general profile of malaria infection for well-established frontier settlements in the Amazon region [23]. The study site and malaria transmission patterns have been described in detail elsewhere [7]. Although P. vivax and P. falciparum are transmitted year round, P. vivax is responsible for about 90% of malaria cases in the region. Similar to other parts of the Brazilian Amazon area [24], a continuous decrease in the number of malaria cases has been reported in the Rio Pardo community; in 2008, the local Annual Parasitological Index (API) was 131 cases per 1000 inhabitants, while in 2009 the API was 54.6 (Health Surveillance Secretariat of the Brazilian Ministry of Health, SVS/MS). In the study area, precarious living conditions, including houses with partial walls and roofs made of tree leaves, increase human-vector contact and reduce indoor residual spraying efficacy [23]. However, while the availability of curative services is limited, a government outpost in the area provides free malaria diagnosis and treatment.
The ethical and methodological aspects of this study were approved by the Ethical Committee of Research on Human Beings from the Centro de Pesquisas René Rachou (Reports No. 007/2006, No. 07/2009, No. 12/2010 and No. 26/2013), according to the Resolution of the Brazilian Council on Health-CNS 466/12. In November of 2008, 541 of the 701 residents of the settlement (77.2%) invited to participate in the study accepted by giving written informed consent, which was also obtained from the next of kin, caregivers, or guardians on the behalf of participating minors.
A population-based open cohort study was initiated in November of 2008, with the following procedures: (i) administration of a structured questionnaire to all volunteers to obtain demographical, epidemiological, and clinical data; (ii) physical examination, including body temperature and spleen/liver size, recorded according to standard clinical protocols; (iii) venous blood collection for individuals aged five years or older (EDTA, 5 mL), or blood spotted on filter paper (finger-prick) for those aged <5 years; and (iv) examination of Giemsa thick blood smears for the presence of malaria parasites via light microscopy. The geographical location of each dwelling was recorded using a hand-held 12-channel global positioning system (GPS) (Garmin 12XL, Olathe, KS, USA) with a positional accuracy of within 15 m. At the time of initial enrollment in the study, 222 out of 541 volunteers had no familial relationships with other volunteers, and were consequently selected for HLA genotype and serological assays.
Six and twelve months after the initial survey, two similar cross-sectional surveys were carried out. In total, 336 unrelated subjects were enrolled in the study, with 222 examined in the baseline cohort, 249 examined during the 2nd survey (June, 2009), and 239 during the 3rd survey (October-November, 2009). A total of 244 (72.6%) subjects had consecutive samples taken, and 156 of these (64%) had samples taken in all cross-sectional surveys (baseline, 6 and 12 month follow-up).
Malaria infections were diagnosed by microscopy of Giemsa-stained thick blood smears, and by Real-Time PCR amplification of a species-specific segment of the multicopy 18SSU rRNA gene of human malaria parasites. The Giemsa-stained smears were evaluated by experienced microscopists, according to the malaria diagnosis guidelines of the Brazilian Ministry of Health. For Real-Time PCR, genomic DNA was extracted from either whole blood samples collected in EDTA, or from dried blood spots on filter paper using the Puregene blood core kit B (Qiagen, Minneapolis, MN, USA) or the QIAmp DNA mini kit (Qiagen), respectively, according to manufacturers’ instructions. Real-Time PCR was performed as previously described [25].
Molecular amplification of the alleles of HLA-DRB1, HLA-DQB1 and HLA-DQA1 were performed by the PCR-SSO (polymerase chain reaction, specific sequence of oligonucleotides) technique, with Luminex technology (One Lambda Inc., Canoga Park, CA, USA). Briefly, target DNA was PCR-amplified using group-specific primer sets, after the amplified product was biotinylated, which allowed later detection using R-Phycoerythrin-conjugated Streptavidin (SAPE), and hybridized with microspheres linked to specific conjugated fluorescent probes for each HLA allele group (One Lambda, Canoga Park, CA, USA). The fluorescent intensity varied based on the reaction outcome, with an expected intensity of 1000 or more for positive control probes. Reaction readings were carried out by flow cytometry using Luminex technology (One Lambda). Samples were analyzed through the HLA FUSION software (One Lambda Inc., San Diego, CA, USA).
A conventional enzyme-linked immunosorbent assay (ELISA) for total IgG antibodies to DBPII was carried out using a recombinant protein that included amino acids 243–573 (region II) of the Sal-1 DBPII variant, which is highly prevalent in the study area [23]; the recombinant protein was expressed as a 39 kDa 6xHis fusion protein, as previously described [26]. ELISA was carried out as previously described [27], with serum samples at 1:100 and DBPII at a final concentration of 3 μg/ml. The results were expressed as reactivity index (RI), calculated by dividing the reading values of the test (OD values) by the cut-off (mean reading for the unexposed group plus 3 SD, n = 30). Values of RI > 1.0 were considered positive.
COS7 cells (green monkey kidney epithelium, ATCC, Manassas, VA) were transfected with the plasmid pEGFP-DBPII, which coded for a common DBPII sequence circulating in the Amazon area [13]. Transfections were performed with lipofectamine and PLUS-reagent (Invitrogen Life Technologies, Carlsbad, CA) according to manufacturer’s protocols. Forty-eight hours post-transfection, erythrocyte-binding assays were performed as previously described [10]. Briefly, plasma samples were added at 1:40, and plates were incubated for 1 hr at 37°C in 5% CO2. Human O+ DARC+ erythrocytes in a 10% suspension were added to each well (200 μl/well), and plates were incubated for 2 h at room temperature. After incubation, unbound erythrocytes were removed by washing the wells three times with phosphate buffered saline (PBS). Binding was quantified by counting rosettes observed in 10–20 fields of view (x200). Positive rosettes were defined as adherent erythrocytes covering more than 50% of the COS cell surface. For each assay, pooled plasma samples from Rio Pardo residents characterized as non-responders by ELISA were used as a negative control (100% binding). For this purpose, only plasma that did not inhibit erythrocyte binding was pooled for use as the negative control (usually, 10 plasma samples/pool). The positive control included a pool of plasma from individuals with long-term exposure to malaria in the Amazon area. The percent inhibition was calculated as 100 x (Rc—Rt)/Rc, where Rc is the average number of rosettes in the control wells, and Rt is the average number of rosettes in the test wells. Plasma samples with more than 50% of binding inhibition were considered positive.
To predict HLA-DR/-DQ binding affinities (IC50) we used the P. vivax DBP sequence (XP_001608387.1) from the NCBI database. Each potential 15-mer sequence frame was scored using the NetMHCIIpan-3.1 server (http://www.cbs.dtu.dk/services/NetMHCIIpan-3.1/), an improved version of the tool that permits a much more accurate binding core identification [28]. Binding affinity was given as the log IC50 value in nanomolar (nM), and the defined thresholds for strong and weak binders were <1.7 nM and <2.7 nM, respectively.
Homology models of the three DRB1 variants were generated using Modeller and Macro Model (Schrodinger, New York, NY) using an ensemble of previously solved X-ray crystal structures of the HLA II beta chain (PDB IDS: 1IEB Chain D, 3LQZ Chain B and 1SEB Chain B; 76%, 67% and 90% sequence identity respectively). The alpha chain was modelled using an ensemble of available X-ray crystal structures including PDB IDs: 2Q6W and 4AEN (Chain D and A respectively, 100% sequence identity). As previously described [29, 30], the models were then minimized using the MMF94s forcefield in Sybyl-X 2.1.1 (Certara L.P., St Louis, MO), with the final structure having more than 95% of residues in the allowed region of a Ramachandran plot. The quality of the models was confirmed with Verify3D. The models of the HLA II complex of the alpha and beta chains were built using X-ray crystal structures of the complex (PDB ID: 1IEB, 3LQZ, 1SEB, 2Q6W, 4AEN) to guide protein docking [31]. Two representative DBPII antigenic peptides (H1: FHRDITFRKLYLKRKL; H3: DEKAQQRRKQWWNESK) were modelled into each HLA II complex variant using the available crystal structures of the HLA II complexes to guide docking. Binding affinities were predicted using CSM-lig [32]. Model structures were examined using Pymol.
The structural consequences of each amino acid difference between the DRB1 variants were analyzed to account for all the potential effects of the mutations [33]. The effects of the variations on the stability of DRB1 and the HLA II complex were predicted using DUET [34], an integrated computational approach that optimizes the prediction of two complementary methods (mCSM-Stability and SDM). The effect of the differences on the protein-protein binding affinity between the alpha and beta chains to form the HLA II complex were predicted using mCSM-PPI [35]. The effect of the changes on the binding affinity of the HLA II complex for a model peptide were also analysed using mCSM-PPI, as previously described [36], mCSM-lig [37], and mCSM-AB [38]. These computational approaches represent the wild-type residues structural and chemical environment of a residue as a graph-based signature in order to quantitatively determine the change upon mutation in Gibb’s Free Energy of stability or binding.
A database was created using Epidata software (http://www.epidata.dk). Linear correlation between two variables was determined by using the Spearman’s correlation coefficient. Differences in proportions were evaluated by chi-square (Χ2) test and, differences in medians were tested using either the Mann-Whitney or Kruskal–Wallis tests, with Dunn’s post hoc test, as appropriate. For allelic group comparison, differences in proportion were performed by Z-test or chi-square tests, or Fisher’s exact tests, as appropriate. Alleles frequencies for each locus (DRB1, DQA1, and DQB1) were summarized descriptively using frequencies and percentage for immunological categorical variables.
Overall associations with immunological responses and alleles from each locus of HLA class II were evaluated by comparing the allele frequencies between seronegative individuals and seropositive individuals from the baseline study using standard contingency tables. Based on the humoral immune response to DBPII from the three cross-sectional surveys, the long-term immune responses against DBPII were grouped into three categories: (i) Persistent non-responder (PNR)—absence of antibodies against DBPII in all three cross-sectional surveys; (ii) transient responder (TR)—antibodies detected in at least one cross-sectional survey; (iii) Persistent responder (PR)—individuals with detectable DBPII antibodies in all three cross-sectional studies. The association between HLA class II alleles (or haplotypes) and long-term immune response (PR or PNR groups) was analyzed by standard contingency tables (Chi-square and Fisher’s exact test, as appropriate), with two observations per subject (one for each allele). Alleles with a frequency of less than 0.01 were not included in the analysis.
Additionally, multiple logistic regression models with stepwise backward deletion were built to describe independent associations between covariates and HLA class II alleles or haplotypes and antibodies to DBPII. Covariates were selected for inclusion in the logistic models if they were associated with the outcome at the 15% level of significance in exploratory unadjusted analysis. Logistic regression models included the following covariates: age, gender, exposure to malaria (time of residence in the endemic area), self-reported malaria episodes, recent malaria infection and household location within the study area. Multivariate logistic regression was performed using Stata software version 10 (Stata Corporation, College Station, TX). Only variables associated with statistical significance at the 5% level were maintained in the final models. To avoid type II errors due to overly severe correction, statistical adjustment for multiple tests were not used [39, 40]. Type I errors were reduced by using multiple logistic regression models with stepwise backward deletion. Estimated genotype distribution between the observed and expected allelic frequencies was tested using the method described by Guo and Thompson [41] to verify Hardy-Weinberg equilibrium. Because the gametic phase was unknown, maximum-likelihood estimates of haplotype frequencies were obtained from multilocus genotype data and computed using the expectation-maximization (EM) algorithm [42]. Both procedures were performed using Arlequin software version 3.5 (http://cmpg.unibe.ch/software/arlequin35/) [43].
We evaluated DBPII antibody responses in 336 unrelated subjects with a median of age 41 years and a 1.3:1 male-female ratio (Table 1). Age was significantly associated with a subject’s time of malaria exposure in the Amazon area (r = 0.82; p<0.0001, Spearman’s correlation test). At the time of the first blood collection, the overall prevalence of malaria was 5%, with 14 out of the 17 (82%) infections caused by P. vivax and 3 (18%) by P. falciparum. No P. malariae or mixed Plasmodium infections were diagnosed by either microscopy or Real-Time PCR. The 336 participants were followed up for an average of 7.5 months (10 days to 12 months), thus representing 2,514 person-months of follow-up. Based on parasitological-confirmed cases, the incidence rates of P. vivax malaria were 1.03 episodes per 100 person-months (95% confidence interval [CI] of 0.69–1.49/100 person-months) and 0.19 per 100 person-months for P. falciparum (95% CI of 0.03–0.32/100 person-month).
One hundred and twenty-two (36%) of the individuals enrolled in the study had ELISA-detected IgG antibodies to the main variant of DBPII circulating in the study area (Sal-1) (Table 1). Because not all DBPII IgG antibodies are able to block the interaction between the ligand (DBPII) and its receptor on the RBC surface (DARC), we evaluated the functional properties of the anti-DBPII antibodies. Due to the methodological constraints of performing functional assays, measurement of DBPII binding inhibitory antibodies (BIAbs) was performed on a representative subset of the study population comprising 164 individuals, matched for age, sex, malaria exposure and DARC alleles; 58 (35.4%) of these individuals showed BIAbs response against a predominant DBPII variant circulating in the study area (Table 1).
Since the number of malaria cases varied during the course of the study (Fig 2A), we evaluated the long-term antibody response at different levels of malaria transmission. Over three cross-sectional surveys, at 6-month intervals, between 38 to 40% of individuals developed DBPII IgG antibodies, as detected by conventional serology (Fig 2B). Considering the inhibitory antibody response, there was a slight decrease in the frequency of BIAbs at the time of the 3rd cross-sectional survey (34–35% to 25%) (Fig 2C). Finally, the 12-month follow-up study allowed individuals to be classified as persistent non-responders (PNR), transient responders (TR), or persistent responders (PR) for either conventional serology or BIAbs immune response (Fig 2D). For conventional and inhibitory antibody responses, the frequency of acute malaria infections was similar between the PNR, TR or PR groups (p>0.05 for all comparisons).
Of the HLA class II loci that were genotyped in the study population, we found 13 HLA-DRB1, 6 HLA-DQA1, and 5 HLA-DQB1 allele groups. As expected, HLA-DRB1 was the most polymorphic locus with 46 alleles identified; there were 21 and 13 DQB1 and DQA1 alleles, respectively. For each HLA class II locus, the predominant alleles (frequency ≥ 0.01) were listed in the S1 Fig.
In a preliminary analysis, the effect of HLA class II genes on conventional DBPII antibody response was evaluated at the time of the first blood collection (S1 Table). While three HLA class II alleles (DRB1*13:01, DQA1*01:03, DQB1*06:03) were positively associated with anti-DBPII antibody response, six alleles (DRB1*10:01, DRB1*14:02, DQA1*01:01, DQA1*05:03, DQB1*03:01, DQB1*05:01) were negatively associated. Nevertheless, using multiple logistic regression models, only the DRB1*13:01 (presence) and DRB1*14:02 (absence) alleles were significant predictors of anti-DBPII antibodies (Fig 3A).
Since combinations of HLA alleles are inherited together in the genome more often than expected, we further evaluated the association between ELISA-detected DBPII-specific antibodies and HLA class II haplotypes. In total, 126 combinations of specific DRB1, DQA1, DQB1 haplotypes were found, and for 27 of them (frequency ≥ 0.01) it was possible to estimate the individual probability of developing DBPII antibodies. Adjusted logistic regression analysis identified a single haplotype associated with poor production of DBPII antibodies, with individuals carrying the haplotype DRB1*14:02-DQA1*05:03-DQB1*03:01 5-times less likely to develop a conventional DBPII antibody response (Fig 3A). In addition, for each HLA class II locus analyzed (DRB1, DQA1, and DQB1), the genotype frequencies were confirmed to be in Hardy-Weinberg equilibrium (for the responder vs. non-responder groups).
Next, we investigated whether the status of persistent responder (PR) or non-responder (PNR) was HLA class II-linked at the time of the 12-month follow-up collection. The frequencies of some HLA class II alleles were significantly different between the PR and PNR groups (S2 Table). An adjusted odds ratio analysis confirmed that two alleles were associated with the status of long-term responder, and a single allele was associated with the absence of an antibody response (Fig 3B). More specifically, while individuals carrying either the DRB1*13:01 or DQA1*01:03 alleles had an increased probability of a sustained antibody response, the DQA1*05:03 allele carriers were associated with the status of persistent non-responders (Fig 3B). It is noteworthy that the DQA1*05:03 allele aggregated in a specific haplotype (DRB1*14:02-DQA1*05:03-DQB1*03:01) that was primarily associated with the absence of an antibody response (Fig 3A), and this haplotype was in strong linkage disequilibrium (Δ = 1.0; P = 0).
Further experiments investigated whether HLA class II polymorphisms interfered with the functional proprieties of DBPII antibodies. The three cross-sectional measures of DBPII BIAbs responses were performed on a subset of the study population comprising 164 individuals (Table 1), with responders (n = 58) and non-responders (n = 106) matched for age, sex, and malaria exposure. Three alleles (DRB1*07:01, DQA1*02:01, and DQB1*02:02) were overrepresented in DBPII BIAbs responders (S3 Table), and these same alleles aggregated in a haplotype (Fig 4A), which was in linkage disequilibrium (Δ = 0.90 and 0.94, for responders and non-responders, respectively). Interestingly, the long-term persistence of anti-DBPII responses (determined at the 12 month follow-up analysis) was also associated with those same three HLA class II alleles (Fig 4B). Unfortunately, the small size of the sample precluded use of adjusted odds ratio analyses. Nonetheless, responder and non-responder groups were matched by the confounding variables (age, sex, malaria exposure, and dwelling localization).
Since persistence and functional properties of DBPII antibodies were influenced by class II allelic variants, we investigated whether differences in the affinity of DBPII peptides for class II molecules might contribute to the observed difference in responses. Based on predicted binding affinity between DBPII peptides and HLA-DR/DQ alleles, we found unexpected differences in affinity in favor of the non-responder allele carriers. Actually, the HLA-DR allele linked to non-responders (DRB1*14:02) appeared to have a higher binding affinity for the peptides (low IC50 values) than the HLA–DR alleles that were associated with responders (DRB1*07:01 and DRB1*13:01) (S2 Fig). In spite of that, different HLA-DR binding profiles were found for previously identified DBPII epitopes [44–47]. Of note, the recently described broadly neutralizing DBPII epitopes (2D10/2H2 and 2C6) had low binding affinity for all of the Class II molecules analyzed (S2B Fig). Considering HLA-DQ, the model also showed a higher predicted binding affinity for HLA-DQA1/B1 molecules linked to non-responders.
Structural analysis of the three HLA-DRB1 variants revealed a number of interesting differences in the peptide-binding groove (S3 Fig), with significant alterations to its electrostatic potential, and reduced affinity for the model peptides (Fig 5A–5C), which we propose as an explanation for the differences in subject responses. The antigenic surface of DBPII has a strong positive charge (S4 Fig), which is suggestive of a binding preference for antibodies targeting positively charged epitopes such as those that will be preferentially bound by the DRB1*07:01 and DRB1*13:01 variants. Interestingly, while the DRB1*14:02 and DRB1*13:01 variants are the closest in sequence identity, the docked peptides were most similar between the two responder variants (rmsd of the peptides < 3.8 Å), whilst the non-responder did not dock in a similar way (rmsd of the peptides > 8 Å). This supports the suggestion that the non-responder variant leads to reduced antigen presentation.
Overall, while the non-responder-associated variant (DRB1*14:02) shares 81.2% and 95.5% sequence identity to DRB1*07:01 and DRB1*13:01, respectively, the sequence identity is lower in the groove region within 5 Å of the presented antigen (DRB1*07:01–73.3%; DRB1*13:01–89.6%). While the majority of the differences between the variants are located within the peptide-binding domain, this change in the nature of the antigen-binding groove is evident in differences between their isoelectric points, with DRB1*07:01 and DRB1*13:01 having slightly acidic pI’s (6.5 and 7.0 respectively), and DRB1*14:02 being basic (7.7). It was also reflected in the energy calculations, with DRB1*14:02 having an overall Coulomb energy approximately 1.5% lower than either of the responder variants.
One significant difference between the good and poor responder variants was the presence of a glutamine residue at position 99 (247 according to the Protein Data Bank, PDB) in DRB1*14:02, compared to aspartic acid in DRB1*07:01 and DRB1*13:01. This residue side-chain points into the peptide-binding groove and is located approximately 5 Å from the antigen peptide backbone, making a key hydrogen bond (Fig 5D and 5E). Here, loss of the acidic residue in the peptide-binding groove was predicted to reduce the H1 and H3 peptide binding affinity (ΔΔG < -0.8 kCal/mol), and is likely to lead to altered antigen binding and presentation profiles, and hence the poor response seen for DRB1*14:02 variant carriers.
The HLA molecules encoded by MHC class II genes are responsible for presenting peptide epitopes to CD4+ T helper cells. Consequently, it is reasonable to postulate that polymorphism in the HLA class II region may account for the variation in DBPII antibody responses. Unfortunately, most antibody prevalence data on malaria have been collected by cross-sectional analysis at a single time point, which might led to misclassification of individual immune responsiveness. Therefore, in this study we conducted a longitudinal study, collecting serum from the same individuals over a period of 12 months, to obtain a reliable estimate of DBPII antibody prevalence. The genetic profile of HLA class II in the study population was similar to other populations of the Brazilian Amazon [48], which are characterized by an interethnic admixture with high proportions of European and Amerindian groups [49]. Accordingly, in the study population 44% had European ancestry, followed by 38% Amerindian, and 18% African ancestry. For other Brazilian regions, the contribution of European ancestry has ranged from 40% in the Northeast to >70% in the Southeast and South [50, 51]. As expected, the Native-American ancestry in the study area (38%) was representative of the Amazonian region, where Amerindian ancestry is much higher that in other Brazilian regions (<10%) [50].
Our current study confirms the heritability of antibody responses to DBPII, with genetic variation in HLA class II molecules influencing both the development and persistence of an individual’s anti-DBPII IgG antibody response. Accordingly, multivariate analyses adjusted for potential confounding variables showed effects of alleles linked to the DR and DQ loci on the presence (DRB1*13:01) and persistence (DRB1*13:01 and DQA1*01:03) of ELISA-detected DBPII IgG antibodies. On the other hand, two alleles were associated with DBPII-non-responsiveness, DRB1*14:02 and DQA1*05:03, and these comprised a single haplotype (DRB1*14:02-DQA1*05:03-DQB1*03:01) that significantly reduced the development of anti-DBPII IgG at any time during the follow-up study (baseline, 6 and 12 months later). Interestingly, the alleles in the aforementioned haplotype were in strong linkage disequilibrium, which demonstrated that these poor-responder alleles are inherited together more often than expected by chance. So far, only a single study has investigated the association between HLA and DBPII antibodies [52]. Although in that case the authors were unable to demonstrate an association between HLA type and ELISA-detected DBP IgG antibodies, and the relatively limited number of responders did not allow any final conclusions about the highly polymorphic HLA class II and DBP antibodies.
Here, the assessment of long-term antibody response was essential to strengthen the conclusion that there was an increased susceptibility of DRB1*13:01 carriers to develop and sustain their anti-DBPII IgG antibody response. Furthermore, these data confirmed that individuals harboring the haplotype DRB1*14:02-DQA1*05:03-DQB1*03:01 were persistent non-responders. Due to the overall scarcity of data combining analysis of HLA and immune responses to P. vivax, further confirmation of these associations in other malaria endemic areas is needed. Despite the remarkable lack of data on this subject, systematic review and meta-analysis studies have identified a link between the DRB1*13:01 allele and increased antibody responses to vaccines for other microbial infections, including hepatitis B, influenza virus, serogroup C meningococcus, and MMR-II (measles and rubella virus) [53, 54]. Additionally, the DRB1*14 allelic group has been associated with a poor humoral response to HBsAg vaccination [55].
Many field studies examining the immune response to malaria have focused on measuring the concentrations of antibodies to vaccine candidate antigens, while less attention has been paid to complementary approaches defining the functional relevance of these antibodies. By using an in vitro assay to quantify inhibition of DBPII–erythrocyte binding [9, 56], we demonstrated that DBPII binding inhibitory antibodies (BIABs) were associated with three alleles (DRB1*07:01, DQA1*02:01 and DQB1*02:02), which are in linkage disequilibrium and were found to be part of a single haplotype. Notably, these three alleles were associated with the presence of BIAbs antibodies and were also associated with the persistence of this inhibitory response. Therefore, our observations may explain previous results showing that the majority of people who are naturally exposed to P. vivax do not develop antibodies that inhibit the DBPII-DARC interaction, but once they are acquired these BIAbs seem to be stable under continuous exposure to malaria transmission [11, 13].
Intriguingly, a significant number of pharmacogenetic studies have identified HLA-DRB1*07:01 carriers (in less extension, DQA1*02:01 and DQB1*02:02) as being more susceptible to side effects of biological therapy due to the activation of immune response drug-induced [57–59]. Notably, part of the side effect could be explained by the higher production of neutralizing antibodies against drugs (or their metabolites) in the HLA-DRB1*07:01 and/or DQA1*02 carriers [59, 60]. Although drug-specific antibodies are undesirable in therapies involving biological proteins, these findings reinforce our results of a much higher frequency and persistence of DBPII neutralizing antibodies in individuals harboring those alleles HLA class II. In future studies, functional analysis of a greater number of individuals might allow for more robust statistical comparisons.
It is noteworthy that the class II alleles associated with DBPII inhibitory activity were not associated with the conventional IgG antibody responses. Likewise, alleles (or haplotypes) associated with ELISA-detected IgG antibodies were not associated with DBPII BIAbs. These results are not completely unexpected because quantitative receptor binding assays distinguish between antibodies that recognize DBPII and those that inhibit binding to DARC receptor. Here, the DBPII/DARC interaction was assessed by using an established cytoadherence assay based upon multivalent interactions between DBPII on the surface of COS-7 cells and DARC expressed in RBCs [56]. As a consequence, we and others have demonstrated a moderate correlation between DBPII BIAbs and ELISA anti-DBPII antibodies (revised in [8]). Overall, our results emphasize the relevance of examining functional aspects of the immune response, particularly in the case of immunogens such as DBPII, in which the goal of vaccination would be to enhance broadly neutralizing antibodies targeting invasion-blocking epitopes.
To gain insights into the difference between good and poor HLA responders, we sought to investigate whether natural HLA-DR/DQ allelic differences could be explained with respect to binding affinity of DBPII epitopes. While predicted DBPII epitopes have a unexpected moderate-to-high affinity for non-responder alleles, the binding affinity of previously described DBPII epitopes [44–47] was much more variable, including low binding affinities of recently described DBPII B-cells epitopes associated with strain-transcending immunity [47]. However, it seems inappropriate to extrapolate our findings to conformational B cell epitopes because the prediction analyses used here were largely determined by the primary amino acid sequence of the peptide-binding core. In this context, the development of tools for reliably predicting B-cell epitopes, particularly for predicting conformational epitopes, remains a major challenge in immunoinformatics [61].
Although the predicting peptide-HLA binding affinity method used here (NetMHCIIpan—www.cbs.dtu.dk/services/NetMHCIIpan-3.1) [62] seems to be a suitable predictive algorithm for T cell epitopes [28], we performed a further detailed structural in silico analysis of the HLA-DRB1 variants. Significantly, the majority of the differences between HLA-DRB1 variants (good vs. poor responders) were located within the peptide-binding domain, leading to significant changes in the nature of the antigen-binding groove. A striking structural difference between HLA-DRB1 variants was the presence of a glutamine residue at position 99 in the poor responder allele (DRB1*14:02), as compared to aspartic acid in the good responder alleles (DRB1*07:01 and DRB1*13:01). Remarkably, mutation of the corresponding residue has previously been shown to result in loss of the ability of HLA-DP2 to present the metal beryllium to T cells, in genetically susceptible to chronic beryllium- disease [63]. Notably, this single mutation seems to drive helper CD4 T cells in susceptible individuals to secrete Th1-type cytokines, such as gamma-interferon, but not IL-4, leading to beryllium-induced hypersensitivity and chronic beryllium-disease [64, 65]. Consequently, we speculate that the mutation found here in the peptide-binding groove (D247 vs. Q247) is likely to change the outcome of the CD4+T cells immune response. Accordingly, the loss of the acidic residue in the peptide binding groove was predicted to reduce DBPII-specific peptide binding affinity (H1 and H3), which is expected to lead to altered antigen binding and presentation profiles, and hence poor response of carriers of the DRB1*14:02 variant. It strengthens the findings that DRB1*14:02 could be more frequently involved with a poor antibody production [55], while the DRB1*13:01 allele produces a much more robust antibody response [53, 54]. Future studies are required to determine differences in the functionalities of DBPII epitopes in the context of different HLA molecules.
Notwithstanding the relevance of our results, the current study has some limitations. As we focused on the highly variable HLA class II genes it may not have been possible to discriminate between causal alleles and variation that is due to the linkage disequilibrium (LD) between alleles. In fact, in most association studies it has been difficult to pinpoint the causal variants within this genetic complex due to strong LD, population heterogeneity, and the high density of immune-related genes [66]. Such studies have proven most successful for diseases with one prominent predisposing genetic factor mapping to either the class I or class II region [67]. In addition, the associations described in the present study are most likely multifactorial, and depend on several additional factors related to the parasite and host environment. Although the structural analysis of DRB1 variants described here suggested that specific alleles might influence anti-DBPII antibody responses, these results indicate a first step towards the understanding of DBPII immune response in the context of different HLA class-II variants. We are confident that future cellular assays can be pursued to confirm and identify mechanisms associated with good and poor antibody responders. Finally, knowledge of the relative binding affinities of DBPII peptides for class II molecules associated with good and poor responses to this major P. vivax blood-stage vaccine candidate might lead to strategies for controlling the type of helper T cells activated in response to DBPII.
|
10.1371/journal.pntd.0007336 | Effects of ‘The Vicious Worm’ educational tool on Taenia solium knowledge retention in Zambian primary school students after one year | Taenia solium is a neglected zoonotic parasite endemic throughout many low-income countries worldwide, including Zambia, where it causes human and pig diseases with high health and socioeconomic burdens. Lack of knowledge is a recognized risk factor, and consequently targeted health educational programs can decrease parasite transmission and disease occurrence in endemic areas. Preliminary assessment of the computer-based education program ‘The Vicious Worm’ in rural areas of eastern Zambia indicated that it was effective at increasing knowledge of T. solium in primary school students. The aim of this study was to evaluate the impact of ‘The Vicious Worm’ on knowledge retention by re-assessing the same primary school students one year after the initial education workshops.
Follow-up questionnaires were administered in the original three primary schools in eastern Zambia in 2017, 12 months after the original workshops. In total, 86 pupils participated in the follow-up sessions, representing 87% of the initial workshop respondents. Knowledge of T. solium at ‘follow-up’ was significantly higher than at the initial ‘pre’ questionnaire administered during the Vicious Worm workshop that took place one year earlier. While some specifics of the parasite’s life cycle were not completely understood, the key messages for disease prevention, such as the importance of hand washing and properly cooking pork, remained well understood by the students, even one year later.
Results of this study indicate that ‘The Vicious Worm’ may be an effective tool for both short- and long-term T. solium education of rural primary school students in Zambia. Inclusion of educational workshops using ‘The Vicious Worm’ could be recommended for integrated cysticercosis control/elimination programs in sub-Saharan Africa, particularly if the content is simplified to focus on the key messages for prevention of disease transmission.
| The zoonotic parasite Taenia solium, commonly known as the pork tapeworm, causes substantial public health and economic losses worldwide. It is commonly found in low-income countries where pigs are raised in areas of poor sanitation, including Zambia. The links between the parasite and its different disease forms in humans and pigs are not very well known, and ignorance of the parasite is a known risk factor for infection. Health education can significantly increase knowledge and awareness of the parasite and can inspire behavioral change that reduces disease transmission. ‘The Vicious Worm’ is a computer-based program designed to provide T. solium education in a fun and interactive way. We conducted educational workshops in three primary schools in rural areas of eastern Zambia, and preliminary assessment indicated that the ‘Vicious Worm’ educational content significantly improved students’ knowledge of T. solium. We also conducted follow-up studies in the same students one year later, and discovered that the students’ knowledge was still significantly higher than at baseline. We conclude that ‘The Vicious Worm’ may be a useful educational component to enable targeting of school students, and would recommend its inclusion in integrated T. solium control programs in future.
| Taenia solium is a zoonotic parasite known as the pork tapeworm, which infects over 50 million people worldwide [1]. Invasion of the human brain by the larval stage of the parasite is known as neurocysticercosis (NCC), which can cause neurological deficits including severe progressive headache, stroke and hydrocephalus, and is the world’s leading cause of preventable epilepsy [2]. Other impacts of human infection include treatment costs, productivity losses and social stigmatization of epilepsy sufferers [3]. Porcine infections (porcine cysticercosis, PCC) cause substantial economic losses from carcass condemnation, and reductions to farmer income and food safety that exacerbate the poverty cycle in many developing countries in which the parasite is endemic [4, 5].
Despite global ‘tool readiness’ for control of T. solium [6], high levels of active parasite transmission persist in many endemic countries throughout Latin America, Asia and sub-Saharan Africa, including Zambia. Transmission is to a large extent socially determined, with inadequate sanitation, poor hygiene practices, minimal access to medical or veterinary services, and low levels of health education enabling parasite transmission in areas where pigs are raised. A lack of knowledge of the parasite has been identified as one of the barriers for control, and targeted health education interventions have been shown to be an effective addition to other T. solium control measures [7–11]. Education is recognized by the World Health Organization as an important part of the multisectoral approach needed for control of zoonotic pathogens such as T. solium [12]. Computer-based tools have the advantages of providing standardized educational messages, reduce training costs, are able to be widely disseminated and can be updated more easily, compared to traditional paper-based learning systems [13].
‘The Vicious Worm’ (https://theviciousworm.sites.ku.dk) is a freely-downloadable computer-based educational program designed to provide comprehensive information about T. solium in a fun and interactive way. It is set in a sub-Saharan African context and has different levels of detail to allow tailoring of the educational content to suit the needs of the target audience [13]. Studies with medical and agricultural professionals in Tanzania demonstrated significant knowledge uptake and retention, and reported behavioral changes and knowledge dissemination directly attributable to exposure to ‘The Vicious Worm’ [14, 15]. The program had not previously been evaluated for use in school-going children, who have been shown to be effective ‘health change agents’ capable of effectively disseminating educational messages to family and community members [16, 17]. A preliminary study conducted by the authors of this manuscript in three primary schools in the highly T. solium–endemic Eastern Province of Zambia in 2016 demonstrated significant uptake of T. solium-associated knowledge in adolescent primary school pupils in the short-term [18]. The study at hand revisited the same primary school pupils one year later, to evaluate the longer-term impact of ‘The Vicious Worm’ on T. solium–associated knowledge retention.
The study took place in the Nyembe (Katete district), Chimvira and Herode (Sinda district) communities in the Eastern Province of Zambia. As discussed in [18], the region is highly endemic for T. solium; prevalence of active human and pig infections are among the highest in the world, and over 57% of human epilepsy cases are attributable to NCC [19, 20].
‘CYSTISTOP’ is a prospective, large-scale community-based T. solium intervention study, which commenced in three study arms in the Katete and Sinda districts in the Eastern Province of Zambia in 2015. The study has two intervention arms designed to compare integrated human- and pig-based interventions (elimination study arm) versus pig-only (control study arm) interventions, as compared to a negative control study arm. Health education was also conducted at four- (elimination study arm) and twelve-monthly (control and negative control study arms) intervals (Fig 1). Health educational methods included village-based educational sessions during sensitization, conducted in Chewa (the local language) by a trained bilingual CYSTISTOP program member. These sessions included descriptions of the parasite’s life cycle and ways to prevent its transmission in the villages, and utilized visual aids including a large canvas life cycle poster, a five-meter long ribbon to represent the adult tapeworm, and life-sized plasticine models of human stool demonstrating expelled tapeworm proglottids. Participation in village-based sensitization sessions was higher in the elimination study arm than in the control study arm (89% compared to 46%, [35]), and sessions were primarily attended by women, very young children, and few men (personal observation.)
Large color posters of the parasite’s life cycle were permanently displayed at the rural health centers in each of the three study areas. Simplified A4-sized paper copies of the life cycle poster were also distributed to each household in the two intervention study areas (elimination and control study arms) during the baseline visits in October 2015. The final component of CYSTISTOP’s health education intervention was workshops in primary schools using the ‘The Vicious Worm’ computer program.
The educational workshops were conducted in Nyembe (elimination study arm) in July 2016, and in the Kondwelani (control study arm) and Gunda (negative control study arm) primary schools in November 2016 as described in [18]. The initial workshops comprised a ‘pre’ questionnaire to assess baseline knowledge, an educational session using ‘The Vicious Worm’, followed immediately by a ‘post’ questionnaire to evaluate knowledge uptake (see Fig 2). Follow-up sessions were scheduled in the same primary schools in July (elimination study arm) and early December (control and negative control study arms) 2017, one year after the initial workshops.
There were two questionnaires (QS) used in the sessions as per Hobbs et al [18]: the original questionnaire (QS1), modified from the original questionnaire [14] to include Zambian terminology, was used in the elimination study arm and had 24 questions grouped into eight categories. As the QS was deemed too long and complicated for primary school pupils, a simplified version (QS2) containing 15 questions in three categories was subsequently used in the control and negative control study arms. Both QS were designed to test knowledge of human tapeworm infections, known as taeniosis (TS); human (neuro)cysticercosis (NCC/CC); and PCC, including the linkages between the disease states and methods of transmission, diagnosis, and prevention. (The QS used in the sessions are provided in the data repository.)
All of the pupils who had attended the initial educational workshops in 2016 were invited to return for a follow-up session, conducted in the same primary schools in July (elimination study arm) and December (control and negative control study arms) 2017.
Follow-up sessions were conducted as per the ‘post’ QS used in the initial workshops, as described in [18], and were conducted by one of the same two trained bilingual CYSTISTOP project members as in the original workshops. Briefly, QS were projected onto a classroom wall, and questions and answer options were read aloud in Chewa and repeated at least once for clarity. Using Bluetooth-connected TurningPoint clicker devices, all pupils had to individually submit their answer to each question before the group could proceed to the next question. At the conclusion of each session, the group was taken through the QS again to discuss the correct answers and address any remaining misconceptions. The sessions were between 30 (QS2) and 45 (QS1) minutes in duration.
The differences in the two questionnaires prevented direct comparison of response data, so QS1 data (elimination study arm) were analyzed separately from QS2 (control and negative control study arms). Each question was scored as either correct (1) or incorrect (0), resulting in a maximum score of 24 for QS1 and 15 for QS2. Some questions in QS1 had more than one correct answer; selection of any one of these answers resulted in a ‘correct’ outcome. Group (QS1 and QS2) and individual (QS2 only; a technical problem prevented the collection of individual QS1 data during the initial elimination study arm workshop) responses to each session were exported into an Excel (Microsoft Corporation, 2010) spreadsheet for descriptive statistics. Responses were assessed individually and by category.
Grouped result data for QS1 were analyzed using a generalized linear model, using the number of positive and negative answers as binomial response variable, and study time point as categorical covariate. The absence of individual data did not allow taking the within-respondent correlations across study time points into account. Pairwise comparisons of mean scores by study time point were performed using Tukey’s all-pair comparisons method.
Individual result data for QS2 collected at both baseline and follow-up allowed further analyses. The analysis of the correlated ‘pre’, ‘post’ and ‘follow-up’ scores was carried out using a generalized linear mixed model using individual respondent as random effect, the number of positive and negative answers as binomial response variables, and study time point as categorical covariate. Pairwise comparisons of mean scores by study time point were performed using Tukey’s all-pair comparisons method. Additional multivariable analyses were performed adding the respondents’ age, gender, and school. This model was applied to the total scores and to each of the three categories. The analyses were performed using the lme4 and multcomp packages for R 3.5.1 [21–23].
This study was conducted as part of the ongoing CYSTISTOP project (https://clinicaltrials.gov/ct2/show/NCT02612896). Ethical clearance was obtained from the University of Zambia Biomedical Research Ethics Committee (004-09-15) and the Ethical Committee of the University of Antwerp, Belgium (B300201628043, EC UZA16/8/73). The study was introduced and explained to all project participants, both in village group settings and within individual households, prior to each field visit. Written informed consent to participate in the workshops, voluntarily provided by a parent or guardian, was obtained for each pupil, and attendance at the educational sessions was voluntary. The sessions took place outside of normal school hours. There was no incentive for participation, but light refreshments were provided after the sessions.
A total of 86 pupils participated in the follow-up sessions, of whom 55% were female. Ages ranged from 10–18 years, with a median of 14 years. QS1 was taken by 32 of the original 40 pupils whereas QS2 was taken by 54 of the 59 original pupils (83% and 92% follow-up rates, respectively).
Individual analyses of QS2 data revealed that there were no significant differences based on age, gender or village (control study arm vs negative control study arm). Consequently, QS2 data are presented as a consolidated dataset.
This follow-up study indicates that educational workshops using ‘The Vicious Worm’ may have lasting positive effects on T. solium knowledge uptake and retention in rural adolescent primary school pupils in eastern Zambia. Knowledge levels at ‘follow-up’ were significantly higher than at baseline one year earlier, with increases of 14% and 10% compared to ‘pre’ levels in QS1 and QS2, respectively. Compared to ‘post’ knowledge levels immediately following the educational component one year earlier, however, knowledge at ‘follow up’ was similar (QS1) or significantly lower (QS2). The questions relating to general knowledge of TS and NCC, diagnosis of PCC, and prevention of PCC/TS/NCC were answered very well in both QS at ‘follow-up’, with 63% of categories in QS1 and 66% of categories in QS2 answered correctly by at least 75% of the groups. The knowledge regarding prevention of the parasite’s transmission was both the best answered category, and showed the lowest decrease in knowledge from the ‘post’ round one year earlier. This indicates that although some aspects of the parasite’s life cycle remained imperfectly understood at ‘follow-up’, the pupils generally retained the main aspects of T. solium and the key messages for disease prevention one year after ‘The Vicious Worm’ educational workshops.
The parasite’s life cycle is complex, and certain aspects remained imperfectly understood by the pupils at ‘follow-up’. Transmission of PCC was not well understood, nor was transmission of NCC/CC in humans. Many respondents from both QS selected the incorrect answer responses stating that NCC/CC is obtained via ingestion of raw or undercooked pork that is infected with PCC, which given the complexity of the T. solium life cycle is not surprising. Indeed, many other field studies have demonstrated similar results with adults, farmers and even veterinary and medical professionals showing imperfect understanding of the life cycle despite educational interventions [7, 9, 10, 14, 17, 24]. However, what is of concern from these data is that some respondents apparently believed that people with NCC/CC or specifically epilepsy can transmit the disease to others (24%, QS2). Epilepsy is often stigmatized in many low-income countries including Zambia, and the social and psychological effects of stigmatization can substantially decrease quality of life for epilepsy sufferers and their families [25, 26]. While the majority of other respondents correctly indicated that NCC is not transmissible to others, this message should be particularly emphasized in future educational interventions.
Many pupils again selected destruction of the pig and/or carcass as the most suitable method for management of live or slaughtered pigs with PCC, as was also seen in the initial workshops and discussed in [18]. While the ‘correct’ answers for the purposes of the QS scoring were treating pigs with oxfendazole or properly cooking pork, destruction of and proper disposal of heavily infected pork is in fact the recommended approach mandated by World Organization for Animal Health’s (OIE) Terrestrial Animal Health Code [27] and the Zambian Public Health Act [28], and this should be reflected in the marking of these questions in future workshops. However, the OIE Terrestrial Animal Health Code also states that the meat of carcasses infected with less than 20 cysticerci can be consumed after treatment (that is, freeze- or heat-treatment, with the latter reaching a core temperature of 80°C). As ‘backyard’ animal slaughter is frequently conducted in rural and remote communities in many developing countries including Zambia, meat inspection is often rudimentary or absent. Given the limited availability of nutrition and particularly protein in many rural and remote developing communities, insisting on strict measures pertaining to meat inspection and condemnation is not always realistic, and may foster resistance and/or resentment in some situations. We therefore feel it is important to also highlight the alternative options to carcass destruction, especially considering the nutritional needs of these and many other low-resource communities that are endemic for T. solium. Consequently, we would recommend that future educational messages and workshops should recommend destruction of heavily infected meat and carcasses wherever possible, while also promoting proper cooking of lightly infected meat and/or anthelmintic treatment of pigs as more realistic alternatives for some resource-poor endemic communities.
The reason for the decreased knowledge regarding PCC transmission routes seen in students from the control and negative control study arms at ‘follow-up’ (more students indicating that infection arises after pigs being mated with an infected pig, or after eating moldy feedstuff) is unclear, but may be related to the decreased frequency of educational delivery in these study arms compared to in the elimination study arm.
Adolescent primary school pupils were selected to participate in these educational workshops because studies have shown that school students can be ‘health change agents’ capable of effectively disseminating educational messages to family and community members [16, 17]. A cluster-based education trial in northern Tanzania utilized leaflets and videos containing T. solium-specific health education in primary and secondary schools, and demonstrated generally increased knowledge and attitudes in pupils from intervention schools compared to control schools [17].
Using computer-based programs allows standardization of educational messages, while allowing flexibility and adaptation of the content to specific audiences. The recent release of ‘The Vicious Worm’ as a multiplatform smartphone app and the completed translation of the online version into Kiswahili [29], will allow expansion of the program across the African continent. Other language translations are currently underway (personal communication, C. Trevisan), and with adaptation of the illustrations and contexts for Latin American, Asian or other specific settings, this tool could be implemented worldwide. Other electronic educational media including short animated videos, talking books, songs and DVDs are increasingly used in public health campaigns around the world, with encouraging results [30]. In a Chinese study, a short animated cartoon called ‘The Magic Glasses’ was shown to halve infection rates of parasitic worms in school-aged children (8.4–4.1%, P<0.0001), and observed occurrence of handwashing increased from 54% to 98.9% (P<0.0001) in the intervention group compared to the control group [31]. Tablet-based educational interventions have also been successful at raising awareness and changing behaviors for prevention of other, non-parasitic diseases, including cervical cancer and human papilloma-virus infections [32].
It should be emphasized that increased knowledge and awareness of a topic does not necessarily translate into behavioral change, and there may be underlying sociocultural and/or economic factors contributing to parasite transmission in endemic communities that can override even known adverse health outcomes associated with certain behaviors [33, 34]. Student responses given during these assessment situations may indicate what the students believed to be technically correct answers, rather than reflecting their actual behaviors and beliefs. Feedback from focus group discussions conducted in the elimination and control study arms indicated that behavioral changes have been initiated in the villages since the start of the CYSTISTOP project [35], and follow-up observational visits to the study areas are planned for 2019 to corroborate these reports.
The effectiveness of information transfer from educated individuals to others is difficult to quantify, and evaluation of such knowledge transfer was not within the scope of this study. A primary school-based health education trial in Tanzania demonstrated significant knowledge uptake in pupils from intervention schools compared to control schools, whereas evaluation of knowledge transfer to the community showed mixed results [36]: some parents reportedly implemented behavioral changes such as building toilets and boiling drinking water based on knowledge passed on from their children; others reportedly wished to do more but lacked resources to do so; and some parents found it improper for children to instruct their parents. Mwidunda et al [17] reported that secondary school students are often more respected in their families and communities than primary school pupils, and suggested that focusing health educational messages on secondary schools may increase effects of knowledge transfer to communities. No secondary schools are present in the study areas, as is typically the case for many remote and rural regions of Zambia, but conducting Vicious Worm workshops in secondary schools would be encouraged where possible.
This study has limitations. The project activities including health education were conducted more frequently in the elimination study arm (four-monthly) than in the control and negative control study arms (annually), which could have been at least partially responsible for the seemingly better knowledge retention at ‘follow-up’ demonstrated by the elimination study arm students (QS1). The use of two different QS prevented direct comparison of knowledge uptake and retention from individuals across all three study arms, which would have allowed even more robust analyses. In addition, because the technical error in the initial elimination study arm workshop prevented collection of individual response data, we only had grouped result data for QS1, and were consequently not able to take the within-respondent correlation across study time points into account. This led to an underestimation of variances, and consequently an increased probability of (falsely) detecting significant associations. The comparisons across study time points for QS1 should therefore be interpreted with caution. The loss of twelve of the original students to follow-up in this study is another limitation, however statistical significance was nevertheless achieved. Evaluating the effects of knowledge uptake on behavioral change or the extent of knowledge transfer from students to others was outside the scope of this study, but would be useful to attempt in future studies.
In future educational workshops using ‘The Vicious Worm’ it may be beneficial, as per the authors’ previous recommendations [18], to modify the educational component to focus on the main methods for prevention of disease transmission, rather than detailing the T. solium life cycle. Tailoring educational materials to the specific sociocultural context, including use of non-textual media to include individuals with low literacy skills, may further enhance education uptake in endemic communities. The use of locally-broadcast radio programs or simple, illustrative printed material such as posters, leaflets and comic books may also add value to educational programs [7, 8, 11, 37], especially in areas where access to smartphones or computers is limited. Some standardized educational posters are available for T. solium education [38], including several recently published online by the European Network on Taeniosis/Cysticercosis (CYSTINET, COST Action TD1302, http://www.cystinet.org/) (see S1 File).
The results from this follow-up study demonstrate that educational workshops using ‘The Vicious Worm’ can contribute to significantly increased T. solium knowledge in rural Zambian primary school students in both the short- and long-term. Despite some confusion regarding the precise relationships between TS, NCC/CC and PCC, in general the data indicate that the key messages for prevention of disease transmission, including the importance of hand washing and of proper cooking of pork, remained well understood by the students one year after the educational sessions. The flexible nature of ‘The Vicious Worm’ program, combined with recent and ongoing translations into languages other than English and the development of the app for smartphones, provides standardized educational content that can be tailored to the specific educational and sociocultural context of the target audience. For village-level educational interventions in rural endemic communities it may be advised to simplify or omit the more scientific aspects of ‘The Vicious Worm’ in favor of promoting key behavioral messages, to enhance knowledge uptake and retention. Focusing education on school-going children as key change agents may also increase community awareness and engagement. Tailored ‘Vicious Worm’-based educational interventions should be considered for incorporation with integrated T. solium control or elimination programs in future.
|
10.1371/journal.ppat.1005004 | Clearance of Pneumococcal Colonization in Infants Is Delayed through Altered Macrophage Trafficking | Infections are a common cause of infant mortality worldwide, especially due to Streptococcus pneumoniae. Colonization is the prerequisite to invasive pneumococcal disease, and is particularly frequent and prolonged in children, though the mechanisms underlying this susceptibility are unknown. We find that infant mice exhibit prolonged pneumococcal carriage, and are delayed in recruiting macrophages, the effector cells of clearance, into the nasopharyngeal lumen. This lack of macrophage recruitment is paralleled by a failure to upregulate chemokine (C-C) motif ligand 2 (Ccl2 or Mcp-1), a macrophage chemoattractant that is required in adult mice to promote clearance. Baseline expression of Ccl2 and the related chemokine Ccl7 is higher in the infant compared to the adult upper respiratory tract, and this effect requires the infant microbiota. These results demonstrate that signals governing macrophage recruitment are altered at baseline in infant mice, which prevents the development of appropriate innate cell infiltration in response to pneumococcal colonization, delaying clearance of pneumococcal carriage.
| Infants are particularly susceptible to infections, though why is not well understood. One important cause of infant mortality worldwide is infection with Streptococcus pneumoniae, the pneumococcus. All pneumococcal disease begins with asymptomatic colonization of the upper respiratory tract. Infants are also more likely to carry pneumococci, and on average each carriage event has a longer duration. Here, we used an infant mouse model of pneumococcal colonization to study the mechanisms underlying delayed clearance of carriage. We found that infant mice were unable to recruit the effector cells of clearance, macrophages, into the lumen of the upper airway, and that this delay was accompanied by an inability to produce a macrophage chemoattractant in the nasopharynx. We attribute this defect to a dysregulation in the expression of these chemokines and show this effect results from the commensal bacterial flora of infants. Our findings provide an explanation for why infants are more susceptible to being colonized with and infected by pneumococci.
| Many infectious diseases target infants, although our understanding of the host factors that contribute to the increased susceptibility of early childhood remains incomplete [1]. Prominent among the causes of infection of the infant period is Streptococcus pneumoniae, the pneumococcus. Despite effective antibiotics and vaccines, pneumococci are responsible for more than 1 million deaths annually, predominantly in the developing world [2]. Worldwide, this gram-positive bacterium is a common cause of pneumonia at all ages. The spectrum of pneumococcal disease ranges from local infections such as acute otitis media, acute rhinosinusitis and pneumonia, to invasive infections including meningitis and sepsis. In all these diseases, the pathogenic pneumococci initially disseminate from the nasopharynx, a single common site of colonization and carriage [3,4].
Disease, however, represents an evolutionary dead-end for the pneumococcus, since transmission to a new host occurs only via respiratory secretions from the reservoir of bacteria colonizing the nasopharynx [5,6]. Clinical studies and experimental colonization in humans have revealed that different pneumococcal serotypes can colonize repeatedly and concurrently. Each carriage event is maintained for weeks to months before being cleared [3,7,8].
Pneumococcal colonization is particularly common in young children, with a peak prevalence of 55 percent in children 3 years old, declining to 8 percent of 10 year olds and an even smaller proportion of adults [4,9]. Carriage is not only more frequent in children, but is also prolonged. Multiple studies across 3 continents demonstrate a consistent 2-fold increase in the duration of a given pneumococcal colonization event in children compared to adults [10–12]. The mechanism for delayed pneumococcal clearance by infants is not clear, however. One proposed explanation for more efficient clearance with increasing age is the development of antipneumococcal antibodies following clearance of pneumococcal carriage. These anticapsular antibodies cannot be the sole mediator of acquired protection against pneumococci, however, as pneumococcal disease decreases in childhood for all serotypes at a similar rate, a finding that would not be expected if each serotype would need to be carried to generate type-specific anticapsular antibodies [13]. This analysis implies that non-serotype specific mechanisms are responsible for the faster clearance of pneumococcal colonization that occurs with increasing age.
The molecular mechanisms underpinning pneumococcal clearance have been studied in an adult mouse model that faithfully recapitulates multiple aspects of human carriage, including the duration of carriage [14]. Clearance of colonization is independent of the acute inflammatory response and neutrophil influx into the nasopharynx, and furthermore does not require the development of anticapsular antibodies [15,16]. Rather, clearance depends on the recruitment of macrophages into the airway lumen, a process that requires Th17 cell immunity and the expression and sensing of the monocyte chemoattractant chemokine (C-C motif) ligand 2, or CCL2 (MCP-1) [17–19]. Chemokine production and macrophage recruitment occur in response to sensing by pattern recognition receptors TLR2 and Nod2, [18,19] as well as the macrophage scavenger receptor MARCO [20]. It is unclear how these pathways that normally lead to clearance of colonization from an adult host are absent or altered in infant mice.
Here, we show that infant mice are delayed in clearing pneumococcal colonization, and that this prolonged carriage is accompanied by slower macrophage recruitment. We demonstrate that increased macrophage chemoattractant expression due to acquisition of the infant microbiota prevents the formation of a chemokine gradient, and that this lack of chemokine gradient delays macrophage recruitment and pneumococcal clearance.
To determine whether pneumococcal carriage is prolonged in infant mice, adult (6 week old) and infant (7 day old) mice were intranasally inoculated with a pneumococcal isolate that does not cause systemic infection in mice (S1 Table) using a small volume and without anesthesia to prevent aspiration into the lower respiratory tract. At different timepoints, bacterial density was measured by plating nasal lavages. Adult mice started to clear colonization within a week after bacterial challenge, and had fully cleared pneumococci from the nasopharynx by 21 days postinoculation. By contrast, mice inoculated as infants maintained pneumococcal carriage for >6 weeks, and only started to clear colonization at 21 days postinoculation (Fig 1A). Delayed clearance in infant mice was not a strain-specific effect, as it was also seen with a clinical isolate of a different pneumococcal serotype (Fig 1B). The effect of age at inoculation waned over time, as seen by the gradual increase in clearance by 21 days postinoculation in mice inoculated at 7, 14, 24 or 42 days old (Fig 1C). Half of all mice inoculated as infants had cleared colonization by 45 days postinoculation, while half of all mice inoculated as adults cleared colonization completely at approximately 18 days postinoculation.
Since pneumococcal clearance in adults is dependent on the sustained presence of macrophages into the nasopharynx, [18] we examined whether macrophages infiltrated the airway lumen of infant mice by using flow cytometry to quantify different cell populations in nasal lavages. Macrophage influx into the nasopharynx was delayed in infant mice compared to adults (Fig 2A). Inflammatory responses were not completely absent in infants, however, as the neutrophil influx into the nasopharynx in the first week post-inoculation was equivalent between adult and infant mice (Fig 2B). Myeloid cell maturation was not impaired in infant mice, as myeloid cells present in the nasal lavages from adult and infant mice had the same level of CD11b surface expression. (Fig 2C). Further evidence for a lack of age-related difference in macrophage maturation came from analysis of macrophages isolated from the peritoneal cavity of adult and infant mice. These had equal expression of surface receptors MHC class II, CD36 and MARCO, as well as the alternatively-activated macrophage polarization transcript Rtnla (S1 Fig). Furthermore, infant and adult mice were equally capable of mounting a humoral immune response to colonization, as measured by serum titers of antibodies specific to the colonizing strain of pneumococci (Fig 2D). In adult mice, the pattern recognition receptors Nod2 and TLR2 play redundant roles in macrophage recruitment and eventual clearance following pneumococcal colonization [19]. The infant clearance defect was epistatic with these pathways, as there was no additional delay in clearance of colonization in infant mice deficient in these pattern recognition receptors, implying that the infant clearance defect was redundant with these pathways (Fig 2E).
Previous work in adult mice demonstrated that induction of the monocyte/macrophage chemoattractant protein CCL2 (MCP-1) during colonization temporally correlated with clearance and occurs in macrophages in culture following exposure to pneumococci [19]. Colonized infants, however, did not upregulate Ccl2 expression in the upper respiratory tract relative to expression in mock infants, as measured by qRT-PCR on RNA isolated from nasal lavages with RLT lysis buffer (Fig 3A). Due to the small volumes and dilution, it was not possible to reliably measure chemokine concentrations directly in lavage fluid. When directly comparing Ccl2 expression in mock-infected adult and infant mice, baseline levels were significantly higher in the infant URT than the adult (Fig 3A). As with other chemokines, a concentration gradient of CCL2 is required to attract macrophages [21]. The lack of induction of Ccl2 expression in infant mice during colonization suggested the concentration gradient of this macrophage-attracting chemokine was insufficient. Serum CCL2 levels were also elevated in infant mice compared to adults, potentially contributing to the failure to induce a concentration gradient from low CCL2 levels systemically to high levels at the site of colonization (Fig 3B).
We next wanted to identify the source of increased baseline CCL2 in infant mice. Since macrophages were not abundant in the nasal cavity, we turned to a distal site, the peritoneal cavity. We examined macrophage-intrinsic CCL2 signaling by eliciting macrophages from the peritoneal cavity with thioglycollate injection followed by peritoneal lavage 3 days later. Macrophages were purified by adherence, RNA was harvested from cells, and qRT-PCR performed. Ccl2 expression was higher in infant macrophages than adults at baseline (Fig 3C). In the adult nasopharynx, stimulation by pneumococcal colonization led to increased Ccl2 expression, while Ccl2 expression in the infant nasopharynx did not increase above an already elevated baseline (Fig 3A). Macrophage-intrinsic Ccl2 expression followed the same pattern. When stimulated with bacterial lysates, adult peritoneal macrophages increased Ccl2 production, while infant peritoneal macrophages maintained the same elevated level of Ccl2, without further upregulation (Fig 3C). Therefore, cultured macrophages in isolation were sufficient to recapitulate the same pattern of CCL2 expression and upregulation as found in the nasopharynx. This tonic increase in Ccl2 production by infant systemic macrophages was accompanied by a decrease in infant Ccr2 production (Fig 3D).
Baseline expression of the related macrophage chemoattractant Ccl7 (Fig 3E) and proinflammatory cytokine Il6 (Fig 3F) were also elevated in infants compared to adults. In contrast, expression of the neutrophil chemoattractants Cxcl1 (Kc) and Cxcl2 (MIP2) was not elevated in the infant nasopharynx (Fig 3G and 3H). Together these findings suggested an inflamed state in the infant mucosa and that elevated serum and mucosal levels of macrophage chemoattractants in infants compromise the generation of a concentration gradient leading to the nasopharynx.
To induce a concentration gradient of CCL2 in infant mice, we infected 4 day-old infant mice with adeno-associated viral (AAV) vectors expressing GFP (mock/vector control) or murine CCL2. Three days later, mice were colonized with pneumococci. At 7 and 21 days postinoculation, mice were sacrificed, nasal lavages obtained and flow cytometry performed to assess the macrophage influx into the nasopharyngeal lumen. CCL2 overexpression in the URT increased the local gradient in CCL2 concentration, as mice infected with the CCL2-expressing vector exhibited increased Ccl2 transcription (Fig 4A) and CCL2 levels (Fig 4B), while CCL2 concentration in serum was unchanged (Fig 4C). CCL2 protein measurements in nasal lavage fluid are an underestimate, since lavage fluid is at least a 67-fold dilution of the fluid lining the nasal airway surface [22]. The impairment in macrophage recruitment in infant mice was partially recovered by ectopic CCL2 overexpression (Fig 4D). We assessed whether partial macrophage recruitment was sufficient to accelerate pneumococcal clearance by measuring colonization density at 21 days postinoculation, and found a small but significant recovery of the infant defect in clearance (Fig 4E).
We next sought to explain the elevated CCL2 expression in infants that was associated with higher baseline Ccl2 expression in the URT, delayed macrophage recruitment and persistent pneumococcal colonization. Among the changes infants experience during normal development is the acquisition of a stable microbiota [23]. We examined whether the microbiota contributed to C-C motif chemokine expression by treating the drinking water of mice with antibiotics to deplete the flora. Adult mice were directly exposed to antibiotic-treated drinking water, while infant mice were exposed indirectly by treating the water of the dams from which the infants nursed. This indirect exposure was sufficient to decrease the commensal flora of the infant upper respiratory tract (Fig 5A). The magnitude of the depletion of the URT flora was consistent with the decrease in gut microbiota previously found in indirectly exposed infants [24]. Antibiotic treatment had no effect on URT expression of Ccl2 in adults, but decreased baseline infant Ccl2 expression to adult levels (Fig 5B). The microbiota was also responsible for the elevated infant baseline levels of Ccl7 (Fig 5C). Limiting baseline C-C motif chemokine expression in the URT of infants allowed for normal responses to pneumococcal colonization, as macrophages were recruited into the nasopharynx of antibiotic-treated infant mice following 7 days of pneumococcal colonization, unlike tap-water treated mice (Fig 5D). Prior antibiotic treatment accelerated pneumococcal clearance, even 15 days after antibiotics were discontinued (Fig 5E). Even though antibiotics were removed from the drinking water starting 24 hrs before pneumococcal challenge in these experiments, it was still possible that any residual antibiotics could have direct effects on pneumococcal density in the nasopharynx. To exclude this possibility, we measured bacterial load at 7 days postinoculation, before the onset of clearance, and found no effect of antibiotics (Fig 5F). Nasopharyngeal expression of Ccl2 was suppressed in germ-free infants compared to tap-water treated infants, confirming the microbiota was responsible for tonically elevated Ccl2 expression in infants (Fig 5G).
Pneumococcal colonization and disease are more common in children than adults, but the mechanism underlying this predisposition has not been clear. Here, we demonstrated that an infant mouse model of carriage recapitulates the human delay in pneumococcal clearance. Using this model, we found that infant mice have delayed macrophage responses during colonization, which correlated with a failure to upregulate CCL2 signaling. Infant mice had tonic CCL2 production in the URT, indicating a compromised chemokine gradient. Reestablishing a gradient by ectopic overexpression of CCL2 partially restored macrophage recruitment and contributed to pneumococcal clearance. We found that the microbiota contributed to tonic macrophage chemoattractant expression, as depleting the commensal flora lowered expression of CCL2 and CCL7, restored normal macrophage responses and accelerated clearance of pneumococcal colonization. This effect was apparent even 14 days after stopping antibiotic treatment.
Higher pneumococcal loads in the infant nasopharynx have been previously reported, [25] but prior work did not examine innate immune responses in vivo that could explain delayed bacterial clearance, such as macrophage recruitment or expression of a CCL2 concentration gradient. Another study found delayed pneumococcal clearance in elderly mice, which correlated with decreased monocytic phagocyte recruitment and an increased inflammatory state at baseline in elderly mice [26]. This study did not find a role for elevated CCL2 expression in aged mice, however [26]. There is a growing understanding that overly exuberant inflammatory responses can be found both early and late in life, both in humans and mice [27]. Our observation of increased IL-6 expression in the infant URT is consistent with a more pro-inflammatory milieu early in life. It would be important to determine whether alterations in macrophage chemoattractant signaling are a consequence of this more generalized inflammatory state.
We found tonically elevated CCL2 expression in the URT of infant mice, which was dependent on the presence of the microbiota in infant mice. It was not clear whether the effect was restricted to infants due to the recent acquisition, size or composition of the flora in infant mice, or a unique response to the flora of the infant host. The mechanism by which the infant URT responds to the presence or acquisition of the microbiota by increasing production of CCL2, CCL7 and other inflammatory mediators remains unknown. Signaling events in the URT could reflect sensing of the local airway flora, or of commensals at distal sites such as the gut. The infant gut in mice is porous until weaning, [28] which could promote leakage of microbial products outside the containment of the gut lumen into otherwise sterile sites. Constitutive intestinal epithelial NF-κB activity is present in infant mice, and may be associated with endotoxin tolerance [29]. The flora has been shown to systemically prime innate immune responses in both adults [30] and newborns [24]. Our finding that CCL2 levels were elevated in infant serum and in infant macrophages isolated from a sterile site without a local commensal flora of its own, the peritoneal cavity, suggested a systemic effect of the flora on the proinflammatory environment of infants. Alternatively or additionally, sensing of microbial products that stimulate inflammation in infants may occur locally at the site of commensal colonization.
We also found macrophage recruitment to the infant murine nasopharynx was delayed during pneumococcal colonization. There is precedent for this pattern in humans as well, as the number of macrophages recruited to the nasal lumen during URT infections increased with age [31]. This effect, moreover, was independent of the number of prior infections [31]. CCL2 signaling in the human infant airway has not been studied, but there is some evidence for altered CCL2 production. One study found that serum CCL2 levels in normal children were higher than those found in normal adults, [32] consistent with our findings of elevated serum CCL2 in infant mice.
Inducing a gradient of CCL2 by local overexpression in the infant URT partially rescued the defect in macrophage recruitment. This partial recovery was associated with an increase in clearance of pneumococcal colonization. The incomplete recovery may have been due to continued tonically elevated expression of a related macrophage chemoattractant, CCL7 (MCP-3). This chemokine can also bind CCR2, the receptor for CCL2, and both it and CCL2 have additive functions in monocyte homeostasis and recruitment during infection [33,34]. In our study, the microbiota stimulated tonically high expression of both CCL2 and CCL7 in the infant nasopharynx. Together, these data suggest that simultaneous overexpression of CCL7 in addition to CCL2 and possibly other signals could lead to adult-like levels of macrophage recruitment, potentially fully accelerating pneumococcal clearance. CCR2 expression was appropriately suppressed considering the elevated CCL2 levels in infant macrophages, which indicated that the infant defect in CCL2 signaling was not a failure to respond to the ligand.
Altered monocyte/macrophage trafficking and CCL2 signals could be particularly important in mediating infant susceptibility to other infections, such as those with Listeria monocytogenes, which require both CCL2 and recruited monocyte-derived cells for clearance [35]. Infections in infancy are commonly caused by encapsulated bacteria, including opportunistic pathogens that colonize the URT, like the pneumococcus [36]. The delayed clearance of colonization in infant mice resembles tolerance, the failure to respond to an antigen. Elevated inflammatory pathways that cannot be further upregulated could be a mechanism for such tolerance. As a result, the mechanisms described here may reflect a general defect in infant innate immune responses and extend beyond pneumococcal carriage to clearance of other mucosal agents.
This study was conducted according to the guidelines outlined by the Public Health Service Policy on the Humane Care and Use of Laboratory Animals. The protocol was approved by the Institutional Animal Care and Use Committee, University of Pennsylvania Animal Welfare Assurance Number A3079-01, protocol number 803231.
C57Bl/6 mice were obtained from Jackson Laboratory. Germ-free mice were bred and raised in the Penn Gnotobiotic Mouse Facility at the University of Pennsylvania. Procedures were carried out according to an animal protocol approved by the University of Pennsylvania IACUC. For antibiotic treatment, tap water was supplemented with 0.5 g/L ampicillin (Sigma), neomycin (Calbiochem), gentamicin (Invitrogen) and metronidazole (Sigma), as well as 0.25 g/L vancomycin (Santa Cruz Biotechnology), then sterile-filtered. Water was changed every 4–5 days. Mice were sacrificed by CO2 inhalation and cardiac puncture.
Pneumococcal strains used were the clinical isolates TIGR4 (capsule type 4), [37] P1547 (capsule type 6A) and P1121 (capsule type 23F, which is avirulent when injected into the murine bloodstream) [7]. For mouse colonization, pneumococci were grown in tryptic soy broth at 37°C until mid-log phase, then resuspended in sterile PBS. Mice were colonized with doses shown to be sufficient to establish high density colonization, 2x103 CFU for infants and 1x107 CFU for adults [38]. Pilot experiments using the adult dose in both infants and adults showed similar effects on clearance, macrophage recruitment and Ccl2 expression. Mice were sacrificed at indicated timepoints, and nasal lavages obtained with 200 μL sterile PBS, as previously described [19]. Lavages were diluted onto TS agar with catalase (5,000 U/plate) (Worthington Biochemical) and 5 μg/mL neomycin added for quantitative culture overnight at 5% CO2.
Nasal lavages were fixed in 2% paraformaldehyde, and then stained with antibodies to identify macrophages and neutrophils: anti-Ly6G (clone 1A8), anti-CD11b and anti-F4/80. Samples were run on a FACS Calibur instrument (Becton Dickinson) and analysis performed using FlowJo software (Tree Star).
For measurements of anti-pneumococcal antibody titers, pneumococcal strain P1121 was grown and resuspended to an OD620 of 0.1 in coating buffer (0.015 M Na2CO3, 0.035 M NaHCO3), then plated onto Immulon 2HB 96-well plates (Thermo) at 4°C overnight. Plates were washed with 0.05% Brij in PBS, and blocked for 1 hr at 37°C in 1% BSA in PBS. After additional washes, serum samples were added in serial 2-fold dilutions (made in 1% BSA in PBS) and incubated overnight at 4°C. Anti-pneumococcal antibodies were detected by incubating for 1.5 hrs at room temperature with an alkaline phosphatase-conjugated goat anti-mouse IgG antibody, followed by developing for 1 hr at 37°C with p-nitrophenyl phosphatase. Absorbance was measured at 415 nm. The sample dilution at which the absorbance equaled 0.1 was used to calculate the geometric mean titer. For measurements of CCL2 protein levels in serum and nasal lavages, an ELISA kit was used according to the manufacturer’s protocol (eBioscience).
RNA was obtained from URT epithelium by lavage with RLT buffer (Qiagen) with 1% β-mercaptoethanol, or from cultured peritoneal macrophages by lysing cells in RLT buffer with 1% β-mercaptoethanol and frozen at -80°C until used. An RNeasy kit (Qiagen) was used to isolate RNA, and cDNA reverse transcribed by the High-Capacity cDNA Reverse Transcription kit (Applied Biosystems). qRT-PCR reactions were performed with Sybr Green (Applied Biosystems) with 10 ng cDNA and 0.5 μM primers. The ΔΔCT method was used to compare conditions. Primer sequences were as follows: Gapdh-F 5’-AGG-TCG-GTG-TGA-ACG-GAT-TTG-3’; Gapdh-R 5’-TGT-AGA-CCA-TGT-AGT-TGA-GGT-CA-3’; [39] Ccl2-F 5′-AGC-TCT-CTC-TTC-CTC-CAC-CAC-3′; Ccl2-R: 5′-CGT-TAA-CTG-CAT-CTG-GCT-GA-3′; [19] Ccl7-F 5’-GCT-GCT-TTC-AGC-ATC-CAA-GTG-3’; Ccl7-R 5’-CCA-GGG-ACA-CCG-ACT-ACT-G-3’; Il6-F 5’-AGT-TGC-CTT-CTT-GGG-ACT-GA-3’; Il6-R 5’-TCC-ACG-ATT-TCC-CAG-AGA-AC-3’; [40]; Cxcl1-F 5’-CTG-GGA-TTC-ACC-TCA-AGA-ACA-TC-3’; Cxcl1-R 5’-CAG-GGT-CAA-GGC-AAG-CCT-C-3’; [41] Cxcl2-F 5’-CCA-CCA-ACC-ACC-AGG-CTA-C-3’; Cxcl2-R 5’-GCT-TCA-GGG-TCA-AGG-GCA-AA-3’; Ccl2ORF-F 5’-TTA-AAA-ACC-TGG-ATC-GGA-ACC-AA-3’; Ccl2ORF-R 5’-GCA-TTA-GCT-TCA-GAT-TTA-CGG-GT-3’; Ccr2-F 5’-GGT-CAT-GAT-CCC-TAT-GTG-G-3’; Ccr2-R 5’-CTG-GGC-ACC-TGA-TTT-AAA-GG-3’ [42]
Macrophages were obtained by injecting adult and infant mice with thioglycollate, followed 3 days later by peritoneal lavage with cold sterile PBS. Cells were spun down and resuspended in DMEM + 10% FBS. Cells were counted and adjusted to equal concentrations, then plated on 24-well non-tissue culture treated plates. After 2 hrs to allow macrophages to adhere, wells were washed 3 times and then media added back. Cells were used for RNA isolation after an overnight incubation, with or without stimulation with heat-killed bacterial lysates (107 CFU pneumococci in 100 microliters heated to 65°C for 30 min, with an aliquot plated to verify complete killing).
For overexpression, an AAV vector with the capsid from serotype AAV5 was used that expressed the open reading frame of murine CCL2, or GFP for the vector control, under the control of the chicken-beta actin promoter (Vector BioLabs, catalog # AAV-254826 for CCL2, 7006 for GFP). Vectors were concentrated to ~1013 GC/mL, and each mouse was inoculated with 1011 GC of vector.
DNA was extracted from 100 μL nasal lavage samples using the ZR Soil Microbe DNA Miniprep kit according to manufacturer’s instructions (Zymo Research). 16S rDNA copy number was measured using qPCR with a standard curve with a Topo vector containing Escherichia coli 16S rDNA (courtesy of Dr. Frederic Bushman). Reactions were performed using primers, probe and conditions as previously described [43].
Comparisons were made using Prism software (Graphpad). Comparisons between groups for colonization data were made by Mann-Whitney U-test or Kruskal-Wallis test with Dunn’s posttest for two and three or more groups, respectively. All other comparisons were made by unpaired t-test or 1-way ANOVA with Newman-Keuls posttest for two and three or more groups, respectively.
|
10.1371/journal.pntd.0000757 | Emergence of the Asian 1 Genotype of Dengue Virus Serotype 2 in Viet Nam: In Vivo Fitness Advantage and Lineage Replacement in South-East Asia | A better description of the extent and structure of genetic diversity in dengue virus (DENV) in endemic settings is central to its eventual control. To this end we determined the complete coding region sequence of 187 DENV-2 genomes and 68 E genes from viruses sampled from Vietnamese patients between 1995 and 2009. Strikingly, an episode of genotype replacement was observed, with Asian 1 lineage viruses entirely displacing the previously dominant Asian/American lineage viruses. This genotype replacement event also seems to have occurred within DENV-2 in Thailand and Cambodia, suggestive of a major difference in viral fitness. To determine the cause of this major evolutionary event we compared both the infectivity of the Asian 1 and Asian/American genotypes in mosquitoes and their viraemia levels in humans. Although there was little difference in infectivity in mosquitoes, we observed significantly higher plasma viraemia levels in paediatric patients infected with Asian 1 lineage viruses relative to Asian/American viruses, a phenotype that is predicted to result in a higher probability of human-to-mosquito transmission. These results provide a mechanistic basis to a marked change in the genetic structure of DENV-2 and more broadly underscore that an understanding of DENV evolutionary dynamics can inform the development of vaccines and anti-viral drugs.
| Dengue virus (DENV) is a mosquito transmitted RNA virus. One of the most characteristic patterns of DENV evolution is that viral lineages, including whole genotypes, are born and die-out on a regular basis. However, the precise cause of such lineage turnover is unclear. To address this issue we explored the genome-scale genetic diversity and evolution of dengue serotype 2 virus (DENV-2) in Viet Nam, a country with a high burden of dengue disease. We observed that DENV viruses assigned to the Asian 1 lineage were likely introduced into southern Viet Nam in the late 1990's (perhaps from Thailand) and subsequently displaced the resident Asian/American lineage serotype 2 viruses. A similar scenario seems to have occurred in Thailand and Cambodia, such that there appears to have been a region-wide proliferation of Asian 1 lineage viruses. Investigation of Vietnamese patients experiencing DENV-2 infection revealed that Asian 1 viruses also attain higher virus levels in the blood than viruses of the Asian/American lineage. This difference in virus titre is likely to have a profound impact on viral fitness by increasing the probability of mosquito transmission, and therein providing Asian 1 lineage viruses with a selective advantage.
| Dengue viruses (DENV) are vector-borne RNA viruses of the family Flaviviridae and are classified as either four distinct viruses or different serotypes (DENV-1 to DENV-4), each of which can infect humans. Infection with DENV can cause a spectrum of outcomes, ranging from asymptomatic infection through to clinically significant disease. Severe disease in children is commonly characterised by increased vascular permeability, thrombocytopenia and a bleeding diathesis that leads to life-threatening dengue shock syndrome. Each year, approximately 40 million clinically apparent dengue cases occur globally and an estimated two-thirds of the world's population live in areas of risk [1]. There are currently no licensed dengue vaccines or specific interventions to treat the disease.
The error prone RNA-dependent RNA polymerase responsible for genomic replication is central to the generation of DENV genetic diversity, upon which natural selection can act. Although viral fitness is clearly multi-faceted, selection during DENV evolution is likely to be driven in part by the underlying immune status and DENV-infection history of the human population in which the viruses are circulating. For example, patterns of DHF incidence within endemic populations such as Thailand exhibits complex wave-like dynamics with the four viral serotypes co-circulating in a single population and with each serotype dominant on a 8–10 year periodic cycle [2]. An additional layer of complexity exists in the dynamics of serotype oscillations in that there are multiple, genetically distinct viral lineages or genotypes within each virus serotype [3]. Intriguingly, these viral lineages experience a process of ongoing birth and death, possibly because of fitness differences in the human or mosquito host that allow some lineages to survive better than others. For example, in Thailand a turnover of DENV-2 strains was observed between 1980 and 1987 [4] and of DENV-3 strains in the 1990s [5]. Similarly, within-serotype lineage turnover has been documented in Sri Lanka (DENV-3) [6] and Myanmar (DENV-1) [7]. On a wider scale, the introduction of a genetically distinct Asian/American DENV-2 strain into the Americas resulted in the replacement, and possibly extinction, of the resident American DENV-2 lineage [8], with the process particularly well described in Puerto Rico [9]. Lineage replacement can also have epidemiological significance. For example, the lineage replacement events in the Americas (DENV-2) and Sri Lanka (DENV-3) were associated with increases in disease incidence and severity [6], [8] and thereby implying that intrinsic differences in virulence exist between viruses of the same serotype, a phenomenon consistent with earlier field observations [10], [11].
Although the occurrence of lineage replacement is one the most intriguing aspects of DENV molecular epidemiology, its mechanistic basis is unclear. In particular, it is uncertain whether such lineage replacement events reflect (i) large-scale epidemiological processes that are independent of viral genotype, such as random population bottlenecks, perhaps caused by large-scale declines in mosquito numbers during the annual dry season, or (ii) because the viral lineages in question differ in fitness such that one is able out-compete another, perhaps because they possess mutations that allow them to evade cross-protective herd immunity. Choosing between these two models is of central importance because it enables predictions to be made as to what viral lineages are likely to proliferate in the near future, and in so doing allows the more accurate design of vaccines and anti-viral drugs.
The aim of this study was to reveal, for the first time, the changing transmission patterns of DENV serotypes (and genotypes) in southern Viet Nam and their relationship to disease incidence. Rather than focusing on a single gene in isolation, we undertook an expansive genomic approach, resulting in the largest sample of DENV genome (complete coding region) sequences generated to date. Our results reveal a major lineage replacement event that has occurred within Viet Nam specifically, and in South-East Asia more generally, and which has resulted in the dominance of one viral genotype.
The dengue patients from whom virus genome sequences were determined were enrolled into an ongoing (since 2001) prospective, descriptive study at the Hospital for Tropical Diseases in Ho Chi Minh City, Viet Nam. Patients (or their parents/guardians) gave written informed consent to participate in the study. The study protocol was approved by the Hospital for Tropical Diseases and the Oxford University Tropical Research Ethical Committee. Serological investigations (IgM and IgG capture ELISAs) were performed using paired plasma samples using methods described previously [12]. Serology was interpreted as suggestive of secondary infection if DENV-reactive IgG was detected in the capture ELISA in the first week of illness. DENV viraemia levels were determined using an internally-controlled, serotype-specific, real-time RT-PCR assay that has been validated and described previously [13]. Diagnostic culture for DENV was performed by inoculation of 50 µl of plasma onto C6/36 mosquito cells in plastic 75×12 mm tubes. Cultures were incubated at 30°C for 7 days, then blind passaged twice in C6/36 cells for 7 days each.
DENV genomic consensus sequencing was conducted at the Sanger Institute (Cambridge, UK) and the Broad Institute (Cambridge, MA, USA). Two approaches were employed for genomic sequencing. In the first approach (Sanger Institute) (n = 45 viruses), PCR amplification of the DENV-2 genome (C6/36 cells, ≤3 passages) was performed in 3 overlapping amplimers (2.4 kb, 4.5 kb and 4.5 kb, primer sequences available on request). These amplimers were pooled in approximately equimolar concentrations, then randomly sheared by sonication. Sheared DNA fragments were separated according to size by gel electrophoresis and fragments corresponding to a size range of 0.7 to 1.0 kb were removed and shot-gun cloned. From each shotgun library, between 96 and 192 clones were sequenced by dideoxy sequencing using universal and reverse primers. Regions of low or no coverage we filled by specific PCR amplification and sequenced.
In a second approach, viral genomes (Viet Nam, n = 142 and Cambodia n = 39) were sequenced using the Broad Institute's capillary sequencing (Applied Biosystems) directed amplification viral sequencing pipeline (see http://www.broadinstitute.org/annotation/viral/Dengue). This sequencing effort was part of the Broad Institute's Genome Resources in Dengue Consortium (GRID) project. Viral RNA was isolated from diagnostic plasma samples (QIAmp viral RNA mini kit, Qiagen) and the RNA genome reverse transcribed to cDNA with superscript III reverse transcriptase (Invitrogen), random hexamers (Roche) and a specific oligonucleotide targeting the 3′ end of the target genome sequences (nt 10701–10723 for DENV-2). cDNA was then amplified using a high fidelity DNA polymerase (pfu Ultra II, Stratagene) and a pool of specific primers to produce 14 overlapping amplicons of 1.5 to 2 kb in size for a physical coverage of 2X. Amplicons were then sequenced in the forward and reverse direction using primer panels consisting of 96 specific primer pairs, tailed with M13 forward and reverse primer sequence, that produce 500–700 bp amplicons from the target viral genome. Amplicons were then sequenced in the forward and reverse direction using M13 primer. Total coverage delivered post amplification and sequencing was ∼8-fold. Resulting sequence reads were assembled de novo and annotated sing the Broad Institute's in-house viral assembly and annotation algorithms. All genome sequences newly determined here have been deposited in GenBank and assigned accession numbers (Table S1).
In brief, the virus nucleotide sequence from position 817–2520 was amplified in 5 overlapping amplimers of between 403 and 503 nt in length. Amplification was by RT-PCR using a high fidelity polymerase and RNA extracted directly from patient plasma samples as template. Each amplimer was sequenced on both strands by conventional capillary sequencing on an Applied Biosystems 3130 genetic analyzer using specific primers. The resulting sequence reads were assembled and the E gene sequence annotated in AlignX, a software program in the vectorNTI suite (Invitrogen, USA). Primers used for sequence amplification are available from the authors on request. All E gene sequences newly determined here have been deposited in GenBank and assigned accession numbers; GU211738-GU211764, GU434146-GU434159 and GU908494- GU908520).
A sequence alignment was manually constructed (in Se-Al, v2.0) for the complete coding regions (10173 nt) of 187 DENV-2 genome sequences upon which phylogenetic analysis could be undertaken. In addition, to place these data in a wider phylogenetic context, we extracted the envelope (E) gene region from these genomic sequences and combined these data with all those E gene sequences sampled globally and available on GenBank. This resulted in a data set of 941 E gene sequences, 1485 nt in length.
For the 941 E gene sequences we utilized the maximum likelihood (ML) approach available within the PHYML package [14] and incorporating the GTR+Γ4 model of nucleotide substitution. A bootstrap resampling process (1,000 replications) using the neighbor-joining method available in the PAUP* package [15] was used to assess the robustness of individual nodes on the phylogeny, also employing the GTR+Γ4 substitution model and using the parameter settings taken from PHYML. A very similar phylogeny containing the same major groupings was obtained using the ML method within the GARLI package and assuming the simpler HKY85+Γ4 model [16] (tree available from the authors on request).
To place the evolution of the Vietnamese DENV-2 in real time we inferred a Maximum Clade Credibility (MCC) tree of the 187 complete genome sequences using the Bayesian Markov Chain Monte Carlo (MCMC) method available in the BEAST package [17] and employing the exact day of viral sampling. This analysis utilized a strict molecular clock, a GTR+Γ4 model of nucleotide substitution, a different substitution rate for each codon position, and a Bayesian skyline prior as the latter is clearly the best descriptor of the complex population dynamics of DENV. Similar results, with no major differences in topology or coalescent times, were obtained under a relaxed (uncorrelated lognormal) molecular clock and different substitution models (results available from the authors on request). All chains were run for a sufficient length (usually 300 million generations) and multiple times to ensure convergence with 10% removed as burn-in. This analysis also allowed us to estimate times to common ancestry (TMRCA) for key nodes on the DENV-2 phylogeny. The degree of uncertainty in each parameter estimate is provided by the 95% highest posterior density (HPD) values, while posterior probability values provide an assessment of the degree of support for each node on the tree. The BEAST package was also used to infer Bayesian skyline plots for both the Asian I (n = 139) and Asian/American (n = 46) genotypes. This analysis enabled a graphical depiction of changing levels of relative genetic diversity through time (Neτ, where Ne is the effective population size and τ the host-to-host generation time).
Low passage (<4) virus stocks of DENV-2 isolates were prepared on C6/36 cell cultures and quantified by plaque assay on BHK-21 cells with the titre calculated in plaque forming units per ml. All virus inocula for mosquito studies were prepared by infection of C6/36 cells (MOI = 0.1) followed by culture for 4 days at 28°C. Cultures were centrifuged and the culture supernatant used to spike artificial blood meals as described below.
Ae. aegypti from immatures (larvae/pupae) sampled from 5 discrete locations in Ho Chi Minh City (HCMC), Viet Nam during May 2009 were used for generation of low filial number mosquitoes for oral infection studies. Adults were allowed to emerge in the laboratory, mate randomly, and feed on fresh defibrinated sheep blood meal held at 37°C through pieces of parafilm stretched over water-jacketed glass feeders. Adults were allowed to lay eggs on wet filter papers which were then collected and stored on dry pieces of filter paper maintained under high humidity. Larvae were reared on a diet of commercial dog food. Pupae were transferred to screened cages, and adults were fed 10% sucrose in PBS ad libitum. All mosquitoes were maintained in an insectary at 27°C, with a relative humidity of 80% and a 12∶12 light-dark cycle in an environmental chamber. In order to obtain a large number of F2 females, egg batches from multiple gonotrophic cycles were combined and hatched simultaneously.
Artificial blood meals consisted of fresh, defibrinated sheep blood to which serial, ten-fold dilutions of culture supernatant from individual DENV-2 infected C6/36 cells was added at a final dilution of 1/5 per blood meal. Cohorts of 40, 7 day-old female mosquitoes starved of sucrose for 18–24 hrs were offered an infectious blood meal containing either neat or diluted (10−1, 10−2, 10−3) virus culture supernatant for 45 min via membrane feeding as described above. Fully engorged mosquitoes were collected and incubated for 12 days. After 12 days mosquitoes were killed by incubation at −20°C for 1 hr and the mosquito carcass from each individual placed into 0.7 ml of mosquito diluent (MD) consisting of 20% heat-inactivated fetal bovine serum in PBS with 50 µg/ml penicillin/streptomycin, 50 µg/ml gentamicin, and 2.5 µg/ml fungizone. All samples were stored at −80°C before processing. Samples were thawed on ice and homogenized. Infected bodies were detected by quantitative RT-PCR (qRT-PCR) on 140 µl of homogenate (experiment 1) or by NS1 detection using a commercial antigen capture ELISA (BioRad) and 50 µl of homogenate (experiment 1 and 2). NS1 detection was 100% specific and 97% sensitive relative to qRT-PCR for detection of infected carcasses (Simmons et al, unpublished results). The 50% infectious dose values were determined using a three parameter dose-response curve fitted in Prism software (Version 5, Graphpad Software Inc, La Jolla, CA).
All statistical analysis was performed using Intercooled STATA version 9.2 (StataCorp, TX). Significance was assigned at P<0.05 for all parameters and were two-sided. Uncertainty was expressed as 95% confidence intervals. The Mann Whitney test was used for continuous variables. In addition, we compared log10-viremia levels between the two-genotypes using a multiple linear regression model. In addition to the genotype, we adjusted for the following potential confounding factors: age, sex, infection status (primary, secondary, or unknown), and day of illness.
Dengue is hyperendemic in southern Viet Nam. A Ministry of Health coordinated clinical and virological surveillance system operates in the 20 southern provinces of Viet Nam and has revealed oscillations in disease incidence and serotype prevalence in this region since 1996 (Fig. 1A). During 2003–2006, DENV-2 was the most prevalent serotype detected by surveillance and its circulation was temporally associated with increased disease incidence relative to the period 1999–2002 (Fig. 1A). Correspondingly, between 2003–2004, and to a lesser extent 2005–2006, the Hospital for Tropical Diseases in Ho Chi Minh City (HCMC) experienced an increased dengue case burden relative to previous years, most of which was attributable to DENV-2 infection (Fig. 1B).
To understand the evolutionary background to the emergence and decline of DENV-2, we determined the complete coding region (consensus) sequence of this virus from 187 hospitalized patients with residential addresses in or around HCMC and sampled between 2001 and 2008 (accession numbers in Table S1). In total, 143 (77%) coding region sequences were obtained directly from patient plasma, 2 (1.0%) from cerebrospinal fluid and 42 (22%) from cultured virus (all with ≤3 passages in C6/36 cells). Phylogenetic analysis revealed that three major lineages of DENV-2 in Viet Nam, representing the Cosmopolitan (n = 2), Asian/American (n = 46) and Asian 1 (n = 139) genotypes, were present in the sampled population (Fig. 2). The complete coding region phylogeny was also marked by a strong temporal structure, such that viruses belonging to the Asian/American genotype were only sampled between 2001 and 2006, and not thereafter (Fig. 2). In contrast, Asian 1 viruses were only sampled from 2003 onwards but quickly became the dominant genotype in the DENV-2 population (Fig. 2). Bayesian molecular clock analysis suggested the most recent common ancestor of Asian 1 viruses in southern Viet Nam existed between 1996–1998. In contrast, the most recent common ancestor of Asian/American lineage viruses may have originated more than a decade earlier, with a credible range of dates spanning 1988–1991. The dramatic emergence and then dominance of Asian 1 viruses is also apparent on Bayesian skyline plots of viruses from each genotype (Fig. S1). The skyline plot for the Asian/American genotype is characterized by a down turn in relative genetic diversity in recent years, indicative of a decline of incidence, while the equivalent skyline plot for Asian I genotype exhibits both continued growth and a higher level of relative genetic diversity, itself compatible with elevated fitness (Fig. S1). Genome-wide rates of evolutionary change were similar in both viruses at ∼1×10−3 nucleotide substitutions per site, per year. Importantly, the temporal structure of the DENV-2 tree was not related to spatial differences in sampling, with viruses belonging to both clades being sampled from the same geographical area (Fig. S2).
To take a more expansive view of the DENV-2 virus population, we performed a phylogenetic analysis of all globally sampled E gene sequences currently available. This included 68 new E gene sequences generated in the course of this study from DENV-2 viruses sampled from patients in HCMC, Viet Nam between 1995–2009. Furthermore, we included 39 sequences from DENV-2 isolates sampled as part of the Cambodian national dengue surveillance program and also generated in this study (Table S1). Phylogenetic analysis indicated that all available Vietnamese DENV-2 E gene sequences sampled between 1995 to 2002 (n = 45) belonged to the Asian/American genotype (Fig. 3), with Asian 1 viruses being first sampled (by this study) in 2003 and dominating after 2006. The temporal structure in the both the genome (Fig. 2) and E gene trees (Fig. 3), together with the Bayesian molecular clock analysis, supports the contention that Asian 1 genotype viruses were a relatively recent introduction into southern Viet Nam, or circulated at a level below the sensitivity of sampling prior to 2003. Strikingly, there are distinct similarities in the temporal structure of the Vietnamese and Cambodian E gene phylogenies, with Asian 1 genotype viruses also being first sampled in Cambodia in 2003 and apparently displacing the resident Asian/American genotype to the extent that only Asian 1 viruses were sampled in Cambodia after 2005 (Fig. 3). A similar pattern is evident in Thailand, for which sampling is most intense (n = 225), and where both Asian 1 and Asian/American genotype viruses apparently co-circulated in the 1980's, only for Asian 1 viruses to then entirely dominate the sampling since 1991 (Fig. 3). Overall, it is striking that Asian 1 genotype viruses now dominate the sampled DENV-2 population in these three high burden countries. The temporal pattern by which Asian/American and Asian 1 viruses were sampled in Viet Nam, Cambodia and Thailand is summarized in Figure 4. Given the timing and high frequency with which Asian I viruses are sampled in Thailand it is possible that this represents the source population for this genotype, although this will need to be confirmed on a more geographically balanced sample of sequences. More broadly, this E gene phylogenetic analysis reveals that no DENV-2 from South-East Asia sampled after 2006 belonged to the Asian/American genotype, clearly indicating that this viral lineage has suffered a major decline in population frequency in this region. Finally, it is striking that although only four Vietnamese viruses fell into the Cosmopolitan genotype on the basis of E gene sequence data, three of these, sampled from HCMC during the period 2006–2009, clustered together on the tree (Fig. 3). Hence, despite the dominance of Asian I genotype, “refugia” of other viral lineages can persist in localities with large numbers of susceptible hosts.
The branch separating the Vietnamese Asian/American and Asian 1 viruses was characterized by 668 nucleotide and 73 amino acid differences, of which 16 amino acid differences can be considered non-conservative (Table 1). Of the 73 amino acid differences, only three (NS3142K-R, NS3186R-K and NS5563K-R) occurred in previously mapped T cell epitopes [18], suggesting these viruses could have very similar antigenic characteristics for T cells. Two non-conservative amino acid replacements in the E protein – at E226 (T to K) and E228 (G to E) – are of particular note as they had not been previously detected in any of >2000 DENV E gene sequences available on GenBank and occur at amino acid sites that are largely invariant at the global scale. Such an evolutionary pattern is strongly suggestive of a major impact on fitness and it is notable that these changes fall in a surface exposed loop in the E protein [19]. Experimentally determining the fitness effects of these mutations, and perhaps of other that define the Asian I genotype, is clearly an important avenue of future research.
To explore fitness differences in the mosquito host, we first compared the growth dynamics of three viruses from each of the Asian 1 and Asian/American genotypes in Ae. albopictus C6/36 cells in vitro. At multiplicities of infection of 0.1 and 0.01 we did not detect measurable differences in the rate of virus replication or the peak viral titre attained after 6 days of culture (Fig. S3). The 50% infectious dose (ID50) of 3 low-passage virus isolates from each DENV-2 lineage was then measured in low filial cohorts of Ae. aegypti to determine if Asian 1 viruses had a measurable advantage in “infectiousness”. The Ae. aegypti mosquitoes used in these experiments were F2 generation derived from immature forms collected in HCMC, Viet Nam. Twelve days after ingesting a spiked-blood meal, DENV infection was determined by qRT-PCR and detection of NS1 in homogenates of individual mosquito carcasses. The ID50 values determined for each virus were consistent between independent experiments, but there was substantial variation between viruses from the same lineage and there was no apparent trend towards lower ID50 values for Asian 1 lineage viruses (Table 2). These data suggest that infectiousness per se for local Ae. aegypti might not account for the epidemiologically observed fitness differences between these viral lineages.
We hypothesized that the rapid displacement of Asian/American viruses by Asian 1 viruses in Viet Nam reflected a relative fitness advantage that might manifest as a measurable virological feature in patients with dengue. To explore this further, the genotype of DENV-2 virus in 389 pediatric inpatients with DENV-2 infections and admitted to the Hospital for Tropical Diseases between 2001 and 2008 was determined by sequencing of nucleotides 9938-10115 (accession numbers GU211349-GU211737) and alignment with reference sequences. All patients were in a prospective study of dengue and admitted to the same pediatric ward. The serological and demographic and features were not significantly different between the two groups (Table 3). Fever clearance times (a surrogate of duration of viral infection) did not differ significantly between patients with Asian 1 (n = 289) or Asian/American (n = 100) DENV-2 infections (P = 0.27, log-rank test), nor did platelet nadirs (P = 0.07, Mann Whitney test) or maximum haematocrit levels (P = 0.16, Mann Whitney test). Intriguingly, however, viremia levels were significantly higher in patients with Asian 1 DENV-2 infections compared to Asian/American DENV-2 infections at the time of study enrolment (Fig. 5). In a multivariate analysis, log10-viremia levels were significantly higher for the Asian 1 genotype at the time of study enrolment: the adjusted mean difference between the two genotypes was 0.94 (95% CI 0.60 to 1.27; p<0.001). Higher viremias in Asian 1 DENV-2 infections could plausibly lead to an increased rate of human-to-mosquito transmission and hence an elevated fitness.
Oscillations in dengue incidence and serotype prevalence are a characteristic feature of the epidemiology of this virus. Although changes in genotype composition through time are similarly a relatively common observation in studies of DENV evolution, to date there has been generally insufficient data to determine (i) whether changing patterns of genotype prevalence have any association with viral phenotype including the virological manifestation of DENV infection, and (ii) the evolutionary basis of these lineage replacement events, and specifically the respective roles of natural selection versus genetic drift.
Here we demonstrate a rapid and apparently complete lineage (genotype) replacement event within DENV-2 in southern Viet Nam that was temporally associated with an increase in disease incidence. Strikingly, a functional basis for the displacement of resident Asian/American lineage viruses was suggested by higher viremia levels in pediatric patients with Asian 1 DENV-2 infections. The presence of higher viremias in children hospitalized with Asian 1 DENV-2 infections relative to Asian/American DENV-2 infections would likely increase the probability of human-to-mosquito transmission and hence facilitate greater population diffusion. Another possible outcome of higher viraemia levels is an increased incidence of more severe disease. We did not detect significant differences in the extent of capillary permeability or thrombocytopaenia between patients with Asian 1 or Asian/American viruses in the cohort of hospitalized patients (n = 389) here. This suggests that Asian 1 DENV-2 infections were not overtly associated with more severe disease. However, this was a relatively small sample size to detect differences in clinical outcomes and a larger cohort of symptomatic patients, including non-hospitalised individuals, would most likely be needed to answer this question definitively.
Our best estimates suggest the Asian 1 genotype was first introduced into southern Viet Nam in the late 1990's. Our sampling illuminated the replacement of the previously dominant Asian/American genotype by Asian 1 viruses during 2003–2007, a period in which there was an almost doubling of dengue incidence, mostly associated with DENV-2.
A large number of susceptible hosts in the population, and an associated increased force of infection, could help explain the seemingly short period in which genotype replacement occurred. Whether the apparent fitness advantage of Asian 1 viruses could be attributable to “antigenic fitness” in the face of the population-wide immune landscape during this period is unknown. However, there is a precedent for biologically relevant antigenic differences between genotypes of DENV-2. For example, South-East Asian DENV-2 viruses are less susceptible than American lineage viruses to cross-neutralization by antibodies elicited by DENV-1 infection [20]. Population wide seroepidemiology, coupled with a better understanding of correlates of immunity, are clearly needed to understand serotype and genotype replacement in all endemic regions.
Virus traits in the mosquito host might also explain the difference in fitness between the Asian I and Asian/American genotypes. As an example, previous studies have demonstrated “SE Asian genotype” viruses (there was no analysis of differences between Asian 1 or Asian/American genotypes) are more infectious and disseminate faster in Ae. aegypti mosquitoes, and replicate more efficiently in human dendritic cells, than American genotype DENV-2 viruses [21], [22], [23]. However, we were unable to detect a measurable difference between Asian 1 or Asian/American viruses in terms of overall replication rates in C6/36 mosquito cells, or infectiousness for local Ae. aegypti mosquitoes. Other features of the virus-mosquito interaction (e.g. extrinsic incubation time) could be equally or more important than the infectious dose. However, published data from Armstrong et al. suggested that Asian 1 and Asian/American viruses had similar dissemination rates in Ae. aegypti mosquitoes [22]. Further studies are therefore needed to understand the importance of the mosquito as a site for the expression of the fitness differences between these two virus lineages.
Intriguingly, the displacement of Asian/American lineage viruses by Asian 1 viruses has also seemingly occurred in Thailand and Cambodia. In Thailand, the Asian/American genotype most likely co-circulated with the Asian 1 genotype for at least a decade prior to 1991, but is then absent from amongst the 139 Thai DENV-2 viruses sampled between 1992 and 2006, which all belong to the Asian 1 lineage. In Cambodia, despite a smaller sample size, lineage replacement appears to have occurred along a remarkably similar time-frame to that seen in Viet Nam, with only Asian 1 viruses being sampled after 2005. Both Thailand and Cambodia have considerable transport links with Viet Nam and it's conceivable these are sources for the introduction of Asian 1 viruses into Viet Nam.
In sum, we show that a lineage replacement event in DENV is highly likely to be linked to an underlying difference in fitness. This suggests that natural selection may play a more important role in shaping viral dynamics than previously realized. The virus genetic traits associated with the fitness of Asian 1 viruses are difficult to identify definitively on the basis of sequence data alone. Elucidation of the possible functional consequences of the 16 amino acid differences that characterize the Asian I viruses will clearly require complex reverse genetic experiments. However, we predict that the amino acid changes at E226 and E228 will be of particular importance given that they occur at sites that are invariant across all DENV sequences sampled to date (and which suggest that the vast majority of mutations at these sites are strongly deleterious because they have a major impact on fitness). Our documentation of a major lineage replacement event, coupled with the current dominance of Asian I viruses, suggests that this genotype will continue to dominate DENV-2 infections in Thailand, Cambodia and Viet Nam unless there is a major change in the host environment, such as that brought on by changes in serotype (and which themselves exhibit complex population dynamics [24]).
The prevalence of Asian I DENV-2 viruses has multiple implications. First, it is paramount that the DENV-2 component of future dengue vaccines (reviewed in [25]) be competent at eliciting immunity to viruses belonging to this genotype. Similarly, programs to develop anti-viral drugs for dengue should include Asian 1 DENV-2 viruses in their pre-clinical discovery and development programs [26]. Furthermore, we would predict that Asian 1 viruses will continue to outcompete Asian/American DENV-2 viruses. A likely future setting for this event is in the Americas where currently Asian/American DENV-2 viruses predominate, having themselves displaced the American genotype.
|
10.1371/journal.pcbi.1002170 | Identification of a Novel Class of Farnesylation Targets by Structure-Based Modeling of Binding Specificity | Farnesylation is an important post-translational modification catalyzed by farnesyltransferase (FTase). Until recently it was believed that a C-terminal CaaX motif is required for farnesylation, but recent experiments have revealed larger substrate diversity. In this study, we propose a general structural modeling scheme to account for peptide binding specificity and recapitulate the experimentally derived selectivity profile of FTase in vitro. In addition to highly accurate recovery of known FTase targets, we also identify a range of novel potential targets in the human genome, including a new substrate class with an acidic C-terminal residue (CxxD/E). In vitro experiments verified farnesylation of 26/29 tested peptides, including both novel human targets, as well as peptides predicted to tightly bind FTase. This study extends the putative range of biological farnesylation substrates. Moreover, it suggests that the ability of a peptide to bind FTase is a main determinant for the farnesylation reaction. Finally, simple adaptation of our approach can contribute to more accurate and complete elucidation of peptide-mediated interactions and modifications in the cell.
| Linear sequence motifs serve as recognition sites for protein-protein interactions as well as for post-translational modifications. One such motif is the CaaX box located at protein C-termini that serves as prenylation site. This prenylation is critical for many signal transduction related proteins and it is thus an important goal to uncover the range of prenylated proteins. Due to poor generalization ability, sequence based computational methods can only go so far in predicting novel targets. In this study, we introduce a novel structure based modeling approach that allows both recovery of known farnesylation substrates, as well as detection of a new class of farnesylation targets. We demonstrate high accuracy in retrospective discrimination between substrates and non-substrates of farnesyltransferase (FTase). More importantly, in a prospective study, in vitro experiments validate that 26/29 predicted peptides indeed undergo farnesylation. These novel peptides were derived either from actual human proteins, or predicted to bind particularly well to FTase. Other than the discovery of putative novel farnesylation targets in the human genome, as well as possible inhibitors, we provide insights into the main determinants of farnesylation. Our approach could be easily extended to additional peptide-protein interactions and help the elucidation of the cellular peptide-protein interaction network.
| Protein prenylation is a post-translational modification in which a prenyl group (farnesyl or geranylgeranyl) is attached to the protein via a thioether bond to a cysteine at or near the carboxy terminus of the protein (reviewed in [1], [2]). Protein farnesyltransferase (FTase) and geranylgeranyltransferase type I (GGTase-I) are also called CaaX prenyltransferases, due to their ability to catalyze modification of peptides and substrate proteins bearing the carboxy terminal (C’) Cys-aliphatic-aliphatic-variable amino acid (Ca1a2X) motif [3].
Upon binding of the substrate and the C-terminal Ca1a2X motif, the catalytic zinc ion of FTase coordinates the thiol side chain of the cysteine and catalyzes the covalent attachment of the lipid anchor to this residue. A detailed view of this mechanism has been obtained by a series of structures solved at different stages of the reaction [4]. After the covalent attachment of the isoprenoid in the cytoplasm, substrate proteins can undergo further processing, resulting in a C’ structure that is able to serve as a specific recognition motif in certain protein-protein interactions [5] and to direct the modified protein towards incorporation into cellular membranes [6].
A wide range of proteins involved in diverse cellular functions require this post-translational modification for their action [2]. While numerous proteins have been experimentally shown to undergo farnesylation in vivo [7], [8], [9], it is likely that many FTase substrates remain to be discovered. There is a wide interest in the mapping of FTase targets in the genome, in part due to the therapeutic potential of FTase inhibitors against cancer [10], [11], [12], as well as parasitic infection [13], [14]. Identification of new targets might lead to novel therapeutic approaches [15]. Moreover, the elucidation of cellular FTase targets might shed light on the function of various proteins, as well as on the cellular network of interactions.
Computational approaches have predicted FTase targets based on sequence features of known targets [7], [8]. These methods show good performance in terms of sensitivity, i.e. known targets are correctly identified. Thus, prenylation is mainly defined by the last four residues of the protein, although additional weaker sequence constraints have also been identified upstream in the sequence [16]. Other approaches were based on manual inspection and derived from structural features [9].
Substrate specificity has also been examined using peptide libraries. A comprehensive study by Hougland et al. on the farnesylation of a large synthetic peptide library has allowed a detailed characterization of FTase specificity [17]. In addition to compiling a large and clean dataset of peptides that contains both efficient substrates and non-substrates for FTase, this study discovered a third group of sequences that are farnesylated only under single-turnover (STO) conditions ([E]>[S]). Analysis of peptide substrates has also demonstrated that reactivity depends on synergy between the side chains at the a2 and X positions [18]. These findings indicate that FTase substrate recognition is more complex than the simple Ca1a2X motif model, and that non-canonical sequences can serve as substrates.
A large number of structures have been determined for FTase and FTase-substrate peptide complexes [19]. The peptide binding pocket is well-characterized, although a structure of the ternary FTase•farnesyl diphosphate(FPP)•peptide in an active conformation has not been determined [9]. The Ca1a2X cysteine sulfur atom (prior to the product formation) coordinates the catalytic Zn2+ ion together with side chains (D297, C299 and H362) of the FTase β-subunit. The a1 side chain points out of the binding pocket and faces the solvent, while the a2 side chain is buried within the binding pocket and interacts both with the farnesyl chain of FPP and the residues lining the pocket. The C’ X position interacts with residues mostly from the FTase β-subunit and is considered the main determinant for the specificity between FTase and GGTase-I 9. Finally, two highly conserved hydrogen bonds are formed: 1) between the C-terminal carboxylate group and the side chain of FTase Q167α and 2) between the a2 backbone carbonyl oxygen and the side chain of FTase R202β (Figure 1). Despite this detailed structural information, only a handful of different peptide sequences have been solved in complex with FTase.
We previously developed a scheme for modeling the structures of peptide-protein complexes (Rosetta FlexPepDock [20]), which is incorporated within the Rosetta modeling suite framework [21]. This protocol is the starting point for the development of a structure-based scheme for the prediction of peptide binding specificity (FlexPepBind). Specifically, to refine FlexPepBind for the prediction of FTase binding peptides, we have incorporated constraints derived from the conserved features in solved FTase structures and adapted the energy function to distinguish between reacting and non-reacting tetrapeptides (based on an underlying assumption that tetrapeptides that bind will react, while those that do not bind will not react). We trained and tested this protocol against the recent dataset published by Hougland et al. [17]. Validation of the protocol against several independent sets showed accurate prediction of peptides that could be farnesylated, both under multiple turnover (MTO) and single turnover (STO) conditions. Evaluation of all possible Cxxx peptides identified a previously uncharacterized class of farnesylation targets that contain an acidic C-terminal residue. The 13 peptides predicted to bind with best affinity were experimentally shown to indeed undergo farnesylation in vitro. Finally, a genomic scan for novel FTase targets revealed 77 novel putative FTase targets previously undetected by sequence-based approaches. Among these, 13 out of 16 selected novel putative farnesylation targets were indeed farnesylated by FTase in an in vitro experimental validation. FTase-peptide binding is a model system for our approach to peptide-protein binding specificity prediction and design. Our protocol can easily be adapted to additional peptide-protein interactions where both experimental structure and affinity data are available, thereby providing a mechanism to identify targets not detectable by sequence conservation only.
Recently Hougland et al. performed a large-scale study, in which they characterized a TKCxxx peptide library for reactivity with rat protein farnesyltransferase (rat FTase) [17]. Out of an initial library of 213 sequences, 77 peptides are farnesylated under multiple turnover (MTO) conditions, and 51 sequences are not farnesylated under any conditions. Interestingly, the remaining 85 sequences are farnesylated under single turnover (STO) conditions but not under MTO conditions.
We set out to use FlexPepBind and the structural data available for FTase to discriminate MTO sequences from non-reactive (NON) peptide sequences, using the 77 MTO and 51 NON peptide sequences as our training set (128 peptides in total; Dataset S1A). Towards this end, we used the high resolution structure of human FTase in complex with a peptide derived from the carboxy terminus of Rap2a and a farnesyl diphosphate (FPP) analog (PDB: 1tn6 [9]) to create a starting model. The bound peptide was truncated to include only the terminal Ca1a2X motif. Different peptide sequences were then threaded onto the peptide backbone and used as starting structures.
Initially, we modeled peptide-FTase complex structures for different peptide sequences by applying the Rosetta FlexPepDock protocol to the threaded starting models. This protocol was developed previously in our lab for the modeling and refinement of peptide-protein complex structures to high resolution [20]. Our simulations included three constraints, namely the conservation of the 2 structurally conserved hydrogen bonds (C’ carboxylate - FTase Q167α; a2 backbone carbonyl oxygen - FTase R202β) and the location of the cysteine sulfur atom coordinating the Zn2+ ion (Figure 1, see Methods for more details).
For each simulation, the energy of the best scoring Cxxx peptide was extracted (see Methods for further details). Figure 2A shows the Receiver Operating Characteristic (ROC) plot for the ability of the peptide energy to discriminate between MTO sequences and non-substrate sequences. The plot shows very good discrimination with an Area Under the ROC Curve (AUC) value of 0.915 on our training set.
These results demonstrate that a structure-based evaluation of the peptide energy can distinguish very well between farnesylated and non-farnesylated peptide sequences. Since the known constraints restrict the simulation to a closely defined region in the binding site, we reasoned that a simpler and faster protocol might be able to model the peptides with similar accuracy. Our simplified protocol therefore includes only a minimization using the Rosetta energy function [21], [22] under constraints to retain the 2 structurally conserved hydrogen bonds and the cysteine sulfur atom location coordinating the Zn2+ ion (see above and Methods for more details). This protocol yielded similar results with an AUC value of 0.875 on the training set. A peptide energy threshold of -0.4 (i.e. sequences with energy below/above -0.4 are predicted to be binders/non-binders and therefore farnesylated/non-farnesylated, respectively) corresponds to a 69% True Positive Rate (TPR) and 8% False Positive Rate (FPR). A more stringent threshold of -1.1 energy units corresponds to a 44% TPR and 2% FPR (Figure 2A). With the two protocols exhibiting similar performance, we decided to proceed further using the fast minimization protocol. (Performance on the training set using additional sampling and scoring schemes is summarized in Table S1.)
To assess FlexPepBind using the selected thresholds, we evaluated performance on three independent test sets (Dataset S1B-D online).
Using FlexPepBind, we modeled all of the 8000 possible Cxxx sequences and scored them according to our protocol. The thresholds for the discrimination of MTO/NON predict that 1349 (17%; stringent threshold = −1.1) and 2309 (29%; threshold = −0.4) of all tetramer peptide sequences could be possible substrates (see Figure 3). This set of putative farnesylation targets suggest a much more versatile binding motif than previously accepted (see Figure 4): while position a2 of the Ca1a2X motif is still prominently aliphatic (ILE/VAL/LEU/PHE), positions a1 and X are less restricted than previously reported (compare Figure 4C to Figures 4A&B). In particular, we identify within this set a novel class of farnesylation targets that contain an acidic residue at the C-terminus (238/1349 putative targets; ∼20%; see Figure 4D).
Figure 4C indicates that the minimization-based protocol tends to miss larger residues at the C-terminal X position. Indeed, assessment of the prediction accuracy for this position on the training set shows that only 1/8 CxxF and 0/3 CxxW sequences are correctly predicted with the chosen protocol (CxxM peptides are predicted with higher accuracy: 10/14). Using the FlexPepDock based protocol, performance increases to: 6/8 CxxF; 2/3 CxxW and 11/14 CxxM, demonstrating that CxxF peptides are indeed rescued by the additional backbone flexibility. Therefore, it might be advisable to use the FlexPepDock based protocol for peptides that contain a bulky C-terminal side chain.
We compared our predictions to the PrePS [7] prediction of prenylation targets on the initial training set of peptides. Regarding the discrimination of MTO substrates from non-active peptides, PrePS results are comparable to FlexPepBind (AUC of 0.92, with a threshold corresponding to 60% TPR for 2% FPR). However, the performance for STO peptides is significantly better for our structure-based approach: while FlexPepBind recovers 47% and 32% of the STOs with the loose and stringent thresholds concordantly, PrePS predicts only 14% of these sequences as substrates.
Since our retrospective studies indicated that our approach can very accurately retrieve actual farnesylation targets, we were interested in testing it prospectively – could novel targets be indeed identified? We selected the 13 best scoring peptides (i.e. predicted tightest binders), yet previously uncharacterized for experimental validation. These are mostly ‘non-canonical’ peptides, including 5 peptides with an acidic C-terminal residue. Indeed, PrePS [7] predicts only 2 out of the top-scorers to be FTase substrates. In vitro farnesylation assays indicate that all of these peptides indeed undergo farnesylation catalyzed by FTase: 10 under MTO conditions and 3 under STO conditions (Table 1A). These results demonstrate the robustness of our protocol and its exceptional accuracy. Importantly, they confirm the novel class of farnesylation substrates that contain a negatively charged C-terminal residue (Figure 4D).
Structural investigation of this novel class of substrates suggests that the negatively charged C' side-chain is stabilized by FTase residue His 149βwhile accepting a hydrogen bond from Trp102β (GLU) and creating an additional hydrogen bond with the side-chain of Ser99β (GLU & ASP) (see Figure S1). Additional polar interactions with water molecules are possible but were not explicitly modeled.
Equipped with a score that can predict both known and novel FTase targets, we set out to scan the human genome for proteins that may undergo farnesylation. Our protocol was developed based on experimental assays on rat FTase (and the structure of human FTase [9]). Since rat and human FTases show very high sequence identity (92% and 96% for subunits α and β respectively), and none of the sequence differences are located at or near the peptide binding site, we are confident that our prediction scheme can be applied to human farnesylation as well.
We identified 756 unique proteins in human SwissProt [24] that contain the Cxxx motif at their carboxy terminus. 167 and 309 of these protein sequences obtained scores lower than the −1.1 and −0.4 threshold values, respectively, indicating that these proteins might be farnesylated by FTase. We focused on the group of 167 proteins detected with the more stringent threshold.
Could these proteins indeed be FTase substrates? Several indications support our predictions: First, amongst the 167 candidates, 42 contain a Cxxx motif of a known FTase substrate. Secondly, the Gene Ontology (GO) [25] cellular compartment annotation for most of these 167 proteins is Membrane related (see Figure S2; see Methods for more details). This supports their association with membranes, possibly by farnesylation (albeit this localization annotation might have been inferred from sequence similarity). Furthermore, peptide library studies have demonstrated FTase-catalyzed farnesylation (under STO or MTO conditions) of 50 of these Cxxx motifs (representing 66 human proteins) [17]. Finally, analysis of the putative target proteins with the PrePS server predicts that most of them (90/167) are indeed FTase targets, while the other 77 are not predicted to be farnesylated (see Figure S3). To further characterize the latter, we proceeded with in vitro experimental validation of selected sequences.
Among these 77 proteins (containing 72 unique Cxxx motifs), 39 motifs had not yet been tested for in vitro farnesylation. The second set chosen for experimental validation consisted of 16 top-scoring peptides selected from these 39 motifs. Of the 16 tested peptides, 9 and 4 peptides are farnesylated in vitro under MTO and STO conditions, respectively, while only 3 were not farnesylated by FTase (Table 1B). None of the 16 sequences in this second set are predicted to serve as farnesylation targets by PrePS. Interestingly, for 9 of these 16 sequences, PrePS predicts that the upstream context of the motif is suitable for farnesylation. In these cases, the PrePS negative prediction is based on the sequence of the Cxxx motif. This suggests that improved characterization of the contribution of the 4 C-terminal residues to farnesylation can identify more farnesylation targets. Finally, for 8 of these 16 sequences, PrePS would predict farnesylation of the Cxxx motifs in the background of the favorable H-Ras upstream sequence. The balance between the upstream signal and the C-terminal Cxxx motif is therefore an interesting subject for future research.
Most of the proteins identified by this study as novel FTase substrates have not been well characterized to date. Consequently, in vivo experiments that evaluate the cellular localization and prenylation status of these proteins, in conjunction with the in vitro farnesylation demonstrated in this study, will advance their functional characterization.
We present here a simple and accurate structure-based scheme for prediction of the sequence of FTase-binding peptides. We have validated our protocol against several test sets, and predictions were experimentally verified in vitro to reveal novel putative FTase substrates and potential tight binders. This approach has expanded our understanding of farnesylation, both within the context of the reaction itself, as well as in the greater context of cellular biology. Furthermore, this protocol presents an advance in the computational prediction of binding specificity in general.
The protocol that we developed essentially estimates the binding affinity of FTase for Cxxx peptides, using a training set of reactive peptides, rather than predicting the farnesylation activity of these sequences. This has several implications and limitations. Remarkably, the ability to discriminate peptides that undergo MTO reaction from non-active peptides according to binding energy suggests that the non-active peptides may bind weakly or not at all to FTase (see Figure 3). This finding is supported by results from an in vitro inhibition experiment in which none of the tested non-active peptides inhibited FTase-catalyzed farnesylation of a known substrate [17]. In turn, the members of the small class of FlexPepBind false positive peptides may bind to FTase with high affinity but still not be farnesylated. These false positive peptides could therefore serve as FTase inhibitors and represent an interesting set to characterize in future work.
Previous studies have shown that the sequence immediately upstream of the conserved cysteine residue may also play a role in substrate selectivity [16]. These sequences modulate peptide affinity and reactivity with FTase, i.e. a high-affinity terminal tetramer sequence does not necessarily ensure farnesylation of the protein. For half of the proteins tested in the study, the PrePS [7] program predicts favorable upstream sequences. This result coupled with the high-affinity -Cxxx motif predicted by FlexPepBind (see Results and Table 1B) increases the confidence that the human proteins containing the said Cxxx motif could be farnesylated in vivo. In turn, a favorable upstream sequence might compensate for a weak C-terminal signal. Our future work will therefore further characterize the balance between these two signals in determining farnesylation.
One puzzling aspect of FTase substrate recognition is the large number of peptides that exhibit single turnover activity. The single turnover rate constant, kfarn, reflects all of the rate constants up to but not including the release of the farnesylated product [4], [26], [27], [28]. Therefore, the STO peptides bind to FTase and are readily farnesylated, but the product dissociates very slowly so multiple turnover activity is very slow. Consistent with this, FlexPepBind achieves an AUC value of 0.776 in the discrimination between STO and non-active peptides on the training set, indicating that STO peptides have higher affinity for FTase than the non-active peptides (see Figure 3). Our protocol thus identifies STO peptides much better than sequence-based methods (see Results and Hougland et al. [17]).
What then discriminates between MTO and STO peptides? Hougland et al. postulated that the farnesylated STO peptides might bind more tightly to FTase than farnesylated MTO peptides, and as a consequence FPP-catalyzed product dissociation is slow [17]. However, binding energy, as approximated by our approach, seems to be a poor discriminator between MTO and STO peptides (AUC value of 0.625 on the training set – Dataset S1B). That is, estimation of the binding affinity of peptides in the context of static conformations of the protein cannot explain the difference in reactivity. Furthermore, application of this approach to models of MTO and STO peptides at different stages of the reaction sequence (pre-farnesylation, post-farnesylation with the farnesyl group in the exit groove) was not able to account for this difference as well. Hence, rather than binding affinity, a parameter related to the dynamics of product dissociation might dictate turnover. We therefore conclude that a dynamical approach, such as molecular dynamics, will be required to explain the mechanism that distinguishes STO from MTO peptides.
Past in vitro peptide farnesylation experiments with FTase have measured kcat/KMpeptide under MTO conditions and kfarn rate constants under STO conditions [17]. The estimated reactivity of MTO and STO peptides (see Methods) measured in this work falls within the range of previously measured activity [17]. Therefore, these peptides have comparable reactivity to other substrates, including peptides that correspond to proteins that are farnesylated in vivo.
Measured under MTO conditions, the kinetic parameter kcat/KMpeptide is termed the specificity constant and best reflects the reactivity of an enzyme in the presence of multiple substrates, as observed in vivo [29]. In a cell, the reactivity of a protein substrate with FTase depends on the value of kcat/KMpeptide as well as on the concentration of the substrate within the cytosol. Although a protein substrate with a higher value of kcat/KMpeptide is more likely to be farnesylated in vivo, it is unclear what level of in vitro activity corresponds to a true FTase substrate in vivo. Furthermore, in vivo the optimal levels of farnesylation of a given substrate may vary and a low fraction of modification may still be biologically relevant. Additionally, a substrate must be localized to the proper cellular locale in order for modification to occur and the C-terminus of the protein must be structurally available. Peptide library studies and this work have aided in determining potential FTase substrates and have also identified already known substrates, but more work is needed to characterize the reactivity of these substrates in vivo.
As for the STO-only peptides, these substrates are readily farnesylated but the product does not dissociate rapidly. One possibility is that these proteins function as FTase inhibitors and consequently play a regulatory role within the cell [17]. However, both FPP and peptides have been implicated in catalyzing product dissociation of farnesylated STO peptides [17], [30], [31] and therefore it is possible that other cellular components could activate product dissociation allowing rapid farnesylation of these proteins in vivo. Therefore, competition or synergy among different FTase substrates could play an important functional role for modification and localization of proteins. Improved identification of STO peptides using the structure-based FlexPepBind approach presented here will expand our understanding of regulatory aspects of this reaction. In addition, the overlap in substrate preference of FTase and GGTase-I [3] indicates that modulation of the type of prenyl modification (e.g. changes in relative enzyme availability or magnesium concentration) might be functionally important as well. Our future focus on structure-based characterization of GGTase-I specificity will allow an improved investigation of this regulatory feature, complementary to sequence-based studies [7], [8].
Scanning the human genome for putative FTase targets using our structure-based approach revealed many putative, not yet detected, farnesylated proteins. These new farnesylation substrates may provide novel disease targets for farnesyltransferase inhibitors. Moreover, the prediction that these proteins are farnesylated might shed light on their function. As an example, the putative proteins Q8NA34, A6NHS1, and P0C7P2 (UniProt identifiers [24]) all contain C' sequences strongly predicted to serve as farnesylation targets suggesting that the proteins are membrane localized. Additionally, our method also predicts FTase substrates that have recently been identified from in vivo experiments. For example, Kho et al. used a tagging-via-substrate proteomic approach to discover novel farnesylation targets [32]. They found a total of 18 farnesylated proteins: 13 are well known, and of the remaining 5 our approach predicts 4 to be farnesylated, including one hypothetical protein. Furthermore, it was recently found that pathogens can hijack the host farnesylation machinery to their own advantage, for example, anchoring effector proteins to the membrane of Legionella-containing vacuoles [33], [34], [35]. Thus, in addition to the identification of putative new farnesylation targets in the human genome, FlexPepBind can be used to scan pathogen genomes for farnesylation as well.
13/16 motifs derived from human proteins tested for in vitro farnesylation indeed undergo the reaction. Will this also happen in vivo? In the following we compile additional available details on these targets that might help answer this question.
One way to assess the in vivo relevance of the observed in vitro ability to undergo farnesylation of the C-terminus of a protein is to look for homologous proteins that also undergo farnesylation. Such information can easily be retrieved from PRENbase [8]. A search in this database revealed that Kinesin-like protein KIF21B variant (Q2UVF0; CFLT) maps to a cluster of 9 highly similar eukaryotic sequences (E-val<e-20) that are all predicted to undergo farnesylation by PrePS. Similarly, Ankyrin repeat and BTB/POZ domain-containing protein BTBD11 (A6QL63-3; CWLS) maps to a cluster of 25 sequences of related proteins in PRENbase. Zinc finger protein 64 homolog (Q9NTW7-3; CYVA) also contains a number of conserved homologs in PRENbase, however in this specific isoform the target cysteine is part of the Zinc-finger structural motif, and therefore it might not readily be farnesylated.
Another interesting putative farnesylation target that we have identified is the short isoform of Intersectin-2 protein (Q9NZM3-3; CCLS). This protein is involved in clathrin-mediated endocytosis [36], [37], and farnesylation could be a mechanism for regulation and localization to the membrane, similar to the prenylation of Rho GTPases for endocytosis [38]. In particular, the long isoform of intersectin-2 contains additional domains [39], including a PH domain known to bind phosphoinositides [40], and a C2 domain known to be involved in Ca-dependent and independent binding of phospholipids [41]. Consequently, in the short isoform that lacks these domains, farnesylation might indeed be used as an alternative way to achieve membrane proximity and attachment. While the localization of some Rho GAP proteins (e.g. p190 [42]) is regulated by phosphorylation, the short isoform of Rho GTPase-activating protein (GAP) 19 (Q14CB8-5; CSLI) exposes a new C' motif that may target it to the membrane (while keeping the Rho GAP domain intact). The same goes for MAPKAP1 isoform 6 (Q9BPZ7-6; CKLA), a subunit of mTORC2. While the full length protein was shown to contain a functional PH and Ras binding domains [43], the truncated isoform reveals a C' putative farnesylation motif instead. Thus, for all but three MTO sequences we could gather additional information that supports actual in vivo farnesylation. We further discuss alternative splicing as a regulatory mechanism below.
Four motifs were found to undergo in vitro farnesylation under STO conditions. The Homeobox protein ESX1 (Q8N693; CPFF) is cleaved into an N' and C' domain; while the N' enters the nucleus, the C' domain is localized to the cytoplasm where it inhibits cyclin degradation[44]. A search for homologues in PRENbase produced a cluster with 2 sequences predicted to undergo farnesylation by PrePS. While the latter could support actual farnesylation of this protein, in this case this modification would serve for purposes other than membrane association, such as the interaction with new partners [5]. Isoform 2 of the integral membrane protein solute carrier family 7 member 13 (Q8TCU3-2; CHFH) is missing an intracellular domain, and therefore places its C' in proximity to the membrane. Here farnesylation could play a role in targeting this transmembrane protein to a specific membrane compartment [45], resulting in different membrane distributions for alternative spliced isoforms. Decaprenyl-diphosphate synthase subunit 1 isoform (Q5T2R2-2; CTTE) is a nuclear encoded mitochondrial protein. If indeed farnesylated, this would be a first example where an isoform of a mitochondrial protein is farnesylated in the cytosol. Finally, the proton-coupled amino acid transporter 1 (Q7Z2H8; CAFI) is likely not a farnesylation target, since mutation of the target cysteine to alanine did not affect its function [46]. As discussed above, the biological role of farnesylation under STO conditions is not yet clear; furthermore, if these proteins are farnesylated in vivo, the function is likely more complex than localization to the membrane.
For the three motifs that were not farnesylated under in vitro conditions, additional information about the cognate proteins indeed suggests that the C-terminal cysteines are likely not farnesylated in vivo. The target cysteines of Growth/differentiation factor 15 (Q99988; CHCI) and the extracellular C-type lectin domain family 2 member D isoform (Q9UHP7-3; CLFE) are part of a conserved disulfide bridge and therefore most likely not farnesylated in vivo.
In this study, we chose peptide motifs for in vitro experimental characterization based on their predicted ability to bind FTase and their novelty (i.e. not predicted by PrePS, and not yet experimentally tested). While our post-hoc literature analysis reinforces some of the predictions, other targets will apparently undergo farnesylation only in vitro. The latter represent an interesting set of proteins that allow the investigation of additional factors that regulate the actual farnesylation in vivo, and that therefore distinguish between the ability of a protein to undergo farnesylation in vitro and in vivo. In any case, future in vivo validation is required for all putative targets to unequivocally define their functional importance in the cell.
Approximately half of the proteins strongly predicted by FlexPepBind to undergo farnesylation (86/167) appear in alternative splicing isoforms (according to Swissprot [24]; the actual number of isoforms is expected to be higher, as more experimental data accumulate from large scale sequencing efforts). Among these 86 proteins, most (61) contain the Cxxx motif only in some of the isoforms. This may present a second layer of regulation for the localization of such proteins, in which a protein can reside in different cellular compartments as a function of the isoform expressed at a given time or a given tissue and therefore perform different functions. This form of regulation may be a consequence of the irreversible nature of farnesylation. On the other hand, farnesylation can be maintained despite alternative splicing. For example, in Rab28 the two reported isoforms (hRab28S, hRab28L) differ only by a 95-bp insertion within the coding region [47]. This insertion generates two alternative sequences in the 30 C-terminal amino acids, which strikingly both contain a high-affinity farnesylation motif (CSVQ – L isoform, CAVQ – S isoform) at the C-terminus. This is similar to the case of KRas that also expresses as two splice variants with strong farnesylation motifs (CIIM - 2A isoform, CVIM - 2B isoform) and different upstream sequences. In this case one upstream sequence harbors an additional palmitoylation site, and may thus lead to different distribution in the membrane [48].
FlexPepBind is a framework for designing peptides that bind to a given protein, as well as for the prediction of peptide binding specificity. It is based on our previously developed modeling protocol FlexPepDock for peptide-protein structures [20]. Inclusion of constraints derived from known structures with bound peptides allows for the definition of backbone flexibility that is appropriate for the specific system of interest, and optimization of the energy function is based on a given set of binding and non-binding peptides.
How much conformational freedom should be given to the peptide in order to sample the correct conformation, without introducing too much noise? What is the best score for discrimination of active and non-active peptides? While Grigoryan et al. were able to design peptides that bind to specific members of the bZip family [49], Goldschmidt et al. identified fibril-forming peptides on a large scale [50], and Kota et al. defined a binding motif for type I HSP40 peptide substrates [51] using fixed backbone conformations, the incorporation of backbone conformational flexibility has generally improved computer-aided design of functional protein interactions, as well as structure-based prediction of peptide-protein and protein-protein interaction specificity [52]. In particular, a range of backbone conformations created by the backrub method [53] improved computational sequence recovery of experimental phage display results on human growth hormone [54], and variation along normal modes allowed improved optimization of binding between the anti-apoptotic protein BCL-xl and BH3 helical ligands [55]. Modeling of the structure of HIV protease – peptide targets using a flexible docking protocol allowed the distinction between peptides that are cleaved from those that are not, opening new avenues towards the design of HIV protease inhibitors [56].
In our modeling study of FTase binding peptides, side-chain repacking alone that restricts sampling to a discrete rotameric representation results in a low AUC value of 0.606 over the training set. Simple minimization that allows for very subtle backbone, side chain, and rigid-body adjustments relieves clashes that cannot be resolved with a simple rotameric side-chain search, and indeed improves performance significantly (AUC = 0.875). Much more extensive sampling with Rosetta FlexPepDock [20] produces even better AUC values (up to 0.94). Therefore, the more we sample, the better we perform. On the other hand, restricted sampling can also improve performance: the incorporation of conserved structural constraints into the simulations, as well as the inclusion of the FPP farnesyl analog, significantly improves the identification of farnesylation targets. The performance of different sampling and scoring schemes is summarized in Table S1.
Incorporation of additional FTase backbone conformations from additional FTase-substrate complex structures could enhance the predictions. To examine this, we evaluated the FlexPepBind protocol with two additional backbone templates, and assessed for each the performance on the training set. Using PDBs 1tn7 [9] and 2h6f [57], we achieve comparable and slightly worse AUC values of 0.85 and 0.75, respectively. Combining the scores based on 1tn6 and 1tn7 gave a marginally better performance (AUC = 0.88) and could indeed represent an avenue for future improvement of the protocol.
In addition to sampling, calibration of the energy function can also improve the prediction of binding peptides. In a study on PDZ-peptide interactions, Kaufmann et al. optimized the Rosetta energy function on 28 peptide interactions with PDZ domain 3 of PSD-95 for binding prediction. The resulting interface energy using an increased contribution of the hydrogen bond term produces a ROC plot with an AUC value of 0.78 on a general set of 144 peptide-PDZ interactions [58].
In our study we find that scoring with the Rosetta energy provided by the peptide provides the best results for the discrimination of active and non-active peptides. This energy includes the internal peptide energy as well as the interface energy, minus a reference energy term that had been previously introduced to optimize sequence recovery in the design of globular proteins [46]. De-facto, removal of this term favors (in decreasing order) C,W,F,H,Y,V,I,A,P and disfavors R,Q,N,E,D,K,S,M,T,G,L. Consequently, without this term, hydrophobic residues will be favored, and performance on the training set improved (probably due to the significant proportion of hydrophobic residues in this set, see Figure 4B). Inferior results are obtained using the Rosetta energy score provided by the interface, as well as the total protein structure. In addition, we would like to note that when using FlexPepDock for sampling, averaging the scores of the best 10 models always gives better results than using merely the top-scoring model (see Table S1 for the performance of different scoring functions).
While the FlexPepDock based protocol gives better results, it is computationally expensive, however, and would impede large-scale characterization (even though it may be the method of choice to make specific decisions once a threshold has been determined from the training set). We found that simple minimization worked well for FTase specificity prediction (and is about 500 times faster than the full FlexPepDock-based protocol). This is due to the restricted nature of the binding - three very strong limitations constrain the peptide backbone orientation. Other systems will probably benefit from increased modeling of backbone flexibility.
In summary, proper calibration of the energy function together with conformational sampling provides efficient structure-based characterization of peptide-protein interactions. It has been estimated that up to 40% of the cellular protein-protein interaction network is mediated by peptide-protein interactions [59]. FlexPepBind is generic in the sense that very little prior knowledge is needed in order to predict the specificity profile for a certain peptide-protein interaction. Given a structural template and a small set of known examples, prediction can be made to identify additional putative targets. We therefore anticipate that this approach can be expanded to a large scale by adapting it to additional peptide-protein interaction motifs in the cellular peptide-protein interaction network.
Human SwissProt [24] was downloaded from IPI [61] (newest version available as of 19.01.10), and was scanned for sequences containing a Cxxx regular motif as the last 4 residues in the protein sequence.
Gene Ontology [25] terms were associated with each of the 167 identified candidates for farnesylation (see Results). Enrichment for different cellular compartments, evaluated using DAVID [62], extracted a subset of 93 proteins that are enriched with 18 GO cellular compartment terms, most of them related to the membrane (see Figure S2).
We used the PrePS web-server [7] to obtain sequence-based predictions on our set of 167 selected proteins. For each protein suggested by our protocol to undergo farnesylation, we calculated its prenylation ability using 30 C-terminal residues as input to the server.
Farnesylation screens were performed using radioactivity assays. Different conditions were used to assess the ability of Cxxx sequences to undergo farnesylation under multiple turnover (MTO) and single turnover (STO) conditions, as detailed below. Peptides that do not undergo farnesylation under either of these conditions were defined as NON (see Hougland et al. [17] for more details).
|
10.1371/journal.pntd.0003186 | Sources and Distribution of Surface Water Fecal Contamination and Prevalence of Schistosomiasis in a Brazilian Village | The relationship between poor sanitation and the parasitic infection schistosomiasis is well-known, but still rarely investigated directly and quantitatively. In a Brazilian village we correlated the spatial concentration of human fecal contamination of its main river and the prevalence of schistosomiasis.
We validated three bacterial markers of contamination in this population by high throughput sequencing of the 16S rRNA gene and qPCR of feces from local residents. The qPCR of genetic markers from the 16S rRNA gene of Bacteroides-Prevotella group, Bacteroides HF8 cluster, and Lachnospiraceae Lachno2 cluster as well as sequencing was performed on georeferenced samples of river water. Ninety-six percent of residents were examined for schistosomiasis.
Sequence of 16S rRNA DNA from stool samples validated the relative human specificity of the HF8 and Lachno 2 fecal indicators compared to animals. The concentration of fecal contamination increased markedly along the river as it passed an increasing proportion of the population on its way downstream as did the sequence reads from bacterial families associated with human feces. Lachnospiraceae provided the most robust signal of human fecal contamination. The prevalence of schistosomiasis likewise increased downstream. Using a linear regression model, a significant correlation was demonstrated between the prevalence of S. mansoni infection and local concentration of human fecal contamination based on the Lachnospiraceae Lachno2 cluster (r2 0.53) as compared to the correlation with the general fecal marker E. coli (r2 0.28).
Fecal contamination in rivers has a downstream cumulative effect. The transmission of schistosomiasis correlates with very local factors probably resulting from the distribution of human fecal contamination, the limited movement of snails, and the frequency of water contact near the home. In endemic regions, the combined use of human associated bacterial markers and GIS analysis can quantitatively identify areas with risk for schistosomiasis as well as assess the efficacy of sanitation and environmental interventions for prevention.
| People tend to live close to natural water bodies, and often these water bodies are used as waste disposal in many regions of the world. The consequences of this are often studied with regard to bacterial and viral infections, but rarely for parasitic infections. In this study the authors examined a rural community settled along a river in Brazil, and found that the concentration of fecal bacteria in water accumulates as the river runs downstream. Molecular methods were able to show that most of these fecal bacteria were of human origin rather than from local livestock or other domestic animals. To assess the impact of fecal contamination of surface waters, the authors investigated its association with schistosomiasis, a parasitic infection transmitted by snails exposed to water contaminated by human feces. Similar to the distribution of fecal contamination, the proportion of people with schistosomiasis was higher in areas located downstream. A model combining concentration of human fecal bacteria in water and Geographic Information Systems (GIS) analysis of schistosomiasis prevalence showed that areas of increased concentration of human feces correlated with areas in the village at higher risk for schistosomiasis. This research provides insight into the dynamics of fecal contamination of rivers and its spatial impact on a human parasitic disease.
| The culture of common fecal organisms such as coliforms and enterococci from surface waters has historically been used as a proxy for the risk of infection with viral, bacterial, and parasitic pathogens [1]. It forms the standard for the United States Environmental Protection Agency's criteria for water quality [2]. Despite the well-studied association between fecal contamination of water and acute enteric and skin diseases [2], [3], a correlation between these bacterial proxies and specific disease causing organisms has been difficult to demonstrate in the absence of a point-source such as sewage outflows [4]. Known limitations that could explain this weak association include the short survival of some fecal indicator organisms in water [5], their presence in environmental sources including soils and sediments [6], [7], contributions from non-human sources, and low sensitivity of detection methods for some pathogens [6], [7]. The short incubation and shedding periods of these infections may also cause the pathogenic organism to no longer be present in sampled water by the time an investigation is undertaken. Molecular methods have been developed to address some of these weaknesses. Most current approaches involve PCR amplification of bacterial rDNA taken directly from specific hosts or sources of fecal contamination without prior culture. Fecal anaerobic bacteria are some of the most promising alternative indicators to Escherichia coli and enterococci. They are more abundant than coliforms, they do not multiply in the water column, and some sub-species or strains are more specific for host sources [8]–[12]. One of the most commonly employed and reliable indicators for human fecal pollution are human Bacteroides originally described as the HF8 cluster [13], [14]. A concern remains as to the distribution of these markers originally developed in the US and other developed countries and whether they can be associated with disease risk in other parts of the world [15].
In contrast to most waterborne bacterial and viral infections, schistosomiasis is a chronic parasitic infection that results from skin contact with water as opposed to ingestion. It is a global disease that is transmitted in 78 countries with 240 million people infected [16]. In Brazil, where it is the second most common cause of morbidity and death due to parasitic infection [17], Schistosoma mansoni is the only human species transmitted. Its transmission in common with other waterborne diseases is dependent on human fecal contamination of fresh water. Thus, in Brazil at a national scale the distribution of schistosomiasis maps to areas with the poorest level of sanitation [18].The parasite is able to establish a long-term infection (5–40 years) that produces hundreds to thousands of eggs per day, most of which will be eliminated in the feces [19]. An obligate step in transmission is development in a host snail, so that there can be no direct temporal connection between fecal contamination and human infection. Snails movements, however, are extremely restricted geographically [20] and in this way past contamination and infection events are registered locally in chronic human infections.
Given the complex life cycle of this parasite and its long-term survival in a community, bacterial indicators that track human sources of fecal contamination in water may contribute much to our understanding of the transmission dynamics of the parasite. Since snail infection with S. mansoni is dependent on human fecal contamination of surface waters, the probability of snail infection and differences in the spatial distribution of human schistosomiasis is likely to correlate with differences in the concentration and distribution of this contamination. We tested this in one small community in Northeastern Brazil.
The study was conducted in a Brazilian village. Demographic and schistosomiasis prevalence data was obtained from a population and fecal survey whose results have been previously published [21], [22]. DNA was extracted from stool samples of individuals who tested positive for schistosomiasis. Animal fecal samples were also collected for DNA extraction. We validated use of three host-indicative fecal bacterial markers in this population by high throughput sequencing of 16S rRNA gene and qPCR of human and animal feces. Water samples were collected along the village's main river. Fecal water contamination was assessed by traditional culture as well as by qPCR of host-indicative fecal markers. Sequence data from the microbial communities found in river water was used to compare the relative abundance of >20 bacterial families. Finally, by use of a linear regression model, we tested the correlation of household proximity to concentration of human fecal water contamination with prevalence of schistosomiasis.
The Committee on Ethics in Research of the Oswaldo Cruz Foundation of Salvador, Bahia, the Brazilian National Committee on Ethics in Research, and the Institutional Review Board for Human Investigation of University Hospitals Case Medical Centre, Cleveland, Ohio approved the study design. Study participants provided written informed consent. Animal owners provided permission to collect the stool samples of their animals.
The village of Jenipapo in the state of Bahia, Brazil was selected for study because of its high prevalence of S. mansoni infection, the geographic distribution of its human population around surface waters, and its relative isolation from other settlements. The village is split north and south by the Jiquiriçá River and a two-lane highway. The Brejões River descends from the north, borders part of the village on the west, and enters the Jiquiriçá River at approximately the village midpoint (Fig. 1a). Within Jenipapo, the Jiquiriçá River measures 5–10 meters across and less than 1 m deep, with areas of bare rock as well as thick aquatic vegetation. The Brejões is narrower and shallower, but still perennial. Most houses are located within 20 meters of these rivers. Topographically, the region is a narrow river valley with approximately equal elevations on both sides. Commercial activity is primarily devoted to raising livestock along with planting cassava, beans, and bananas. Demographic data and prevalence of schistosomiasis was obtained from interviews and a fecal survey of all residents of the village in 2009. The description of the community has been published previously [21], [22].
The location of each home and human water contact sites in the community was registered with a hand-held Trimble/Nomad GPS unit (Model 65220-11). The course of the river within the village was surveyed by walking along one bank. Data were imported into ESRI ArcGIS 10.0 (Redlands, CA) for mapping and analyses. Kernel density estimation was used to assess and display the spatial density of the human population, schistosome infection, and river use for sewage disposal. The Moran's I statistic was calculated using the Spatial Autocorrelation Tool in ArcGIS to assess spatial clustering.
The collection and processing of human fecal samples of the residents of Jenipapo was described previously [22]. Briefly, all inhabitants of the community >1 year of age were asked to enroll in the study, and in addition to answering a questionnaire, they provided a stool on 3 different days. A single slide was prepared by the Kato-Katz method [23]. Microscopists trained for parasitological studies read one slide per sample, and from this the number of eggs per gram of stool (epg) in each sample was determined. Participants with one or more egg-positive stools were treated with praziquantel at dosages recommended by Brazilian Ministry of Health.
One ml of each sample was placed in culture media (3M Petrifilm E. coli/Coliform Count Plate, 3M, Saint Paul, MN) for 24 h at 37°C and counted for colony forming units (CFUs) of total coliforms and E. coli.
Six different qPCR assays were used in extracted DNA of all collected water samples for identification and quantification of fecal bacteria indicative of human or ruminant sources (Table 1). All qPCR assays were amplified in 25 µl reactions using 12.5 µl TaqMan Master mix, 1.0 µl 25 µM primer mixtures, 1.0 µl 2 µM probe mixtures, 5.5 µl water and 5.0 µl of DNA. Assays were carried out as previously described in the referenced literature in Table 1. All assays were run in duplicate.
Deep sequencing using the Illumina MiSeq platform was carried out at the Josephine Bay Paul Center of the Marine Biological Laboratory. A comprehensive microbial community profile was generated for five river samples, ten human fecal samples, and all collected animal fecal samples. The V6 hypervariable regions of the 16S rRNA gene were amplified in each of the samples using previously described primers and protocols [25]. Sequences were trimmed, controlled for low quality and contaminated reads, and then aligned. Nearly 27 million bacterial sequence reads were generated (∼1 million reads per sample). The sequence data were further processed and stored in the Visualization and Analysis of Microbial Population Structures (VAMPS) database (http://vamps.mbl.edu) [26]. Taxonomic assignments were made for all sequences using Global Alignment for Sequence Taxonomy (GAST) [27]. Further analysis of sequence data is reported in [28] and sequence data is available in the National Center for Biotechnology Information (NCBI) Short Read Archive under the accession number SRP041262.
To assess the proportion of bacterial community members that are potentially amplified by the human-indicative fecal indicator assays, a BLAST search was performed against the Illumina sequence data sets with the HF8 and Lachno2 primers [29]. Since the primers for the human-indicative assays are in regions of the 16S rRNA gene different from or only encompasses a larger region that the V6 sequences, the HF8 and Lachno2 primers were BLASTed against the complete reference sequences dataset that corresponded to the shorter V6 sequences. The V6 sequencing reads, each a proxy for a bacterial community member, were then binned within the HF8 cluster or Lachno2 cluster if their corresponding reference sequences contained both the forward and reverse primers and the probe sequences for the assays.
To examine the association between proximity to fecally contaminated water and schistosomiasis, a linear regression model was created with SPSS version 19 and ArcGIS 10.1. For the model, we made the following simplifying assumptions: 1) infection occurs at the common water contact sites, 2) probability of infection depends only on proximity of place of residence to a water contact site, 3) distribution of snails along the river is homogeneous, and 4) prevalence of snail infection is proportional to degree of human fecal contamination in water.
In order to associate spatial distribution and the prevalence of S. mansoni, the residential area of Jenipapo was mapped as a grid of 200 m2 blocks. The density of human population and density of number of cases of schistosomiasis per block was calculated using the point density tool in ArcGIS. The prevalence of schistosomiasis within each block was obtained by calculating the ratio of these two values using the Map Algebra tool. All houses within a 200 m2 block along the river were assigned the mean prevalence for that block. This spatial prevalence for each household then comprised the dependent variable in the linear regression. The independent variable was the risk of exposure to fecal contamination. To assign a value for fecal exposure for each household, spatial interpolation of two fecal marker DNA concentrations measured from the eight-water sample sites was performed using Inverse Distance Weighting (IDW). The village was divided into a two-dimensional grid of cells whose values were a function of their distance from a water contact site and the concentration of a fecal contamination marker at that site. A power of 2 was determined to be the best value for the weighting exponent by distance with a cell size of 20 m. Since human fecal contamination of water is necessary for transmission of S. mansoni, we hypothesized that a human-indicative fecal marker (Lachno2) would be a better predictor of schistosome infection than a general fecal marker, i.e. E. coli. Consequently, E. coli and Lachno2 estimated concentrations at each resident's location were extracted from the IDW generated surface to obtain an E. coli-IDW and Lachno2-IDW value for each home. Water samples were not taken from the Brejões River, thus, the section of the community bordering the Brejões was not included in the analysis. The relationship of fecal contamination to prevalence of schistosomiasis was then assessed by standard linear regression. The model significance was determined by bootstrapping with 1000 resamples at the household level.
In 2009, Jenipapo consisted of 128 houses with 482 residents. Twenty-three residents had no house assigned, hence were not included in the analysis (Table 2). More than 98% of residents had tap water and an indoor flush toilet. There was access to adequate sanitation for 201 (43.8%) via home septic tanks, while sewage drained directly into the river for the remaining 258 (56.2%). Schistosomiasis was found in 209 individuals (45.5%) by examination of 3 stools collected on different days [22]. The geometric mean of intensity (57 epg) indicates that infections are generally light and comparable to other studies in Brazil [30]. Ten percent of the infections were heavy (>400 epg).
The Jiquiriçá River flows from west to east, and its course measures 1542 meters from the upstream sampling site to the downstream site. There is one formal bridge across the river at the point where the Brejões River enters the Jiquiriçá. Human water contact sites were primarily used to cross the river as well as for bathing, washing clothes, or fishing (Fig. 1a). The majority of houses are located on the south bank. Kernel density estimation shows clustering of human population density at both ends of the village along the south side of the Jiquiriçá River, but the greatest density clustered along the Brejões (Fig. 1b). The distribution of houses with sewage draining directly into the river (Fig. 1c), however, were not clustered based on Moran's I statistic, which fell within the 95% confidence interval of the null hypothesis (random spatial distribution) as indicated by a low z-score (Moran's index = −0.013, p = 0.847, z = −0.19). In contrast, for the prevalence of schistosomiasis, kernel density estimation shows clusters located along the Brejões River and for people living along the most downstream segment of the village (Fig. 1d). The positive Moran's index with a high z-score and low p value indicated that the distribution of schistosomiasis was not random (Moran's index = 0.042, p = 0.006, z = 2.74).
The 16S rRNA sequencing reads of extracted DNA from fecal samples of ten humans and all collected animal fecal samples were normalized against their maximum number of reads and queried for the human-indicative Bacteroides HF8 and Lachno2 clusters. Overall, humans had considerably lower amounts of Bacteroides in relation to Lachnospiraceae or more specifically, Blautia (one genera within Lachnospiraceae from which the Lachno2 assay was designed). Despite the low amount of overall Bacteroides in humans, the HF8 sequence represented 28% of all Bacteroides sequences. Overall, the proportion of sequence reads matching the HF8 cluster in humans was 10-fold higher than for pigs and dogs, and 100-fold higher than for horses and cows. The Lachno2 cluster showed even higher specificity with the proportion of reads in humans ∼100-fold higher than three animal sources, but ∼10-fold higher than for horses (Table 3).
Using qPCR, the concentration of the Bacteroides-Prevotella group was at its lowest (<2.7×105 copies/100 ml) from site S1, located upstream of the first house of the village, through site S3 (Fig. 2). There was a steep increase at S4 to 4.8×105 copies/100 ml. The highest concentration was found at S5 (5.4×105 copies/100 ml), where the Brejões River joins the Jiquiriçá. Its concentration then decreased gradually and by S8, located downstream of the last house, the concentration of this general fecal marker had returned to a value similar to S1 (2.7×105).
The human-indicative markers (HF8 and Lachno2) followed a similar distribution, however, concentrations increased one site further downstream compared to the Bacteroides-Prevotella group marker. The HF8 marker was undetectable until site S5, at which point it also reached its peak (0.9×104 copies/100 ml), followed by a gradual decline. Lachno2 was detectable in minimal quantities at sites S1 to S4 (maximum concentration 684 copies/100 ml), and also had a marked increase by site S5. The peak Lachno2 concentration was at site S7 (1.6×104 copies/100 ml), which is the last site downstream in Jenipapo that humans utilize to cross the river, and declined by S8. The ruminant-indicative marker was undetectable until S5 and remained in low concentrations without significant variation between sites thereafter. The E. coli marker showed a smaller degree of increase after S5. By contrast, the source of drinking water located 4.8 km north of the village had no copies of the HF8 human-indicative marker; other assays were not performed on this sample. Colony counts for coliforms, and less so for E. coli, also increased as the river moved down stream and declined sharply past the last house in the village (Fig. 3).
Sequence data from the microbial communities found in river water was used to compare the relative abundance of >20 bacterial families. Consistent with the qPCR results, the proportion of Prevotellaceae and Lachnospiraceae increased significantly in the downstream portion of the village (Fig. 4). Ruminococcaceae and Enterobacteriaceae, two other families associated with fecal communities, also increased. These combined fecal families increased their representation from ∼3% to ∼9% of all bacterial community between upstream to downstream sites. Families associated with sewage- contaminated water - Moraxellaceae and Aeromonadaceae, specifically Acinetobacter spp. and Aeromonas spp. [31]–[34] - also increased at sites six and seven. Comamonadaceae, a bacteria common to the environment and freshwater [35], was the most abundant family on average, accounting for over 40% of the microbial community populations at each sampling site. Bacteroidaceae, which includes the genera Bacteroides, were in low abundance and are not represented in Figure 4.
The 200 m2 schistosomiasis prevalence grid for Jenipapo produced 7 blocks (Fig. 5). Each house was assigned a value for exposure to fecal contamination based on proximity to a water sample site and the fecal marker concentration at site. The relationship of risk for infection with S. mansoni to the concentration and proximity to fecal contamination was modeled and tested statistically using the data from Jenipapo. Linear regression of prevalence of schistosome infection against fecal contamination yielded an r2 of 0.28 for the E. coli-IDW value (two-tailed p<0.001, 95% CI 0.22–0.35) and 0.53 for the Lachno2-IDW value (two-tailed p<0.001, 95% CI 0.48–0.58). These results can be interpreted as local concentration of human fecal contamination explaining over 50% of the variance in risk for schistosomiasis.
Although the village of Jenipapo is small, it is typical of many villages of Latin America. It also shares a pattern of development common with larger communities and even the great metropolises of Brazil. The village grew up along the two rivers that meet at its center, and most homes border these rivers in part to have access to a ready form of sewage removal. The community's drinking water supply is 4.8 km away where a dammed stream forms a small reservoir. Jenipapo's geometry is a simple, mostly linear distribution of residences and water contact sites, and this made it ideal for studying the dynamics of fecal contamination and its relationship to acquisition of schistosomiasis.
Putting the degree of fecal contamination of the Jiquiriçá River within Jenipapo in context, the geometric mean CFUs for E. coli (113 CFU/100 ml) was at the upper limit of the EPA's 2012 Recreational Water Quality Criteria value of 100 CFU/100 ml [2]. This level of contamination was estimated to result in 32 gastrointestinal illnesses per 1000 primary recreational contacts.
We were further able to identify human waste as the major contributor to this contamination. We validated both the HF8 and Lachno2 genetic markers as human-indicative by directly assaying the resident population. Interestingly, both human-indicative markers were identified from humans in the US, but were also significantly associated with humans in Brazil. The frequency of members of the Prevotella complex were higher in this rural Brazilian population than in communities in countries like the US and Italy where fat and protein form more of the diet than cereals, with Bacteroides a minor component of the human samples. In all human communities, including hunter gatherers, the Lachnospriaceae group, however, is more similarly represented [36]. In comparison with other human indicative markers, Lachno2 showed a high signal in the water sample and all human feces, but near absence in cows, the other major animal contributing to fecal contamination of the river. These markers indicated that human waste was the major contributor of fecal contamination in this section of river. Overall, human-indicative fecal indicators contribute important quantitative information on water quality that could be used for surveillance to gauge specific sanitation interventions.
The nearest community to Jenipapo is 8 km upstream with a population of 353 and similar level of sanitation, and there are few intervening houses, but many areas of pasture. Twelve km further upstream there is a town of 12000. Despite nearby populations, quantitative tracking of human fecal contamination in this study suggests a predominance of local effects. The qPCR markers for human and other fecal contamination, as well as coliform colony counts, are very low at the entrance to the village and significantly increase as the river continues downstream. Inflow for the village has significant levels of the Bacteroides-Prevotella group, but is very low for human fecal contamination indicating that most influence from communities upstream has dissipated. We presume this is not the result of the HF8 marker being sensitive to environmental degradation, since experimentally the duration of signals from Bacteroides ranges from days to several weeks [37]. In addition, the other marker of human fecal contamination (Lachno2) shows a similar pattern. Within Jenipapo, the entry of sewage is not clustered to one area of the community, and we noted the concentration of contamination is cumulative as the river moves downstream through the village.
The analysis of bacterial communities was based on number of sequence reads and is consistent with the qPCR genetic marker data. The study is limited in the relatively small number of samples taken, sampling only ∼50 m beyond the community's houses and a lack of household water samples. Also the human and animal samples were handled differently, but the distribution of bacterial families is consistent with other studies of the human gut biome [38]. A major strength of the study is the use of markers highly informative for the presence of human feces. Traditional indicators, such as E. coli culture and PCR, are able to demonstrate some of the same distribution pattern as HF8 and Lachno2, but the better model fit using Lachno2 demonstrates that higher tier assessments like qPCR for human indicative markers may provide better linkages between disease and human sources.
Human fecal contamination of water and the presence of snails are prerequisites for transmission of schistosomiasis. Snails are known to have a limited range of movement [39]. Proximity to water bodies where there are infected snails is a known risk factor for schistosomiasis [40], [41]. However, all inhabitants in Jenipapo are essentially equidistant from the river, and finding and determining which snails are infected can be laborious. In this study we show that, in a village with high prevalence of schistosomiasis, the risk of acquiring the infection is driven not only by proximity to surface water but also by its degree of human fecal contamination. The model explained a large amount of variation without including data on snail populations. We also observed that parasite populations were genetically more similar among infected members of the same household compared to parasite populations of all infected individuals in the village [21], which further supports the local, household level of acquisition. The variation not explained by our model was likely due to violations of our simplifying assumptions. Snails are not likely to be evenly distributed, and infection risk is influenced by more than distance to a contact site (age, type of activity, etc.). Some infection occurs outside of contact sites or not at the nearest contact site.
Although the human population disperses widely over this area, the local opportunities for exposure near the home may dominate the infection risk profile. Since awareness of schistosomiasis has been raised in the community and well before the analysis of fecal contamination, we have heard reports that teenage boys now prefer to enter the river upstream of the village. This may be a wise precaution, although the better solution will be to remove the contamination from the river rather than remove the boys and girls.
|
10.1371/journal.ppat.1001175 | Translation Elongation Factor 1A Facilitates the Assembly of the Tombusvirus Replicase and Stimulates Minus-Strand Synthesis | Replication of plus-strand RNA viruses depends on host factors that are recruited into viral replicase complexes. Previous studies showed that eukaryotic translation elongation factor (eEF1A) is one of the resident host proteins in the highly purified tombusvirus replicase complex. Using a random library of eEF1A mutants, we identified one mutant that decreased and three mutants that increased Tomato bushy stunt virus (TBSV) replication in a yeast model host. Additional in vitro assays with whole cell extracts prepared from yeast strains expressing the eEF1A mutants demonstrated several functions for eEF1A in TBSV replication: facilitating the recruitment of the viral RNA template into the replicase complex; the assembly of the viral replicase complex; and enhancement of the minus-strand synthesis by promoting the initiation step. These roles for eEF1A are separate from its canonical role in host and viral protein translation, emphasizing critical functions for this abundant cellular protein during TBSV replication.
| Plus-stranded RNA viruses are important pathogens of plants, animals and humans. They replicate in the infected cells by assembling viral replicase complexes consisting of viral- and host-coded proteins. In this paper, we show that the eukaryotic translation elongation factor (eEF1A), which is one of the resident host proteins in the highly purified tombusvirus replicase complex, is important for Tomato bushy stunt virus (TBSV) replication in a yeast model host. Based on a random library of eEF1A mutants, we identified eEF1A mutants that either decreased or increased TBSV replication. In vitro studies revealed that eEF1A facilitated the recruitment of the viral RNA template for replication and the assembly of the viral replicase complex, as well as eEF1A enhanced viral RNA synthesis in vitro. Altogether, this study demonstrates that eEF1A has several functions during TBSV replication.
| Genome-wide screens for host factors affecting RNA virus infections have led to the identification of several hundreds host proteins in recent years [1], [2], [3], [4], [5], [6], [7]. These works demonstrated complex interactions between the host and plus-stranded (+)RNA viruses, the largest group among viruses. (+)RNA viruses contain relatively small genomes and greatly depend on the resources of the infected hosts in many steps during the infection process. These viruses recruit numerous host proteins to facilitate their replication and spread [8], [9], [10]. Many host RNA-binding proteins have been implicated in replication of (+)RNA viruses, including ribosomal proteins, translation factors and RNA-modifying enzymes [8], [9], [10], [11], [12], [13], [14]. In spite of the extensive effort, the actual function of host factors in (+)RNA virus replication is known only for a small number of host factors [8], [10], [15], [16], [17].
Tomato bushy stunt virus (TBSV) and other tombusviruses are model plant RNA viruses with 4.8 kb genomic (g)RNA coding for two replication proteins, termed p33 and p92pol, and three proteins involved in cell-to-cell movement, encapsidation, and suppression of gene silencing [18], [19]. Yeast (Saccharomyces cerevisiae) expressing p33 and p92pol replication proteins can efficiently replicate a short TBSV-derived replicon (rep)RNA [20], [21]. The tombusviral repRNA plays several functions, including serving as a template for replication and as a platform for the assembly of the viral replicase complex [19], [22], [23]. The viral RNA also participates in RNA recombination [6], [18], [24], which likely plays a major role in virus evolution.
One of the major advantages of studying TBSV replication is the availability of genomic and proteomic datasets on virus-host interactions [4], [5], [6], [7], [10], [15], [25], [26], [27]. For example, systematic genome-wide screens of yeast genes have revealed that TBSV repRNA replication is affected by over 100 different host genes [5], [7]. Additional genome-wide screens with TBSV also identified ∼30 host genes affecting TBSV RNA recombination [4], [6], [28]. The identified host genes code for proteins involved in various cellular processes, such as translation, RNA metabolism, protein modifications and intracellular transport or membrane modifications [3], [5], [7].
Additional global approaches based on the yeast proteome microarray (protein array) have led to the identification of over 100 host proteins that interact with viral RNA or the viral replication proteins [25], [26]. Also, proteomics approaches with the highly purified tombusvirus replicase has determined at least seven proteins in the complex, including the viral p33 and p92pol, the heat shock protein 70 chaperones (Hsp70, Ssa1/2p in yeast), glyceraldehyde-3-phosphate dehydrogenase (GAPDH, encoded by TDH2 and TDH3 in yeast), pyruvate decarboxylase (Pdc1p), Cdc34p ubiquitin conjugating enzyme [14], [26], [27] and eukaryotic translation elongation factor 1A (eEF1A) [25]. The functions of GAPDH and Hsp70 have been studied in some detail [14], [29], [30], [31], but the roles of the other host proteins, such as eEF1A, in the replicase complex are currently undefined.
eEF1A is a highly abundant cellular protein with a role in delivering aminoacyl-tRNA to the elongating ribosome in a GTP-dependent manner. Many additional functions have been ascribed to eEF1A including quality control of newly produced proteins, ubiquitin-dependent protein degradation, and organization of the actin cytoskeleton [32], [33]. Although eEF1A has been shown to be part of replicase complexes of several RNA viruses [16], [34], [35], [36], studies on determining its functions in virus replication are hindered by several major difficulties. These include (i) genetic redundancy: yeast has two eEF1A genes (TEF1 and TEF2), whereas animals and plants have 2–7 genes and several isoforms of eEF1A. (ii) eEF1A provides essential functions for cell viability and mutations could have pleiotropic effects on protein translation, actin bundling and apoptosis. (iii) eEF1A is a very abundant protein that constitutes 1–5% of total cellular proteins, making it difficult to completely remove eEF1A from biochemical assays using cell extracts. (iv) eEF1A is also required for the translation of viral proteins in infected cells, making it difficult to separate its effect on translation versus replication, processes that are interdependent.
The first evidence that translation elongation factors, such as EF-Tu and EF-Ts, play a role in (+)RNA virus replication was obtained with bacteriophage Qbeta [34]. The eukaryotic homolog of EF-Tu, eEF1A was found to bind to many viral RNAs, including the 3′-UTR of Turnip yellow mosaic virus (TYMV) [37], West Nile virus (WNV), Dengue virus, Tobacco mosaic virus (TMV) and Turnip mosaic virus (+)RNA [35], [38], [39], [40]. In addition, eEF1A has also been shown to interact with various viral replication proteins or the replicases, such as the NS5A replication protein of Bovine viral diarrhea virus (BVDV) [41], NS4A of hepatitis C virus (HCV) [42], the TMV replicase [43], and the Gag polyprotein of HIV-1 [44]. It is also part of the replicase complex of vesicular stomatitis virus, a negative-stranded RNA virus [45].
The actual biochemical functions provided by eEF1A for (+)RNA virus replication are currently poorly understood. In case of WNV, eEF1A is co-localized with the WNV replicase in the infected cells and mutations in the WNV (+)RNA within the mapped eEF1A binding site have led to decreased minus-strand synthesis [46]. On the contrary, eEF1A was shown to enhance translation but repressed minus-strand synthesis of TYMV in vitro [37], [47], [48]. Overall, eEF1A likely plays a role in the replication of many RNA viruses. The interactions of eEF1A with viral RNAs and viral replication proteins and its high abundance in cells might facilitate recruitment of eEF1A into virus replication.
eEF1A has been shown to interact with the components of the tombusvirus replicase, including the 3′-UTR of the repRNA, as well as the p33 and p92pol replication proteins [25]. eEF1A is also known to interact with the yeast Tdh2p (GAPDH) [49], which is also a component of the tombusvirus replicase. Overall, the multiple interactions of eEF1A with various components of the tombusvirus replicase could be important for eEF1A to regulate yet unknown functions of the viral replicase complex.
In this paper, we characterized the functions of eEF1A in TBSV replication based on identification of functional eEF1A mutants in yeast as well as using in vitro approaches. The obtained data support the model that eEF1A plays several roles during TBSV replication, including facilitating the assembly of the viral replicase complex. Moreover, using in vitro replication assays, we demonstrate that eEF1A enhances minus-strand synthesis via stimulating the initiation step of the viral RNA-dependent RNA polymerase. Since eEF1A is also associated with several other viral replication proteins or binds to viral RNAs, it is possible that the uncovered functions of eEF1A might be utilized by other RNA viruses during their replication as well.
To determine the functions of eEF1A during tombusvirus replication, we generated ∼6,000 yeast strains expressing eEF1A with random mutations (see Fig. S1A) and tested the level of TBSV repRNA accumulation in a high-throughput assay [50]. In this assay, we used yeast strains, in which the two wt eEF1A genes (TEF1/TEF2) were deleted from the chromosome, while the wt or mutated eEF1A was expressed from plasmids. Importantly, a given eEF1A mutant is the only source of eEF1A in the yeast cells used. Using the high-throughput assay, we identified one yeast strain (N21) expressing an eEF1A mutant that supported reduced TBSV repRNA replication, while the other three strains with eEF1A mutants (named C42, C53 and C62) showed increased level of repRNA accumulation (Fig. 1A and S1B–G). Interestingly, the eEF1A mutants supporting increased steady-state level of repRNA accumulation did not increase the relative level of p33 and p92pol replication proteins (Fig. 1A, bottom panel; S1D–E). Thus, these eEF1A mutants likely affect TBSV replication directly. Accordingly, affinity purification of the solubilized tombusvirus replicase complex from yeast cells, followed by in vitro replicase activity assay revealed that the replicase from C42, C53 and C62 mutant eEF1A-expressing yeast strains had ∼2-fold increased activities when compared with wt eEF1A-expressing yeast strain (Fig. 1B, lanes 1–6 versus 7–8). The amounts of replication protein p33 and the co-purified eEF1A were comparable in the purified replicase samples (Fig. 1B, bottom panel), indicating that the differences in replicase activities in the mutants are likely due to enhanced replicase functions, and not due to altered proteins levels in the replicase complexes. Testing the ability of C42, C53 and C62 mutant eEF1As to bind to the viral RNA or to the p33 and p92pol replication proteins in vitro (Fig. S2B–C) did not reveal significant differences between the mutants and the WT. This further supports that these eEF1A mutations likely increase the function of the viral replicase without altering the protein and RNA components in the replicase.
Placing the identified mutations in the three novel gain-of-function mutants of eEF1A (V301D, L374V/N377K, and F413L; Fig. 1C, indicated with yellow balls), which exhibited increased tombusvirus replication, over the known structure of eEF1A [51] revealed a cluster on one face of eEF1A (namely, the actin bundling domain III), away from the domains known to bind to tRNA and translation factor eEF1Bα. On the other hand, the new reduced function mutant (A76V, Fig. 1C, indicated with green balls) and the previously identified T22S [25], which exhibited decreased tombusvirus replication, showed a distinct and separate localization.
Since eEF1A is part of the tombusvirus replicase complex [25], it is possible that C42, C53 and C62 eEF1A mutants might affect the assembly/activity of the tombusvirus replicase. To test this idea, we prepared cell-free extracts (CFE) from yeast strains expressing selected eEF1A mutants in the absence of the wt copy of eEF1A. These yeast extracts contained comparable amount of total proteins as well as the amounts of eEF1A, ALP, PGK and Hsp70 (Ssa) yeast proteins were comparable (Fig. 2A). The advantage of the CFE extracts is that they can then be programmed with the TBSV (+)repRNA in the presence of purified recombinant p33 and p92pol obtained from E. coli that leads to the in vitro assembly of the viral replicase, followed by a single cycle of complete TBSV replication, resulting in both (−)-stranded repRNA and (+)-stranded progeny [31], [52]. Therefore, this assay can uncouple the translation of the viral proteins from viral replication, which are interdependent during (+)RNA virus infections.
Using CFEs from yeast expressing one of the three mutant eEF1As resulted in ∼3-fold increased TBSV repRNA accumulation when compared with the extract obtained from yeast expressing the wt copy of eEF1A (Fig. 2A, lanes 2–4 versus 5). These data suggest that the viral replicase complex containing the mutant eEF1A can support in vitro TBSV repRNA replication more efficiently than the replicase with the wt eEF1A. In contrast, CFE from N21 yeast supported TBSV repRNA replication to similar extent as the CFE containing wt eEF1A (Fig. 2A, lanes 1 versus 5), indicating that N21 eEF1A mutant can perform the same functions as the wt eEF1A in vitro, when the same amount of p33 and p92pol was provided.
To test if the increased TBSV repRNA replication in vitro was due to enhanced (+) or (−)-strand synthesis, we analyzed the replication products under non-denaturing versus denaturing conditions (Fig. 2B). These experiments showed that the amount of dsRNA [representing the 32P-labeled (−)RNA product hybridized with the (+)RNA] increased ∼3-fold in case of C42, C53 and C62 mutants (lanes 3–8, Fig. 2B) in comparison with the wt (lanes 9–10). The dsRNA nature of these products was confirmed by the ssRNA-specific S1 nuclease digestion assay (Fig. 2C). On the other hand, the ratio of dsRNA and ssRNA did not change in the various CFEs containing the eEF1A mutants or the wt (Fig. 2B). These results are consistent with the model that the replicase complex carrying the eEF1A mutants increased mostly the level of (−)RNA production, which then led to proportionately higher level of (+)RNA progeny.
Cell-fractionation assay, followed by the cell-free TBSV replication assay demonstrated that the soluble fraction from the C42, C53 and C62 mutant yeasts stimulated the in vitro replication of TBSV repRNA by ∼3-fold, while the membrane fraction when derived from C42, C53 and C62 mutant yeasts had a lesser effect (Fig. 2D, lanes 12–14 versus 7–9). These data are in agreement with the expected mostly cytosolic distribution of eEF1A, albeit eEF1A is also present in the membrane fraction in a smaller amount (Fig. 2D, bottom panel).
To test directly if eEF1A could stimulate RNA synthesis by the viral RdRp, we chose the E. coli-expressed recombinant p88pol RdRp protein of Turnip crinkle virus (TCV), which is unlike the E. coli-expressed TBSV or CNV p92pol RdRp, does not need the yeast cell-free extract to be functional in vitro [53], [54]. The template specificity of the recombinant TCV RdRp with TBSV RNAs is similar to the closely-related tombusvirus replicase obtained from yeast or infected plants [21], [54], [55], [56]. However, the recombinant TCV RdRp preparation lacks co-purified eEF1A, unlike the yeast or plant-derived tombusvirus replicase preparations, facilitating studies on the role of eEF1A on the template activity of a viral RdRp. When we added the highly purified wt eEF1A to the RdRp assay containing TCV RdRp protein and a TBSV derived (+)RNA template, which is used by the TCV RdRp in vitro to produce the complementary (−)RNA product (Fig. 3A, lanes 3–4) [55], we observed a ∼6-fold increase in (−)RNA synthesis by the TCV RdRp (lanes 11–12), while, as expected, we did not detect new (+)RNA progeny (not shown). This suggests that eEF1A can greatly stimulate TCV RdRp activity in vitro, confirming a direct role for eEF1A in (−)RNA synthesis by a viral RdRp.
Since it is known that eEF1A can bind to the 3′-UTR of TBSV (+)RNA as well as to the tombusvirus replication proteins [25] and to the TCV RdRp (Fig. S2A), we wanted to test if the above stimulating activity of eEF1A in the in vitro RdRp assay was due to binding of eEF1A to the (+)RNA template and/or to the TCV RdRp protein. Pre-incubation of the purified wt eEF1A with the TCV RdRp prior to the RdRp assay led to a ∼5-fold increase in in vitro (−)RNA synthesis (Fig. 3A, lanes 9–10), while pre-incubation of the purified eEF1A with the TBSV (+)RNA template prior to the RdRp assay led only to a ∼2-fold increase in (−)RNA products (lanes 7–8). Also, pre-incubation of the TCV RdRp with the (+)RNA template prior to the RdRp assay containing purified eEF1A led only to a ∼2-fold increase in (−)RNA synthesis (lanes 5–6), suggesting that eEF1A can stimulate (−)RNA synthesis less efficiently after the formation of the (+)RNA-RdRp complex. Overall, data shown in Fig. 3 imply that eEF1A stimulates (−)RNA synthesis most efficiently when it forms a complex with the viral RdRp prior to binding of the template RNA to the eEF1A-RdRp complex.
To test if eEF1A stimulates the rate of initiation of (−)RNA synthesis, we analyzed the amount of abortive RNA products, which are generated during de novo initiation of RNA synthesis by the TCV RdRp [57]. We found that the amount of the 5–11 nt long abortive RNA products increased by 3.5-fold in the presence of purified eEF1A in the TCV RdRp assay (Fig. 3B, lanes 3–4 versus 1–2). We also tested the RdRp activity in the presence of eEF1A using a (+)RNA template with a mutation opening the closed structure in the promoter region that leads to increased template activity [58]. The mutated template also showed 2-fold increased abortive RNA products in the RdRp assay with eEF1A (Fig. 3B, lanes 5 versus 6). These data strongly support the model that eEF1A stimulates the de novo initiation step in the RdRp assay.
To test if eEF1A stimulates the rate of RNA synthesis in the absence of de novo initiation, we analyzed the amount of 3′-terminal extension (3′-TEX) RNA products, which are generated from an internal primer by the TCV RdRp (Fig. 3D) [55]. Addition of purified eEF1A did not increase the amount of 3′TEX products (lanes 2, 4, 6, Fig. 3D), suggesting that the elongation step during complementary RNA synthesis is not affected by eEF1A. Altogether, the obtained in vitro TCV RdRp data suggest that eEF1A can mostly stimulate the initiation step during de novo viral (−)RNA synthesis.
To further test the function of eEF1A in TBSV replication, we used chemical inhibitors of eEF1A, including Didemnin B (DB) and Gamendazole (GM). DB inhibits the activity of GTP-bound eEF1A during translation by binding to a pocket in eEF1A involved in the interaction with the aminoacylated tRNA and the nucleotide exchange factor eEF1Balpha [59], [60]. GM has been shown to inhibit the actin bundling function, while it does not inhibit protein translation or GTP binding functions of eEF1A [61]. We found that both DB and GM efficiently inhibited TBSV repRNA replication in the in vitro assay with CFE, which contains the endogenous eEF1A (Fig. 4A). Time-course experiments revealed that the inhibition by DB was the most effective when the inhibitor was added at the beginning or during the first 10–15 min of the assay (Fig. 4B, lanes 2–5), while GM inhibited the cell-free replication of TBSV repRNA when added not only at the beginning, but up to 40 min after the start of the assay (lanes 12–17). It is known that the recruitment of the viral RNA and replication proteins as well as the assembly of the viral replicase complex take place during the first 40–60 min in the cell-free assay [31]. Since DB could inhibit translation, we also tested the effect of another translation inhibitor, namely cycloheximide, which did not affect TBSV repRNA replication in our assay (Fig. S3). These data suggest that the inhibition by DB and GM is unlikely through decreased translation in the replication assay. Therefore, the above data are consistent with the model that DB and GM interfere with the assembly of the viral replicase complex in the CFE. Also, GM seems to be a more potent inhibitor of TBSV replication than DB.
To further test if DB and GM can interfere with the assembly of the tombusvirus replicase complex, we performed a two-step in vitro assembly/replication assay, also based on CFE containing endogenous eEF1A [31]. In this assay, first, we only provide ATP and GTP in addition to the replication proteins, the (+)repRNA and CFE, which can support the assembly of the replicase, but cannot perform RNA synthesis due to the lack of CTP and UTP [31]. After 1 hr incubation, once the replicase assembly had taken place, we collected the membrane fraction of the CFE by centrifugation and removed the supernatant containing the unbound p33, p92, repRNA as well as the cytosolic fraction of the CFE. Then we added all four rNTPs (including 32P-labeled UTP) to the membrane fraction of the CFE to allow for RNA synthesis by the pre-assembled replicase complex (second step, Fig. 4C) [31]. Interestingly, adding either DB or GM during the first step resulted in robust inhibition of TBSV repRNA synthesis during the second step of the assay (Fig. 4C, lanes 2–3 versus 1), whereas providing the same amount of DB and GM at the beginning of the second step did not result in inhibition of repRNA replication (lanes 4–6). These data support a model that DB and GM could inhibit the assembly of the tombusvirus replicase complex, but not the RNA synthesis by the already assembled replicase. Similarly, DB and GM failed to inhibit TBSV RNA synthesis in an in vitro assay with a highly purified RdRp from yeast (Fig. S4A).
Since the assembly of the tombusvirus replicase also depends on events prior to the replicase assembly step, such as template RNA binding by the viral replication proteins/host proteins (such as eEF1A), and template recruitment to intracellular membranes [21], [52], we also tested the effect of DB and GM on these processes as well based on purified recombinant eEF1A. We found that GM strongly interfered with the binding of eEF1A to the viral RNA in an EMSA assay (Fig. 4D, lanes 3–6 versus 2), whereas DB did not affect the binding under the assay conditions (lanes 9–12). Since DB binds only weakly to eEF1A in solution, but it binds much more effectively to eEF1A in the presence of GTP and the ribosome [62], we also performed in vitro co-purification experiments. First, 35S-labeled eEF1A was produced in an in vitro translation system (containing ribosome and GTP) and, second, biotin-labeled viral (+)repRNA was added. After short incubation in the absence or presence of various amount of DB, we performed affinity-purification of the viral RNA. Phosphoimaging revealed that eEF1A was co-purified with the viral RNA and the amount of protein co-purified with the viral (+)repRNA was inhibited by increasing amount of DB in the assay (Fig. 4E, compare lane 1 with 2–5). This demonstrated that DB inhibits the binding of eEF1A to the viral repRNA. Moreover, both DB and GM interfered with the recruitment of the viral template RNA to the membrane of the CFE containing endogenous eEF1A (Fig. 4F, lanes 5–8 versus 3–4). On the other hand, DB and GM do not seem to affect the interaction between eEF1A and p33 or p92 replication proteins in vitro (Fig. S4B). Altogether, these data suggest that inhibition of eEF1A function by DB and GM could block several steps during the assembly of the tombusvirus replicase complex, including template binding by eEF1A and viral RNA recruitment into replication.
(+)RNA virus replicases contain viral- and host-coded components, which likely provide many yet undefined functions to facilitate robust virus replication in infected cells. Translation factors, such as eEF1A, are among the most common host factors recruited for (+)RNA virus replication. eEF1A is an integral component of several viral replicases, including the highly purified tombusvirus replicase complex. Since eEF1A is an essential G protein involved in translation elongation, it is difficult to obtain evidence for its direct involvement in virus replication in living cells. Indeed, down-regulation of eEF1A in cells has led not only to decreased TBSV repRNA accumulation, but also reduced p33 levels [25]. However, using a small set of functional eEF1A mutants defective in various functions revealed that eEF1A is involved in stabilization of p33 replication protein in yeast [25].
Based on the previous successful strategy of analyzing eEF1A mutants, here we generated ∼6,000 random mutants covering the entire eEF1A sequence and found four mutants, which greatly affected TBSV repRNA accumulation in yeast (Fig. 1A). Among these mutants, C42, C53 and C62 increased TBSV repRNA replication. Importantly, this effect by the eEF1A mutants was not due to changing the translation efficiency of p33/p92pol, but likely via directly altering viral replication and affecting the activity of the viral replicase. On the other hand, N21 mutant of eEF1A resulted in decreased TBSV RNA accumulation and also led to reduction in the level of p33 replication protein. This is reminiscent of the previously characterized GDP-binding mutant T22S [25], which supported greatly reduced level of viral RNA replication and p33 accumulation due to shortened half-life of p33. Overall, N21 mutant further supports that one of the functions of eEF1A in TBSV replication is to stabilize the p33 replication protein in yeast.
In addition to this genetic evidence on the relevance of eEF1A in TBSV replication in yeast, we also obtained additional supporting data by showing that chemical inhibitors of eEF1A, such as DB and GM, strongly inhibited replication of TBSV repRNA in the cell-free replication assay (Fig. 4A). Since we used the same amount of purified recombinant p33/p92pol in this in vitro assay (i.e., translation in the CFE is not needed for production of p33/p92pol), the role of eEF1A in TBSV replication must be separate from its role in protein translation. Altogether, these data strongly established that eEF1A is directly involved in TBSV replication, independent of the role of eEF1A in protein translation.
The identified eEF1A mutants were also useful to dissect the functions of eEF1A in TBSV replication. Based on a cell-free TBSV replication assay in CFE prepared from yeast expressing the C42, C53 or C62 mutants, we found that the minus-strand synthesis was enhanced by ∼3-fold, while the rate of plus-strand synthesis was proportionate with (−)RNA synthesis, resulting in ∼10-fold more (+) than (−)RNA products for wt and each mutant.
We confirmed a direct role for eEF1A in RNA synthesis in vitro by using a highly purified eEF1A and the recombinant TCV RdRp, which is closely homologous with the TBSV p92pol. Interestingly, it seems that eEF1A stimulates the RdRp activity directly, since pre-incubation of eEF1A and the RdRp prior to the RdRp assay led to the highest level of stimulation of (−)RNA synthesis (Fig. 3A). On the other hand, pre-incubation of eEF1A with the TBSV-derived template RNA led only to ∼2-fold increase in RNA synthesis in vitro (Fig. 3A). Analyzing the amount of short abortive RdRp products, which are produced through initiation followed quickly by abortive termination [57], in the in vitro assays revealed that eEF1A strongly enhanced the initiation of minus-strand synthesis (Fig. 3B). Although the actual mechanism of stimulation of RdRp activity by eEF1A is currently unknown, we propose that eEF1A might facilitate the proper and efficient binding of the RdRp to the 3′ terminal sequence of the viral RNA prior to initiation of (−)-strand synthesis (Fig. 5). Accordingly, eEF1A was shown to bind to the so-called replication silencer sequence (RSE) in the 3′-UTR, which is required for the assembly of the viral replicase complex [22], [58]. The binding of eEF1A-RdRp complex to the RSE might assist in placing the RdRp over the 3′-terminal promoter sequence, thus facilitating the initiation of (−)RNA synthesis starting from the 3′-terminal cytosine. Similar function of eEF1A in stimulation of (−)RNA synthesis has been proposed for WNV, based on mutations in the viral RNA within the eEF1A binding sequence that reduced the binding affinity of RNA to eEF1A and inhibited (−)RNA synthesis in infected cells [46].
Recent intensive work revealed that the assembly of the viral replicase complex is a regulated process involving viral- and host factors, cellular membranes and the viral (+)RNA [8], [10], [19], [63], [64], [65]. The assembly of the viral replicase also depends on steps occurring prior to the actual assembly process, such as selection of the viral template RNA and the recruitment of (+)RNA/protein factors to the sites of assembly. Although our current understanding is rather poor about the factors involved and their functions during replicase assembly, rapid progress is being made in this area due to the development of a new cell-free assay based on yeast CFE [30], [31]. The yeast CFE is capable of assembling the tombusvirus replicase complex in vitro in 40–60 min in the presence of recombinant p33/p92pol and the viral (+)repRNA [31], allowing for studies on direct roles of various factors. We find that inhibition of eEF1A activity by either DB or GM also inhibited the assembly of the tombusviral replicase complex based on time-course experiments (Fig. 4B) as well as a direct replicase assembly assay (Fig. 4C). On the contrary, the replicase activity was not inhibited by these compounds after the assembly took place (Fig. 4B–C). It is possible that after the formation of the eEF1A-RdRp-repRNA complex DB or GM are not effective in inhibiting the stimulatory effect of eEF1A on the RNA synthesis by the viral RdRp. Additional in vitro experiments with purified tombusvirus replicase preparations confirmed the lack of inhibition of RNA synthesis by DB or GM (Fig. S4A) on pre-assembled virus replicases.
The inhibition of the tombusvirus replicase complex by DB or GM might come from the ability of these compounds to inhibit the template RNA recruitment step (Fig. 4F). If the recruitment of the viral (+)RNA is inhibited, then the assembly of the viral replicase cannot take place in yeast or in vitro [21], [22], [31]. A target for GM and DB could be the inhibition of binding between eEF1A and the viral (+)RNA (Fig. 4D, E). Since the actual steps during the replicase assembly process are not yet known, it is possible that eEF1A might play additional roles in the assembly of the viral replicase complex.
The presented data are also in agreement with the function of eEF1A as a chaperone of the viral RdRp. Binding between the eEF1A and RdRp might alter the structure of the RdRp that favors de novo initiation for RNA synthesis. Indeed, the chaperone activity of eEF1A and its bacterial homolog EF-Tu has been shown before [66], [67]. Moreover, the EF-Tu-EF-Ts complex is thought to function in the Qbeta replicase complex as a chaperone for maintaining the active conformation of the RdRp protein [68].
Overall, the current work demonstrates two major functions for eEF1A in TBSV replication (Fig. 5): (i) stimulation of the assembly of the viral replicase complex, likely by facilitating the recruitment of the viral RNA template into the replicase; and (ii) enhancement of the minus-strand synthesis by promoting the initiation step. These roles for eEF1A are separate from its canonical role in host and viral protein translation.
Saccharomyces cerevisiae strain BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) was obtained from Open Biosystems (Huntsville, AL, USA). Plasmid-borne TEF1/2 TKY strains (MATα ura3-52 leu2-3, 112 trp1-Δ1 lys2-20 met2-1 his4-713 tef1::LEU2 tef2Δ pTEF2 URA3) were published before [69], [70], [71], [72]. The plasmid pESCHIS4-ADH-His33/CUP1-DI-72 expressing Cucumber necrosis virus (CNV) p33 and the TBSV replicon RNA, called DI-72, was described earlier [25]. The LYS2-based plasmid pRS317-Tet-His92, expressing CNV p92 under the control of Tetracycline-regulatable (Tet) promoter was constructed as follows: the Tet promoter sequence was obtained from pCM189-His92/Tet [73] by digestion with EcoRI and BamHI, and CNV p92 coding sequences from pGAD-His92 [7] digested with BamHI and PstI, followed by ligation into pRS317 vector treated with EcoRI and PstI. To generate mutations within TEF1 coding sequence by random mutagenesis, we constructed the TRP1-based plasmid pRS314-pTEF1-TEF1, which expressed TEF1 under the control of its native promoter. The TEF1 promoter sequence, the TEF1 coding region and the Cyc1 terminator sequences were amplified by PCR with the following primer pairs, #2764 (CCGCGAGCTCATAGCTTCAAAATGTTTCTAC)/#2765 (CCGCGGATCCGTAATTAAAACTTAGATTAGATTGC), #2768 (CCGCGGATCCAAAATGGGTAAAGAGAAGTCTC)/#1877 (CCGCCTCGAGTTATTTCTTAGCAGCCTTTTGAGCAGC), and #2769 CCGCCTCGAGGAGGGCCGCATCATGTAA/#2770 (CCGCGGTACCAGCTTGCAAATTAAAGCCTTC), respectively. This was followed by cloning the PCR products into pRS314 digested with SacI and KpnI.
The mutagenic PCR conditions were as follows: 50 mM KCl, 10 mM Tris (pH 8.3 at 25°C), 7 mM MgCl2, 0.3 mM MnCl2, 1 mM dCTP and dTTP each, 0.2 mM dGTP and dATP each, 0.2 µM of each primer, 20 pM of template DNA and 10 units of Taq polymerase in a 10 µl reaction volume in 10 aliquots. The PCR was performed for 30 cycles at 94°C for 1 min, 50°C for 1 min, and 72°C for 1 min in a conventional thermal cycler. Three overlapping ∼300–500 bp N-, central- and C-terminal segments of the TEF1 gene were amplified separately by PCR using primer pairs: #2767 (GTTTCAGTTTCATTTTTCTTGTTC)/#2788 (GAGTCCATCTTGTTGACAG), #2787 (CATCAAGAACATGATTACTGGTAC)/#2790 (GACGTTACCTCTTCTGATTTC) and #2789 (CGGTGTCATCAAGCCAGGT/#2771, (TTCGGTTAGAGCGGATGTGG), respectively.
Yeast strain TKY102 was co-transformed with constructs pESCHIS4-ADH-His33/CUP1-DI-72 and pRS317-Tet-His92 to induce TBSV repRNA replication according to standard Lithium acetate-PEG protocol [74]. The transformed yeast cultures were grown in a Synthetic Complete (SC) media with 2% glucose lacking leucine, histidine, lysine and uracil (SC-ULHK−) by shaking at 29°C overnight. To completely suppress TBSV replication before induction, 1 mg/ml Doxycycline was added to the media to inhibit the expression of p92. The plasmid pool carrying the randomly mutated TEF1 gene was introduced into the yeast cells already transformed with the two virus expression plasmids by in vivo gap repair mechanism via homologous recombination (Fig. S1A) [75]. Briefly, pRS314-pTEF1-TEF1 was digested with enzymes to truncate the TEF1 coding sequence, and then the digested plasmid was recovered. The gapped plasmid (5–10 µg) was transformed together with overlapping PCR (20 µg) products carrying the TEF1 mutations created by random mutagenic PCR (see above). The transformed yeast cells were selected on SC media lacking uracil, tryptophan, leucine, histidine and lysine. The colonies were further streaked onto SC media plate lacking tryptophan, leucine, histidine and lysine (SC-TLHK−) with 0.1% (w/v final) 5-Fluoroorotic Acid (5-FOA) media to select against the URA3-based wild-type TEF1 plasmid (Fig. S1A). This selection was repeated once and the loss of URA3 plasmid was confirmed by the inability of the yeast strains to grow on uracil-minus media. The yeast cells carrying the randomly mutated TEF1 were grown at 29°C for 24 h in SC-TLHK− media with 50 µM CuSO4 to induce virus replication. Total RNA extraction from yeast cells and Northern blotting and Western blotting were done as previously described [7], [25].
Whole cell yeast extract capable of supporting TBSV replication in vitro was prepared as described [31]. The in vitro TBSV replication assays were performed in 20-µl total volume containing 2 µl of whole cell extract, 0.5 µg DI-72 (+)repRNA transcript, 400 ng purified MBP-p33, 100 ng purified MBP-p92pol (both recombinant proteins were purified from E. coli), 30 mM HEPES-KOH, pH 7.4, 150 mM potassium acetate, 5 mM magnesium acetate, 0.13 M sorbitol, 0.4 µl actinomycin D (5 mg/ml), 2 µl of 150 mM creatine phosphate, 0.2 µl of 10 mg/ml creatine kinase, 0.2 µl of RNase inhibitor, 0.2 µl of 1 M dithiothreitol (DTT), 2 µl of 10 mM ATP, CTP, and GTP and 0.25 mM UTP and 0.1 µl of [32P]UTP [31]. The reaction mixture was incubated at 25°C for 3 h. The reaction was terminated by adding 100 µl stop buffer (1% sodium dodecyl sulfate [SDS] and 0.05 M EDTA, pH 8.0), followed by phenol-chloroform extraction, isopropanol-ammonium acetate precipitation, and a washing step with 70% ethanol as described [52]. The newly synthesized 32P-labeled RNA products were separated by electrophoresis in a 5% polyacrylamide gel (PAGE) containing 0.5× Tris-borate-EDTA (TBE) buffer with 8 M urea. To detect the double-stranded RNA (dsRNA) in the cell-free replication assay, the 32P-labeled RNA samples were divided into two aliquotes: one half was loaded onto the gel without heat treatment in the presence of 25% formamide, while the other half was heat denatured at 85°C for 5 min in the presence of 50% formamide [31]. S1 nuclease digestion to remove single-stranded 32P-labeled RNA was performed at 37°C for 30 min in a buffer containing 5 mM sodium acetate (pH 4.5 at 25°C), 0.28 M NaCl, 4.5mM ZnSO4 and 40 U S1 nuclease (Boehringer).
Fractionation of the whole cell extract was done according to [52]. The total extract was centrifuged at 21,000× g at 4°C for 10 min to separate the “soluble” (supernatant) and “membrane” (pellet) fraction. The pellet was re-suspended and washed with buffer A (30 mM HEPES-KOH pH 7.4, 150 mM potassium acetate, and 5 mM magnesium acetate) followed by centrifugation at 21,000× g at 4°C for 10 min and re-suspension of the pellet in buffer A. In vitro TBSV replication in the fractions was performed as described [31].
Expression and purification of the recombinant TBSV p33 and p92 and TCV p88C replication proteins from E. coli were carried out as described earlier with modifications [54]. Briefly, the expression plasmids were transformed separately into E. coli strain BL21 Rosetta (DE3). Protein expression was induced using isopropyl β-D-thiogalactopyranoside (IPTG) for 8 h at 16°C, then the cells were collected by centrifugation (5,000 rpm for 5 min). The recombinant TCV p88C protein was purified on an amylose resin column (NEB), as described [54]. The cells were suspended and sonicated in MBP column buffer containing 20 mM Tris-Cl pH 8.0, 150 mM NaCl, 1 mM EDTA, 10 mM β-mercaptoethanol and 1 mM phenylmethylsulfonyl fluoride (PMSF). The sonicated extract was then centrifuged at 27,000 g for 10 min, followed by incubation with amylose resin (NEB) for 1 h at 4°C. After washing the resin 3 times with the column buffer and once with a low salt column buffer (25 mM NaCl), the proteins were eluted with a low salt column buffer containing 0.18% (V/W) maltose and 6% (V/V) glycerol and stored at −80°C. MBP-p33 and MBP-p92pol were purified as above, except 30 mM HEPES-KOH pH 7.4 was used instead of 20 mM Tris-Cl pH 8.0. eEF1A was purified from yeast as described [76] and stored in aliquots at the vapor temperature of liquid nitrogen. Protein fractions used for the replication assays were 95% pure, as determined by SDS-PAGE.
Yeast strains (WT, C42, C53, C62) were transformed with plasmids pESCHIS4-ADH-HF33/CUP1-DI-72 expressing 6XHis- and Flag-tagged CNV p33 and the TBSV DI-72 repRNA, and pRS317-Tet-His92, expressing CNV p92 under the control of Tet promoter [25]. Co-purification was done according to a previously described procedure with the following modification [25]. Briefly, 200 mg of yeast cells were resuspended and homogenized in TG buffer [50 mM Tris–HCl [pH 7.5], 10% glycerol, 15 mM MgCl2, 10 mM KCl, 0.5 M NaCl, 0.1% Nonidet P-40 (NP-40), and 1% [V/V] yeast protease inhibitor cocktail (Ypic)] by glass beads using FastPrep Homogenizer (MP Biomedicals). The yeast cell lysate was cleared by centrifugation at 500× g for 5 min at 4°C to remove unbroken cells and debris. The membrane fraction containing the viral replicase complex was collected by centrifugation at 21,000× g for 15 min at 4°C and then solubilized in 1 ml TG buffer with a buffer containing 1% NP-40, 5% SB3–10 [caprylyl sulfobetaine] (Sigma), 1% [V/V] Ypic via gentle rotation for 1 h min at 4°C. The solubilized membrane fraction was centrifuged at 21,000× g for 15 min at 4°C and the supernatant was incubated with 20 µl anti-FLAG M2-agarose affinity resin (Sigma) pre-equilibrated with 0.7 ml TG buffer. After 2 h of gentle rotation at 4°C, we washed the resin 5 times with TG buffer containing 1% NP-40, the resin-bound replicase complex was eluted in 100 µl elution buffer [50 mM Tris–HCl [pH 7.5], 10% glycerol, 15 mM MgCl2, 10 mM KCl, 0.05 M NaCl, 0.5% Nonidet P-40 (NP-40), 1% Ypic and 0.15 mg/ml Flag peptide (sigma)]. In vitro RdRp activity assay was performed by using DI-72 RI(−) RNA template transcribed in vitro by T7 transcription [25].
EMSA was performed in a 10 µl-reaction containing 20 mM HEPES [pH 7.6], 50 mM KCl, 2 mM MgCl2, 1 mM DTT, 0.1 mM EDTA, 10% [vol/vol] glycerol, 10 U of RNase inhibitor, 10 nM 32P-labeled DI-72 (+) RNA probe and 0.5 µg purified eEF1A protein [76]. Reactions were incubated at room temperature for 20 min and then resolved by 4% nondenaturing polyacrylamide gel as described previously [25].
For in vitro eEF1A-repRNA co-purification, DI-72(+) repRNA was biotin-labeled in standard T7 transcription reaction in the presence of 20 µM Biotin-16-UTP (Roche). After the T7 transcription, the unincorporated biotin-UTP was removed on a Bio-Rad mini gel filtration column. The biotinylated RNA was immobilized on a column containing Streptavidin MagneSphere Paramagnetic Particles (SA-PMPs). Briefly, a 30-µl suspension of SA-PMPs (Promega) was washed three times with 1 ml of water and re-suspended in 1× Phosphate Buffered Saline (PBS). Biotinylated DI-72(+) RNA (5 µg) was then added to the suspension of SA-PMPs, followed by 30 min incubation at 4°C with gentle rotation. The SA-PMPs were collected on the side of the tube in a magnetic stand and washed 3 times with 1× PBS buffer. eEF1A was translated in vitro and labeled with 35S methionine using Rabbit Reticulocyte Lysate (Promega) according to manufacturer's manual. The in vitro eEF1A translation product (10 µl) was pre-incubated in a 200 µl binding buffer (20 mM HEPES [pH 7.6], 50 mM KCl, 2 mM MgCl2, 1 mM DTT, 1 mM GTP, 0.1 mM EDTA, 10% [V/V] glycerol, 1% BSA, 10 U of RNase inhibitor and 0.2% NP-40) with 150 µM Didemnin B (final concentration) or DMSO for 30 min at 30°C and then incubated with biotinylated DI-72(+) RNA-bound SA-PMPs for 1 h at 4°C. The SA-PMPs were collected in a magnetic stand and washed 5 times with the binding buffer, followed by elution with 30 µl SDS-PAGE sample buffer. The eluted protein samples were resolved by SDS-PAGE and then exposed to phosphorimager.
The TCV RdRp reactions were carried out as previously described for 2 h at 25°C [54]. Briefly, the RdRp reactions were performed in a 20 µl reaction containing 50 mM Tris-HCl (pH 8.2), 10 mM MgCl2, 10 mM DTT, 1.0 mM each ATP, CTP, and GTP, 0.01 mM UTP plus 0.1 µl of [32P]UTP, 7 pmol template RNA, 2 pmol affinity-purified MBP-p88C. 20 pmol eEF1A was added to the reaction at the beginning or as indicated in the text and Fig. 3 legend. The 32P-labeled RNA products were analyzed by electrophoresis in a 5% or 15% PAGE/8 M urea gel [57]. The 86-nt 3′ noncoding region of TBSV genomic RNA was used as the template in the RdRp assay [25], [54].
Purified Didemnin B (NSC 325319) was kindly provided by the Natural Products Branch, NCI (Bethesda, MD, USA), while Gamendazole was a generous gift from Dr. Tash (University of Kansas Medical Center). Both chemicals were dissolved in DMSO (the final concentration was 20 mM). The concentrations of chemical and time point of the addition of the chemicals to the in vitro reaction are indicated in the text. The cell-free TBSV replicase assay and the in vitro TBSV replicase assembly assay were performed according to [31]. Briefly, the purified recombinant TBSV p33, p92pol and (+) repRNA were added to the cell-free reaction in the presence of 1.0 mM ATP and GTP in step 1. After incubation at 25°C for 1 h, the in vitro reactions were centrifuged 21,000× g at 4°C for 10 min. The supernatant containing extra p33, p92pol and repRNA, which were not bound to the membranes in the cell-free extract, was discarded, while the membrane pellet was re-suspended in a standard in vitro replicase assay buffer containing [32P]-UTP and ATP, CTP, and GTP, and incubated at 25°C for 3 h [31].
The TBSV viral RNA gets recruited to the membrane from the soluble fraction with the help of TBSV replication proteins and host factors present in the yeast CFE. The in vitro RNA recruitment reaction was performed according to [31], except that 32P-labeled DI-72 (+)repRNA were used and rCTP, rUTP, 32P-labeled UTP, and Actinomycin D were omitted from the reaction. As a negative control, a recruitment-deficient repRNA, termed C99-G mutant, was used (Fig. 4F, lane 2) [23]. This mutant RNA is not recognized by p33/p92 replication proteins and it does not replicate in plants, in yeast or in the CFE in vitro [23], [31], [52], [77]. The RNA recruitment assay results in the assembly of the functional viral replicase, when wt repRNA is used, and nonfunctional replicase when the C99-G mutant is used in the assay (J. Pogany and P.D. Nagy, not shown) [31]. Inhibitors DB and GM were added at final concentration of 150 and 100 µM, respectively. After two hours of incubation at room temperature, 1 ml of reaction buffer was added to the in vitro assay, followed by incubation on ice for 10 min. Samples were centrifuged at 35,000× g for 1h, and the pellet was washed with 1 ml reaction buffer, followed by centrifugation at 35,000× g for 10 min. The membrane-bound repRNA was extracted from the pellet by adding 0.1 ml stop buffer and 0.1 ml phenol/chloroform and vortexing, followed by isopropanol/ammonium acetate precipitation [52]. The RNA samples were analyzed by denaturing PAGE and phophoimaging as described [52].
|
10.1371/journal.ppat.1003002 | Early Mechanisms of Pathobiology Are Revealed by Transcriptional Temporal Dynamics in Hippocampal CA1 Neurons of Prion Infected Mice | Prion diseases typically have long pre-clinical incubation periods during which time the infectious prion particle and infectivity steadily propagate in the brain. Abnormal neuritic sprouting and synaptic deficits are apparent during pre-clinical disease, however, gross neuronal loss is not detected until the onset of the clinical phase. The molecular events that accompany early neuronal damage and ultimately conclude with neuronal death remain obscure. In this study, we used laser capture microdissection to isolate hippocampal CA1 neurons and determined their pre-clinical transcriptional response during infection. We found that gene expression within these neurons is dynamic and characterized by distinct phases of activity. We found that a major cluster of genes is altered during pre-clinical disease after which expression either returns to basal levels, or alternatively undergoes a direct reversal during clinical disease. Strikingly, we show that this cluster contains a signature highly reminiscent of synaptic N-methyl-D-aspartic acid (NMDA) receptor signaling and the activation of neuroprotective pathways. Additionally, genes involved in neuronal projection and dendrite development were also altered throughout the disease, culminating in a general decline of gene expression for synaptic proteins. Similarly, deregulated miRNAs such as miR-132-3p, miR-124a-3p, miR-16-5p, miR-26a-5p, miR-29a-3p and miR-140-5p follow concomitant patterns of expression. This is the first in depth genomic study describing the pre-clinical response of hippocampal neurons to early prion replication. Our findings suggest that prion replication results in the persistent stimulation of a programmed response that is mediated, at least in part, by synaptic NMDA receptor activity that initially promotes cell survival and neurite remodelling. However, this response is terminated prior to the onset of clinical symptoms in the infected hippocampus, seemingly pointing to a critical juncture in the disease. Manipulation of these early neuroprotective pathways may redress the balance between degeneration and survival, providing a potential inroad for treatment.
| Neurodegenerative diseases affect an ever-increasing proportion of the population; therefore, there is an urgent need to develop treatments. Prion disorders belong to this group of diseases and although rare and uniquely transmissible, share many features on a sub-cellular level. Central to disease is progressive synaptic impairment that invariably leads to the irreversible loss of neurons. Understanding this process is undoubtedly essential for rational drug discovery. In this study we looked at neurons very early in disease, when prions are barely detectable and there are no clinical symptoms observed. Specifically, we performed a comprehensive analysis of transcriptional changes within a particularly dense area of neurons, the CA1 hippocampus region, from prion-infected and control mice. In this way we were able to enrich our data for molecular changes unique to neurons and minimize those changes characteristic of support cells such as astrocytes and microglia. We detected the activation of a transcriptional program indicative of a protective mechanism within these neurons early in disease. This mechanism diminished as disease progressed and was lost altogether, concurrently with the onset of clinical symptoms. These findings demonstrate the ability of neurons to mount an initial neuroprotective response to prions that could be exploited for therapy development.
| Prion diseases, or transmissible spongiform encephalopathies (TSEs), are fatal neurodegenerative diseases that affect both humans and animals. The conversion of the normal prion protein, PrPC (cellular prion protein), to the infectious form, PrPSc (Scrapie prion protein), is responsible for the disease pathology which is characterized by deposits of protease-resistant prion protein (PrPRes), extensive vacuolation plus microgliosis and astrocytosis. Although this protein conversion typically occurs over a long pre-clinical incubation period, neuronal dysfunction has been observed at these non-symptomatic stages of disease. In particular, synaptic loss was identified in animal models of prion disease prior to the manifestation of clinical symptoms [1]–[4]. This pathological change must be driven by the activation of molecular responses in prion affected cells during these long pre-clinical sequelae. Currently, the specific cellular responses by neurons to the presence of PrPRes, whether directly or indirectly associated, remain largely unknown. Identifying these specific pathways during prion disease progression, especially at pre-clinical stages of disease, may potentially help guide research in the understanding of prion-induced pathobiology.
Assessing global gene expression patterns to identify perturbed transcriptional networks during disease provides insight into the molecular pathobiology of prion diseases. To this end, numerous gene expression studies have been performed for many prion and mouse genetic strains with the aim of identifying affected pathways during disease [5]–[13]. Many of these studies investigated the changes in gene expression profiles over time from whole brain tissue by looking at both pre-clinical and clinical stages of disease. Nevertheless, the progressive genetic changes that eventually lead to neuronal death, or the mechanism by which neurotoxicity occurs, still remain unresolved. Typically, high-throughput studies on whole brain tissue are very challenging to interpret because of the cellular and functional complexity of the brain. Brain tissue is made up of a myriad of neuronal cell types that work together in intricate cellular networks. Adding to this complexity is the multitude of supporting cells such as astrocytes, microglia and oligodendrocytes that outnumber neurons by at least 10∶1, even in the healthy brain. Therefore, whole brain mRNA profiles are representative of an abundance of cells masking temporal genetic changes that are restricted to similarly affected neurons.
To overcome this challenge, we used laser capture microdissection (LCM) technology to isolate neurons from brain tissue sections [14]–[17]. We chose to isolate neurons distinct to the Cornu Ammonis layer of the hippocampus (CA1) because neurons in this region are known to undergo damage and degeneration during the course of prion disease [2] and represent a relatively homogeneous population of cell bodies for RNA isolation. We used this strategy to isolate this small group of neurons, rather than individual cells as a compromise; testing too small a sample size would increase variability due to heterogeneity of individual neurons [17] and a low amount of starting material would limit sensitivity for detecting lower copy number transcripts. We chose to sample more than one serial section for our studies in order to increase the probability of capturing these disease-affected neurons, especially at early stages of disease when we presume that only some neurons in the region are similarly affected.
We performed a thorough high-throughput screen for messenger RNAs (mRNAs) and microRNAs (miRNAs), a class of small RNAs with important roles as post-transcriptional gene regulators [18], [19]. Despite their recent identification, a substantial research effort has already shown that miRNAs are pivotal in fundamental processes such as neuronal differentiation, development, plasticity and survival (for review, see [20]). Links between miRNA dysfunction and neurodegenerative diseases are also becoming increasingly apparent [21], [22]. This follows from the observation that the loss of miRNA expression in the brain by inactivation of miRNA processing leads to neurodegeneration in a number of animal models [23]–[26]. Augmentation of our mRNA profile with miRNA expression levels is an important step in gaining a more complete understanding of the molecular mechanisms that are affected during the progression of prion disease. Furthermore, given their ability to modulate pathways related to survival, the introduction or knock-down/out of a miRNA may be an avenue for the development of therapeutics.
In this current study, we identified numerous genes and miRNAs, many of which were novel to prion pathobiology, that have altered expression levels in the microdissected CA1 neurons from infected mice. Uniquely, we found that transcript expression followed a temporally distinct pattern revealing the disease to be a dynamic process even at pre-clinical time periods, a characteristic not captured in a whole tissue analysis. More specifically, we identified a bi-phasic molecular response where the presence of a neuronal protective mechanism was up-regulated during early prion disease and this protection was subsequently diminished at late stages of infection, in line with the clinical manifestation.
All procedures involving live animals were approved by the Canadian Science Centre for Human and Animal Health - Animal Care Committee (CSCHAH-ACC) or the University of British Columbia Animal Care Committee according to the guidelines set by the Canadian Council on Animal Care. All protocols were designed to minimize animal discomfort. The approval identifications for this study were animal use document (AUD) #H-08-009 and AUD #H-11-020.
CD1 mice between 4 and 6 weeks of age were inoculated intraperitoneally with the Rocky Mountain Laboratory (RML) strain of scrapie using 200 µl of 1% brain homogenate in PBS from either clinically ill or normal control mice. Animals were sacrificed at intervals following inoculation and at the onset of clinical symptoms. Brains were collected at 6 time points [40, 70, 90, 110, 130 and terminal 153–161 days post infection (DPI)] and processed accordingly for either microdissection or pathological analysis. The terminal time point henceforth is designated as ‘end-point’, or EP, in the manuscript. Clinical signs used to delineate EP were kyphosis, dull ruffled coat, weight loss of 20% or more and ataxia. Samples designated for microdissection were covered in optimal cutting temperature (OCT) medium (Sakura Finetek), flash frozen using a dry ice/methanol mixture and stored at −80°C until processing, while samples planned for pathology were kept in formalin prior to tissue processing and embedding.
Brain samples frozen in OCT were cryo-sectioned into 8 µm thick coronal sections containing the CA1 hippocampal regions, placed on polyethylene-napthalate (PEN) membrane slides (Life Technologies Inc.) and stored at −80°C for less than 4 weeks before processing. The staining and preparation of sections for the laser capture microdissection (LCM) procedure was done using the LCM staining kit (Ambion) following manufacturer's recommendations. The microdissection of the CA1 hippocampal regions were done using the Veritas LCM instrument (Arcturus) where the laser power was set between 50–100 mW and each capture was done with 2 or fewer laser pulses. Overall, about 700–1500 neuronal CA1 cells were captured per cap by combining multiple hippocampal serial sections from each animal.
Total RNA was isolated using the RNAqueous –Micro Kit (Life Technologies Inc.) following the manufacturer's instructions. RNA concentration and quality was assessed for each sample, in duplicate, via the 2100 Bioanalyzer RNA 6000 Pico Kit (Agilent Technologies Inc.) using the protocol provided by the manufacturer. Samples with less than a RIN value of 5.9 were not used for downstream applications (Figure S1). We used the same sample to assess global mRNA and miRNA expression profiles as described below.
To determine gene expression changes, RNA from 4–6 infected and 4–6 control mice were assayed on the Agilent whole mouse genome 4×44K arrays (Agilent Technologies Inc.). Sections of microdissected neurons that were taken from each individual mouse were pooled; this was considered to constitute a biological replicate for subsequent analyses. Due to the low abundance of total RNA isolated from each mouse, two rounds of amplification from 2 ng of total RNA was performed using the Amino Allyl MessageAmp II aRNA Amplification Kit (Life Technologies Inc.) following the manufacturer's protocol. Amplified RNA samples were labeled using either Alexa Fluor 555 (Life Technologies Inc.) or Alexa Fluor 647 (Life Technologies Inc.) as described in the Amino Allyl MessageAmp II aRNA Amplification Kit. Two-colour competitive hybridizations were performed by randomly mixing one control with one infected sample for each time point and hybridized against the same array. Dye swap experiments were executed to remove potential dye bias, resulting in a total of 8 arrays per time point. The hybridization, wash and scanning protocols were followed as described for the two-color microarray-based gene expression analysis (Agilent Technologies Inc.) according to the manufacturer's recommendations. The conversion of the raw image files to data files and subsequent quality control (QC) assessment was performed using the Feature Extraction Software versions 9.1 to 10.5.1.1 (Agilent Technologies Inc.) and only arrays that passed the QC were further considered in the analysis. This resulted in at least 4 array data sets to be compared for all time points (Gene Expression Omnibus # GSE34530).
For Gene Ontology (GO) designation, we filtered out data by selecting genes as ‘present’ based on microarray probe signal intensities above a conservative detection threshold level of 100 units. A high threshold level was chosen due to the variation inherent when using very small amounts of RNA as starting material. This variation particularly affected the detection of low abundance transcripts resulting in the prediction of inaccurate ratios for many genes. The value of 100 units was chosen arbitrarily to remove some of this variability within the data. Significance was assigned to genes exhibiting at least a 2.5-fold change in expression over mock-infected mice and having a false discovery rate (FDR) calculated to be lower than 1%. We performed an initial comparison of enriched gene ontology designations within these lists with respect to prion disease progression using the program ToppCluster [27]. Lists from each time point were submitted to ToppCluster and using a p-value cutoff of 0.05 and the Bonferroni correction method, we determined the most significantly altered “Gene Ontologies” as well as associated “Mouse Phenotypes”. Up-regulated and down-regulated genes are portrayed separately as we noticed that the gene ontology annotations were largely distinct. Networks containing these enriched terms were then constructed to allow visualization of the functional relationships that accompanied disease progression. The ‘Abstracted’ network option in ToppCluster was used for this purpose, resulting in a matrix file compatible with Cytoscape [28], [29]. The network graphics were generated using the Spring Embedded Layout function and significance was based on edge weights. Hierarchical cluster plots were produced using GeneMathsXT (www.applied-maths.com) employing the cosine correction and WPGMC (median linkage) measure.
To assess the differential miRNA expression profiles throughout the time course study, TaqMan low density arrays (TLDA) for Rodent card A (Life Technologies Inc.) were used to profile 335 unique mouse miRNAs. Pre-amplification of these cDNA reactions was performed as per manufacturer's recommendations for low RNA input, which has previously been shown not to affect the miRNA expression profiles [30]–[33]. Briefly, 1 ng of total RNA from each sample was used for the reverse transcription (RT) reaction along with megaplex primer pools for Card A. The RT reactions were then mixed with the TaqMan Universal PCR Master Mix (2×), No AmpErase UNG (Life Technologies Inc.) and loaded onto each TLDA. Overall, 2 scrapie infected and 2 control mouse samples were separately run per TLDA card. All real-time PCR assays, including mRNA and miRNA validations, were run on the TaqMan 7900HT Thermocycler with Sequence Detection System (SDS) software version 2.3 and analyzed by automatically calculating the Ct values via the RQ Manager version 1.2 (Life Technologies Inc.).
TLDA analysis was preceded by an initial inspection of each amplification curve. We considered curves that did not exhibit smooth amplification characteristics as background noise and Ct values for such curves were removed from further analysis. The choice of most appropriate normalization controls and the delta Ct calculations were performed using the real-time StatMiner software version 4.2 (Intergromics). The normalization control for Card A was chosen using a three step procedure. Initially, we narrowed down our list of potential controls based on the smallest Ct range found across all detectors during our time course study of which both snoRNA-135 and snoRNA-202 had a Ct range of 1.8. By comparing U6 (a commonly used control), snoRNA-135 and snoRNA-202 using 3 available stability scoring methods from StatMiner (Normfinder, Genorm and minimum variance median), we found that snoRNA-135 was the most stable endogenous control in our experiment (Figure S2A). Lastly, we considered the abundance of each endogenous control in the system. The limited total RNA input that was used for the qRT-PCR assays further diluted the signal of many miRNAs we were detecting. Hence, we also decided to consider the kinetic differences between low abundance and high abundance miRNAs in the assay when choosing our endogenous control as we expected highly abundant small RNAs to behave differently in the assay than the less abundant miRNAs. Considering that the abundance of many of our assayed miRNAs were exhibiting medium to low expression levels, snoRNA-135 was again the best candidate to fit this criteria (Figure S2B). Therefore, the Ct values of snoRNA-135 for each card were used to normalize the signal from each probe (delta Ct). Fold changes were calculated by the 2−(delta delta Ct) method [34] with standard error and Student's t-test statistics assessing significance. Due to the variability of the samples collected, true significance was represented by p-value≤0.1.
One-dimensional hierarchical clustering was performed on probes resulting in fold changes for at least 6 individual arrays (representing 3 different time points). This was done via GeneMathsXT using the Average Linkage (UPGMA) method and the Euclidean distance (with variance) measure.
We further assessed the expression profiles of 30 genes using quantitative real-time PCR (qRT-PCR). A total of 2 ng of total RNA from 4 infected and 4 control mouse samples for each time point were reverse transcribed separately using the High Capacity cDNA reverse transcription kit (Life Technologies Inc.). cDNA was purified using the ChargeSwitch PCR Clean-Up Kit (Life Technologies Inc.) as per manufacturer's recommendations. Concentration and rough quality was assessed using the Nanodrop ND-1000 Spectrophotometer (Thermo Scientific). Reverse transcribed DNA samples were pooled based on time point and treatment and a total of 50 ng of cDNA was used for each individual real-time reaction, performed in triplicate. For the real-time PCR reaction we used the TaqMan fast universal PCR master mix (2×), no AmpErase UNG (Life Technologies Inc.) following the fast run specifications recommended by the manufacturer. GAPDH was used as the endogenous control and fold changes were calculated by the 2−(delta delta Ct) method. Standard error and Student's t-test statistics were calculated to determine significance for which a p-value≤0.05 was chosen as the cut-off.
We analyzed the expression levels of 7 miRNAs (miR-16-5p, miR-26a-5p, miR-29a-3p, miR-140-5p, miR-132-3p, miR-146a-5p and miR-124a-3p) using a multiplex qRT-PCR approach. Overall, samples from 3–4 mice were pooled per treatment for each time point and 1 ng of the total RNA was used for the multiplex qRT-PCR reaction using the TaqMan Universal PCR Master Mix (2×), no AmpErase UNG (Life Technologies Inc.) as described by the manufacturer. To save on the limited amount of material available from each LCM preparation, we performed a multiplex qRT-PCR. More specifically, each cDNA primer was diluted ¼ to make the final primer stock and 4 µl of this stock was used in place of the 5× TaqMan miRNA RT primer. Real-time assays were run in triplicate, and Ct values for each probe were normalized to the snoRNA-135 control. The analysis was done as described for the mRNA qRT-PCR.
We followed the same immunohistochemistry protocol to stain for IBA1, CREB and phospho-CREB (pCREB). For IBA1, we used rabbit anti-IBA1 antibody (Wako Pure Chemical Industries, Ltd.) at a 1∶1500 dilution in EnVision FLEX antibody diluent (Dako). Total CREB was detected using rabbit anti-CREB (48H2) antibody (Cell Signalling Technology) at 1∶2000 dilution in SignalStain Antibody Diluent (Cell Signaling Technology) while pCREB was detected using the rabbit anti-pCREB antibody (Millipore) at a 1∶800 dilution in EnVision FLEX antibody diluent.
Each brain sample was fixed in 10% neutral buffered formalin and paraffin-embedded from which 5 µm thick coronal serial sections containing the hippocampal region were produced. Sections were, baked overnight at 37°C, deparaffinized and hydrated. Endogenous enzyme activity was blocked by immersing the sections in 2.5% hydrogen peroxide (Fisher Scientific) with 5% ethanol for 10 minutes at 37°C. Subsequent antigen retrieval was done by placing the slides in 10 mM sodium citrate buffer (pH 6.0) and incubating at 121°C for 10 minutes followed by 30 minutes cool down at room temperature. A total of 100 µl of the specific primary antibody was added to the sections and incubated overnight at 4°C. Antibody signal was detected using the Klear Mouse HRP-Polymer DAB Detection System (Golden Bridge International GBI, Inc.) following the manufacturer's recommendation Sections were counterstained with hematoxylin for 20 seconds, dehydrated, cleared and mounted using Paramount (Fisher Scientific).
Infectious PrPSc deposits were identified by using rabbit prion monoclonal antibody (EP1802Y) (Abcam) at a 1∶7000 dilution in EnVision FLEX antibody diluent. Briefly, we executed the same preliminary procedure as described for the immunohistochemistry section, including the antigen retrieval and blocking endogenous enzyme activity. At which point the staining intensity was enhanced by incubating section in 80% formic acid for 10 minutes at room temperature. Slides were rinsed in tap water for 10 minutes and washed twice for 2 minutes each in TBS/Tween (TBS/T). Slides were then treated with 4 M guanidine thiocyanate at 4°C for 2 hours, washed with TBS/T twice for 2 minutes each and treated with diluted rabbit anti-PrP at 4°C overnight. Stain detection was performed as described above using the Klear Mouse HRP-Polymer DAB Detection System. This was followed by counterstaining with hematoxylin, clearing and mounting as described above.
We used a double immunofluorescence staining procedure to detect astrocytes and nuclei in our sections. Briefly, we executed the same preliminary procedure as described for the immunohistochemistry sections, including the antigen retrieval step, at which point sections were blocked in 1∶20 normal goat serum (Cedarlane) that was diluted in EnVision FLEX antibody diluent for 1 hour at room temperature. The slides were incubated with rabbit anti-GFAP antibody (Abcam) at a dilution of 1∶500 in EnVision FLEX at 4°C overnight. Sections were rinsed twice with TBS/T and incubated for 1 hour at room temperature with Alexa Fluor 594 goat anti-rabbit IgG secondary antibody (Life Technologies Inc.) diluted to 1∶1000 in EnVision FLEX. Slides were then rinsed twice with distilled water for 2 minutes each and DAPI (Life Technologies Inc.) was used to stain nuclei at 1∶1000 dilution for 20 minutes at room temperature. Slides were then rinsed with distilled water, dehydrated, cleared and mounted with DPX containing coverslips (Sigma).
Flouro-Jade C (Millipore) staining was employed to detect degenerating neurons. Slides were processed as described above up to the antigen retrieval step at which point they were transferred to Flouro-Jade C solution for 30 minutes. Slides were rinsed twice in distilled water for 2 minutes per rinse followed by dehydration, clearing and mounting as described above.
Immunohistochemistry slides were scanned using a MIRAX MIDI scanner (Zeiss) and regions of interest were exported and analyzed with ImageJ software [35]. Briefly, images were first colour deconvolved to separate haematoxylin and DAB stains. A haematoxylin-stained nuclei channel was used to delineate the CA1 hippocampal region. Isolated nuclear regions were obtained by empirically determining a thresholded pixel intensity to generate a mask. The masked area was measured and then used to segregate the CA1 hippocampal region in the DAB-stained probe channel. Positively-stained regions were thresholded and the masked area was measured. A scoring method was required for some DAB-stained probe channels. Positive signal intensity was binned into 3 ranges by increasing threshold amounts from background to a maximal intensity. Graphing and statistics were performed using GraphPad Prism version 5 (GraphPad Software).
Each section was deparaffinized, rehydrated and wasted twice in DEPC-PBS for 2 minutes each. Slides were fixed with 4% paraformaldehyde (PFA) for 20 minutes at room temperature and washed twice in DEPC-PBS for 5 minutes each. Permeabilization was performed by treating the slides with 10 µg/ml proteinase K (Life Technologies, Inc.) for 10 minutes at 37°C and washed twice in DEPC-PBS for 2 minutes each. Slides were fixed again using 4% PFA for 15 minutes and washed with DEPC-PBS. Pre-hybridization of slides was performed by exposing the slides to the pre-hybridization buffer (BioChain Institute, Inc.) for 4 hours at 60°C. We added 50 nM of the linearized DIG-labeled miRNA-132 LNA probe (Exiqon) in ready to use hybridization solution (BioChain Institute, Inc.) to the slides and hybridized at 55°C overnight. Slides were washed in 2×SSC (Ambion) for 10 minutes at 55°C followed by a 1.5×SSC wash for 10 minutes at 55°C and twice in 0.2×SSC for 20 minutes at 37°C. Slides were then washed in TBS/T twice and incubated in 1× blocking solution (BioChain Institute, Inc) for 1 hour at room temperature. For visualization, we incubated the slides with 1∶500 AP-conjugates anti-digoxigenin antibody in block solution at 4°C overnight. The slides were then exposed to alkaline phosphatase buffer twice for 5 minutes each at room temperature. Slides were incubated with NBT and BCIP for 4 hours at room temperature, rinsed with distilled water and counterstained with Nuclear Fast Red (Vector Laboratories, Inc.). These slides were mounted using aquatic buffer.
TargetScan version 5.2 (June 2011) was used to predict mouse specific miRNA targets. We chose all the conserved target gene lists for each of the 6 miRNAs along with miR-124a-3p target lists for further analysis. We compared all the predicted target genes for each miRNA using ToppFun module from the ToppGene Suite software (http://toppgene.cchmc.org/) [36]. ToppFun detects enrichment of genes for various features from which we chose to focus on Gene Ontology (GO) molecular function, biological processes and cellular component features. For each analysis, we used the Bonferroni correction and a p-value of 0.05 as a cut-off.
Mice were intraperitoneally (IP) inoculated with either the RML strain of mouse-adapted scrapie, or mock-infected, and animals were sacrificed at six time points spanning very early to terminal disease [40, 70, 90, 110, 130 and terminal 153–161 days post infection (DPI)]. Densely packed pyramidal cells from CA1 hippocampal tissue were isolated from 8 µm sections using the LCM. Total RNA was extracted and the relative global transciptome expression levels between infected and control mice was determined at each time point (Figure S3). Agilent whole mouse genome microarrays comprising 44,000 probes were used to detect mRNA expression levels. In total, we found 2,580 genes to be differentially expressed by greater than 2.5 fold during at least one time point with false discovery rates (FDR) of <1% (Table S1).
Prior to performing an in depth analysis of time-dependent transcriptomic changes, we assessed the relative abundances of neurons, astrocytes and microglia in the sampled CA1 region from both infected and uninfected mice. This served as a baseline for the relative proportions of these cells in each sample, and to qualitatively measure the extent of infiltration and activation of astrocytes and microglia within this region over time. We initially inspected the expression levels of numerous cell type-specific marker genes which were further correlated with quantitative immunohistochemistry. The expression of the house-keeping gene GAPDH, like the majority of genes assayed on the array, remained relatively stable throughout the time-course for both control and infected mice (Figure S4); therefore, GAPDH was used as a control for internal variability.
Astrocyte abundance in the LCM generated samples was assessed using glial fibrillary acidic protein (GFAP) expression levels. We found that the GFAP spot intensities were below the cut-off level for detection in all of the CA1 samples dissected from control mice. In turn, a detectable level of GFAP was only evident in RNA samples from RML-infected mice sacrificed at terminal stages of disease, or end point (EP) (Figure 1A). Immunohistochemical staining of GFAP reciprocated these findings; stained cell bodies of astrocytes were normally present in extremely low numbers in the CA1 microdissected region. Significant infiltration of astrocytes to this area was apparent during clinical disease (Figure 1B). We next determined the presence of microglial marker genes within the CA1 hippocampal region. The expression of the microglial marker genes allograft inflammatory factor (AIF1), otherwise known as IBA1, and also F4/80 a marker of murine macrophage populations (data not shown), were detected in both control and infected mice at all time points. Nevertheless, a statistically significant increase in the abundance of these transcripts in infected animals was only evident at 130 days post infection and EP (Figure 1A). Immunohistochemistry using an antibody directed against IBA1 mirrored this data (Figure 1C and D). Microglial staining was evident in all samples and was associated with cells interlaced amongst the densely packed CA1 neurons. Staining was more intense beginning at 110 DPI and the total numbers of cell bodies stained brown was approximately 3-fold higher in infected mice versus control by 130 DPI and EP, thus, mimicking the microarray expression profile.
Gene markers for neuronal cells such as the neurofilament medium and light polypeptides (NEFM and NEFL) plus synaptosomal-associated protein 25 (SNAP25) were confirmed as highly enriched in the sampled tissue, as was expected given the stringent dissections (Figure 2A). We observed a reduction in the signal intensities of these marker genes only at clinical stages of disease in comparison to age-matched, mock-infected controls. Statistically significant reductions of these genes were evident at 130 DPI and EP, however, the decline in expression was relatively small compared to the total signal suggesting that a substantial proportion of neuronal cell bodies were present at EP. We noticed that the neuronal genes showing the most significant decreases in expression, and that were clearly detectable by 130 DPI, coded for proteins that are associated with synaptosomal fractions, such as SNAP25. We further visualized degenerating neurons using Flouro-Jade C, a stain that associates with all neurons undergoing degeneration, regardless of specific insult or mechanism of cell death [37]. Flouro-Jade C staining revealed relatively few degenerating neurons in the sampled CA1 pyramidal layer, even at the end stage of disease (Figure 2B), which agrees with the microarray findings. This mirrors previous studies showing that cell death in this region occurs in the final stages of disease [2], [38]. However, Flouro-Jade C staining did reveal that interneurons in adjacent layers of the hippocampus, located within the stratum radiatum (SR) or stratum lacunosum-moleculare (SLM), exhibited higher proportions of degenerating neurons much earlier in the disease process; by 110 DPI with scarce staining even detectable at 90 DPI (data not shown). These results would suggest that some neurons are more susceptible to damage associated with prion replication and that CA1 pyramidal neurons appear to be particularly robust. However, since the CA1 neurons are intricately associated with adjacent interneurons through synaptic connections, the observed neuronal degeneration in these interneurons would most likely impact the responses of the neurons residing within the CA1 hippocampal region.
The accumulation of PrPRes in the CA1 hippocampal region over time was also confirmed by immunohistochemical staining. Occasional PrPRes deposits associated with CA1 neuronal projections were first detectable at 90 DPI in the stratum lacunosum-moleculare with more consistent buildup evident from 110 DPI until EP (Figure 2C). Although minimal PrPRes presence was directly associated with the cell bodies of CA1 pyramidal neurons during early stages of disease, the dendrites of these neurons pass through the SR and/or SLM regions. Replication here could well contribute to a prion replication-associated response that is reflected in the alteration of mRNA expression within the cell bodies of these neurons.
Overall, we found that the temporal expression changes of cell-specific gene markers from microarrays correlated well with the different cell types and neuronal degeneration observed by immunohistochemical and Flouro-Jade C staining. Given that extensive infiltration of astrocytes and amplification of microglia are only evident at late stages of infection, we believe that transcriptional alterations observed prior to and including 110 days post infection most likely reflect either neuronal events or very early transcriptional changes in microglia precipitating their activation and accumulation.
The most striking finding in our analysis of the microarray data was that alterations within the LCM sampled region followed a strictly temporal pattern (Figure 3A). Differential expressions at 70, 90 and 110 DPI (pre-clinical) are quite dissimilar to those changes that are induced at 130 DPI or EP. The timing of the major sequential changes in CA1 neurons, beginning by 70 DPI, correlates well with detectable prion replication in the hippocampus and suggests that the transcriptome is temporally altered as a direct molecular response to the progressive replication of prions. However, we cannot rule out earlier changes as we did not perform exhaustive analysis prior to this time. Interestingly, between 110 and 130 DPIs, the transcriptional response in the CA1 pyramidal region transitions in two ways; firstly, the initial wave of differential expression either returns to basal levels or follows an inverted pattern, where for example, previously up-regulated genes become down-regulated. Secondly, smaller waves of novel gene expression changes become apparent by 130 DPI, with very strong differential expression detectable at EP.
Our first step in the analysis was to compare those gene-ontology functional designations that were enriched with respect to days post inoculation in order to identify some of the annotated molecular events that accompany prion disease-progression. Briefly, lists of genes that exhibited at least a 2.5-fold change in expression over mock-infected mice with an FDR lower than 1% were compared between each time-point. An initial finding showed the significant delineation between up-regulated and down-regulated genes within ontologically classified genes across the data set. We found that we were able to improve the discrimination of significant ontology's amongst these genes by analyzing these two groups separately. Networks generated using the program ToppCluster are displayed in Figure 3B and C where some of the major annotations in terms of gene functional groups are highlighted (detailed annotation of the networks is provided in (Figure S5).
Up-regulated genes fell into a small number of statistically significant ontological groupings at the early time-points; whilst an overwhelming abundance of these enriched gene groups were found only at the clinical end-point of infection. Although a representative group of ontologies are highlighted in Figure 3B, approximately 85% of the ontological and phenotypic gene categories at clinical end-point reflect the development and function of immune-related cells. The appearance of these genes correlate well with the infiltration of activated GFAP expressing astrocytes plus increased numbers and activation of microglia, and contains the majority of genes previously reported to be differentially expressed in prion disease [5], [7], [13]. A hierarchical cluster plot summarizes the expression pattern of a group of immune-related genes that have been previously identified to be over-expressed in prion disease as well as other neurodegenerative conditions (Figure S6). Interestingly, although the majority of these genes are only strongly expressed during clinical disease, a minority of genes with immunological functions begin a sequential over-expression at 110 DPI such as ICAM1, VCAM1, FYB or S100A4, while some genes show a down-regulated trend such as IRAK1 or CDC42. The adhesion molecules ICAM1 and VCAM1 are typically expressed on endothelial cells and cells of the immune system, including microglia and astrocytes [39]. In this case, microdissection clearly allows us to track the temporal expression of these adhesion molecules well before the detection of immune activation markers such as cytokines and chemokines. These genes are likely expressed in the resting microglia that we found to be closely associated with neurons in this region. They appear to be up-regulated within the CA1 pyramidal region prior to the infiltration of activated immune cells and may be involved in either the activation of phagocytosis and prion-aggregate removal, or in the recruitment of and/or induction of inflammatory pathways in microglia or astrocytes [40], [41]. These particular genes may be useful genetic indicators of very early cellular response(s) to damage.
Up-regulated genes enriched at pre-clinical stages fell into much fewer significant groups of functionally related transcripts. However, the significant groups we identified were found to include genes expressed in response to various stressors as early as 70 days post-infection (Figure 3B). A significant number of these transcripts code for genes known to be regulated by the transcription factor cyclicAMP responsive element binding protein 1 (CREB1). In fact, the IPA analysis tool identified CREB as the top upstream regulator of these genes with a p-value of 10−15 (data not shown). These included a number of highly inducible transcription factors such as FOS, EGR2, EGR4 and NR4A1. No significant change in CREB mRNA was identified by microarrays, which was confirmed by qRT-PCR (data not shown) and immunohistochemical staining of total CREB in the CA1 region (Figure 4A and B). However, it is the phosphorylation of CREB at Ser133 that leads to the recruitment of transcriptional coactivators required for the expression of a large number of genes (for review, see [42]). Staining with an antibody to detect CREB phosphorylated at Ser133 showed that pCREB significantly increased in CA1 sections from infected mice only during pre-clinial disease, as compared to similar sections from uninfected mice (Figure 4C and D). In addition, a number of kinases such as TRIB1 and CAMK2D plus extracellular matrix glycoproteins such as LAMC2 were also strongly up-regulated. We chose 3 of these, EGR2, CAMK2D and TRIB1, to validate by qRT-PCR. All three of these genes were similarly up-regulated in LCM samples at pre-clinical disease taken from a second group of infected mice inoculated in an independent experiment (Table 1).
Networks generated using enriched terms amongst down-regulated genes exhibited considerable functional consistency between pre-clinical and clinical time-points (Figure 3C). Notably, the ontology annotations strongly implicate alterations in the expression of genes that are involved in neuronal projection and dendrite development from early time points onwards. This is consistent with the reduction in dendrites reported as the earliest detectable pathological change in neurons during prion disease [1], [43]. Similarly, perturbations in synaptic transmission and synapse formation are implicated throughout disease progression, from mid-way through the incubation period and beyond [2], [44], [45], coinciding with our 70 DPI time-point. The 70 and 90 days post-infection times also reveal alterations in calcium ion binding and receptor activities while later time-points (130 and EP) show additional down-regulation amongst multiple ontologies relating to vesicle formation and transport, some structural proteins such as those related to cytoskeletal physiology as well as transcripts related to behavioral changes. The reduction in expression of genes specifically involved in synapse formation and/or maintenance, such as synaptophysin (SYP), synaptotagmin (SYT1), alpha-synuclein (SNCA) and SNAP25, plus general dendrite physiology, is particularly striking at 130 DPI and EP.
Interestingly, groups of genes that have been previously related to obvious phenotypic changes in mice (shown in Figure 3B) were almost solely contained within the end-point sampling, correlating closely with the onset of clinical symptoms beginning around 130 DPI. These phenotypes include abnormal synaptic transmission leading to behavior, learning and memory deficits plus abnormal innate immunity; all of which are clinical symptoms of neurodegenerative conditions including prion disease.
Detecting and exploring the earliest transcriptional changes that occur during pre-clinical disease (70–110 DPI) was the particular focus of this work. As previously described, a cluster of approximately 400 genes was significantly dysregulated in the pyramidal layer of the CA1 at these early time points. We performed a literature search to identify published genomic data generated from hippocampal neurons that may be associated with survival/death mechanisms. This meta-analysis proved fruitful; in particular, a comparison with data sets generated by Zhang et al to identify gene expression changes in hippocampal primary neurons in response to NMDA stimulation [46], [47] revealed striking similarities to our dataset. The aim of these two studies was to determine transcriptional programs triggered by the stimulation of either synaptic NMDA receptors (NMDARs) alone, or by overstimulation at extrasynaptic locations. Mounting evidence suggests that the subcellular location of these NMDARs in neurons is critical to the response that follows (for review, see [48]). Synaptic NMDAR activity appears to transduce a signal promoting a neuroprotective genetic cascade, whereas Ca2+ influx through extra-synaptic localized NMDARs appears to directly oppose these effects and promote cell death. In total, 185 neuronal genes were reported to be altered in expression following Ca2+ influx through synaptic NMDARs [46]. Validatory studies revealed that a number of these changes were able to promote neuroprotection in primary neurons. Comparison of this list with our cluster of genes altered between 70 and 110 DPI (FDR<5%) revealed 97 genes in common. Of the rest, 22 genes were below the level of detection in our study, 45 were unchanged early in disease although a number of these were altered during clinical disease, and 21 were not represented on our array (Figure 5A and B).
A core set of 9 genes were validated by Zhang et al [47] as being critical to the neuroprotective response and these were termed Activity-regulated Inhibitor of Death genes, or AIDs. Significantly, we identified 5 out of the 9 AID genes to be up-regulated (BTG2, GADD45G, GADD45B, NPAS4, NR4A1), INHBA did not appear to change in expression and the remaining 3 genes (ATF3, IFI202B and SERPINB2) were below the level of detection in our study (Figure 5B). We also used qRT-PCR to validate the pre-clinical over-expression of 4 of these genes previously unrecognized as playing a role in prion pathobiology (Table 1 and Figure 5C); RASGRF2, TRIB1, MCL1 and HOMER1 were all confirmed to be up-regulated during the same time period post-infection in CA1 pyramidal neurons. We believe that the similar alteration of 60% of detectable genes from the Zhang study including many of the critical AID genes is highly suggestive of the induction of a similar neuroprotective program in the CA1 during early prion disease.
We have previously used whole mouse brain tissue to identify miRNA expression alterations in prion infection [49]. We therefore extended this study to identify changes in miRNA expression levels in the same CA1 hippocampal RNA preparations that we used for mRNA profiling. Differential miRNA expression was determined using a multiplex qRT-PCR platform, TaqMan Low Density Arrays (TLDA). Over the course of the disease, we found 88 miRNAs that had altered expression levels for at least one time-point based on a p-value of <0.1 (Table S2). Analysis of the expression profiles of miRNAs from these samples revealed similar patterns of sequential expression to those exhibited by mRNAs (Figure 6A). From the 88 miRNAs found to be differentially expressed throughout the experimental study, 17 miRNAs were only deregulated early in the course of disease prior to 130 DPI, 57 miRNAs were up-regulated at 130 DPI and/or EP of disease while 8 miRNAs were detected only in infected samples at these time points (Table S3).
Although little is known about the functional roles of many miRNAs and therefore, classical bioinformatics analysis based on ontology is not very definitive; we were able to use prior knowledge to predict the roles of some of the temporally altered clusters of miRNAs. Previously, we showed that gene expression restricted to glial cells is either only detectable, or prodigiously induced, primarily in the LCM generated samples collected at EP. As this was also borne out by immunohistological staining, we hypothesized that glial-associated miRNA expression would follow a similar pattern. Specifically, we compared those 57 miRNAs deregulated or induced in the samples taken at clinical end-point with published data describing miRNA up-regulation in immune stimulated primary astrocytes [50]. Of the 57 miRNAs deregulated at these time points we found that 44, (∼77%), correspond to miRNAs that are induced in activated astrocytes (Table S4). By association, it is likely that the majority of these 44 miRNAs are involved in innate immune related functions within activated astrocytes and possibly activated microglia that infiltrate our dissection area during clinical disease.
We further observed that 8 of these miRNAs were either only detectable or induced during clinical disease based on TLDA results. Specifically, miR-19a-3p, miR-344-3p, miR-34b-3p and miR-497-5p were all detected only in samples from RML infected mice (Figure S7A). This pattern emulates the expression of astrocyte-specific genes such as GFAP. In turn miR-150-5p, miR-26b-5p and miR-410-3p were detectable in all the microdissected samples, although only substantially up-regulated at EP. By analogy with our mRNA data, we postulate that these miRNAs are either expressed at a basal level in microglia prior to their activation or alternatively, in resident neurons and are induced only by severe stress (Figure S7B). Although we cannot rule out the possibility of induction in damaged neurons, comparison of the relative level of expression of each of these miRNAs to a list of those expressed in mouse hippocampal neurons [51] reveals that the majority have very low abundance in neuronal cells (Table S5). Therefore, based on their pattern of expression and low abundance in neurons we conclude that these miRNAs are most likely induced in activated microglia.
In contrast to the other 8 miRNAs in this group we found miR-146a-5p to be up-regulated early in disease as well as at clinical end-point (Figure 6B). MiR-146a-5p was of interest as it has been reported to be up-regulated in brain tissue in a number of neurodegenerative conditions including prion animal models [50] as well as in human cases of Creutzfeldt-Jakob disease (CJD) and Gerstmann-Sträussler-Scheinker (GSS) [52]. Its function as an immune response regulator is well documented [53], and we have determined its induction in activated microglia [54]. The up-regulation of miR-146a-5p during the 70–110 DPI time-period, however, is suggestive to originate from CA1 pyramidal neurons although a specific role of this miRNA in neuronal cells has not been reported.
We next determined the abundance of miR-124a-3p, a miRNA that is well-known to be highly enriched in neurons [55], [56], to see if its expression level mirrored those of the neuronal-specific marker genes tested previously. We did indeed detect high levels of miR-124a-3p in both infected and uninfected mice and saw a significant reduction in miR-124a-3p abundance in the CA1 dissected samples from prion infected mice at the clinical stage of disease (Figure 6C). Interestingly, we noted that this miRNA was also up-regulated in both the 70 and 90 DPI samples collected from infected mice suggesting a potential disease-related function during this period.
Following confirmation that miRNA and mRNA expression profiles are to some extent analogous, our next goal was to validate the expression of the earliest miRNAs found to be deregulated; i.e. during the 70–110 DPI time-period. We had initially identified 17 miRNAs to be deregulated at this stage in disease from the TLDA assays (Table S6). We further confirmed the expression levels of 7 miRNAs using qRT-PCR, miR-16-5p, miR-26a-5p, miR-29a-3p, miR-132-3p, miR-140-5p, miR-124a-3p and miR-146a-5p, all of which were up-regulated in infected samples prior to 130 DPI (Figure 6B, C and D). The overall pattern of expression of the majority of this group was a significant up-regulation throughout pre-clinical disease followed by a drastic decrease in expression for many at 130 DPI and EP. This pattern of expression is analogous to many of the mRNAs that formed our early deregulated cluster. We also used in situ hybridization to confirm the expression level of miR-132-3p because the expression level of this miRNA is known to be induced by pCREB. We found that miR-132-3p appears to be significantly more abundant at pre-clinical time points in the CA1 granule neurons in infected versus uninfected cells (Figure 6E), in conjunction with pre-clinical pCREB expression levels.
To confirm that these miRNAs are likely enriched in neurons, we compared their relative expression to a list of miRNAs expressed at basal levels in the cell bodies of rat hippocampal neurons that was generated by Kye and colleagues [57]. As many miRNAs are highly conserved between mammals (Table S7) this type of interspecies comparison is applicable. Apart from miR-140-5p and miR-146a-5p, all of these early miRNAs are abundantly expressed in rat neurons (Table 2). Next generation sequencing of small RNAs isolated from mouse hippocampus also confirms the endogenous expression of all 7 of these miRNAs in this brain region [52]. In fact, miR-29a-3p, miR-26a-5p and miR-132-3p are particularly highly expressed in the hippocampus, being even more abundant than miR-124a-3p, such that 2.63, 1.54 and 0.43% of total miRNA reads were composed of these miRNAs, respectively (Table 2). Furthermore, miR-146a-5p, miR-16-5p and miR-140-5p, although lower in abundance than miR-124a-3p, were also present in mouse hippocampus in this study. Although some of these miRNAs are expressed at very low levels in neurons, miR-140-5p has been similarly detected by miRNA microarray of rat hippocampus [58] while miR-146a-5p presence has been linked to the CA1 region of the rat hippocampus in one study using in situ hybridization [59].
Although not enough is known about miRNA function in neurons to perform an ontological analysis as is done for mRNAs, we performed a preliminary analysis of the bioinformatically predicted targets of the 7 miRNAs confirmed to be dysregulated during early disease using TargetScan [60]–[62]. We found a total of 396 target mRNAs for at least one of the 7 miRNAs we validated that were also down-regulated between 70 and 110 DPI. To identify the biological processes that these genes are associated with, we ran an enrichment analysis based on GO terms on these genes. In total, 80 GO biological process annotations were identified (Table S8) of which many were associated with neuronal function. Of these, 14 processes are strictly related to neuronal functions (Table 3) where synaptic organization is the most represented pathway containing approximately 11% of the miRNA target genes that are both down-regulated and targeted by at least 1 of the 7 candidate miRNAs. This preliminary data suggests that further investigation into the functions of these miRNAs may help determine regulatory pathways that are affected during early pathogenesis of prion disease.
We have demonstrated that hippocampal CA1 cells undergo a clear temporal transcriptional response to the challenge of infection and propagation of prions. During pre-clinical disease a persistent stimulation of a programmed response was identified that appears to be mediated, at least in part, by synaptic NMDAR signaling. The coordinated program of the induced gene expression suggests the promotion of cell survival and neurite remodeling pathways. This response terminates prior to the onset of clinical symptoms in the infected hippocampus, seemingly pointing to a critical juncture in the disease (Figure 7). The changes described here center on the pre-clinical stages of disease at which time many of the gene alterations we identified are novel to prion disease.
Factors intrinsic to neurons are likely crucial for the initiation of degeneration in prion disease as prion replication in brain cells other than neurons does not appear to trigger neuronal death, and consequently, the clinical stage of disease [63]–[65]. Genome wide analysis of gene expression is a powerful tool for understanding the underlying pathogenic mechanisms of diseases. Several genome wide studies have been performed to date but the information they provide is very broad and encompasses all cell-types specific pathologies, with the most marked expression changes relating to gliosis. Discrimination between alterations in affected neurons and other cell types would increase the specificity of subsequent analysis and enable the determination of temporal changes specific to one cell type.
We chose to study CA1 hippocampal neurons, cells that are known to undergo physiological and morphological changes in mouse models of scrapie beginning at pre-clinical stages of infection. Our analysis firstly demonstrated that we were able to isolate relatively homogenous populations of neurons by LCM from mice sacrificed at various time-points throughout the incubation period of RML scrapie. The expression levels of neuronal marker genes, based on microarray signal intensity, was extremely high in our samples versus microglial and astrocyte markers, which were below the threshold set for detection by our arrays. These data correlated very well with immunohistochemical staining that revealed the presence of a few IBA1 stained microglial cell bodies and even less astrocytes amongst the densely packed neuronal layer. Astrocyte cell bodies stained for the marker GFAP were more abundant in areas through which dendrites of CA1 neurons pass. Although glial cell proliferation and activation is apparent within the diseased brain from the earliest times at which PrPRes is detectable, their cell bodies do not appear to infiltrate the densely packed CA1 neuronal layer until late in disease. Accordingly, microglial and astrocyte genetic markers and the cytokines/chemokines they produce upon activation are not detectable by microarray in our LCM samples prior to EP. Therefore, we are confident that those alterations detected at pre-clinical stages of disease are highly enriched for CA1 neuron expressed genes.
Furthermore, FluoroJade C staining and analysis of neuronal marker genes was not indicative of widespread degeneration and loss of neurons within the LCM region. In fact, the neuronal cell bodies within this area appear relatively robust; other neurons in the hippocampal formation appeared more vulnerable to degeneration. Neurons in other brain regions, especially the thalamus area, also showed significant degeneration that was detectable using FluoroJade C at pre-clinical stages of disease. It will be of interest to compare the transcriptional response of these more vulnerable neurons to those of the CA1 layer. These neurons may well expose the pathways that lead to neuronal cell death from those that modulate the synaptic and morphological changes reported in the hippocampus.
Although it is beyond the scope of this paper to compare and contrast all of the gene expression changes we identify with previous studies, it is important to make some general comments for comparative purposes. Most importantly, direct comparison with our own previously published data confirms that LCM in this case significantly increased the sensitivity for detecting altered genes that are associated with neuronal function, such as the transmission of nerve impulses and neuron projection morphogenesis. Particularly striking was the approximate 4-fold increase in the number of genes detected that are involved in synapse formation and processes relating to behavioral changes in mice. These gene expression changes intricately mirror the alterations in synaptic properties, and the morphology of dendritic spines, that are the recognized neuropathalogic features of early disease [1], [2], [66], [67].
A brief comparison of our data with the comprehensive genomic analysis of prion infected whole brain tissue presented by Hwang and colleagues [5] serves to illustrate some of the ways in which our targeted LCM studies can complement information gleaned on a global scale. For example, Hwang et al reported the novel finding that the global stimulation of the androgen biosynthesis pathway is a general feature of prion disease, implying that altered levels of neurosteroids may play a role in pathogenesis. These biosynthetic pathways were not detected in our study presumptively because cells other than those in the CA1 are carrying out these syntheses. However, multiple neurosteroid-induced signaling pathways such as the aldesterone and corticotropin-releasing hormone signaling pathways were found to be significantly deregulated in our data, interestingly at the earlier time-points (data not shown). These changes likely reflect neuronal physiology or perhaps pre-activation microglial alterations, thus, further supporting a link between perturbed androgen metabolism and glial activation or neuronal-glial signaling. Other metabolic pathways of interest identified by Hwang et al. [5] were those for arachidonate and prostaglandin synthesis. We found these pathways to be significantly deregulated at early time-points as well, suggesting their dysregulation in CA1 resident cells. No doubt, numerous other parallels could be drawn by detailed comparisons of individual genes.
We did look amongst our genes for clear evidence of pathways that may be involved directly in the cell death of neurons. One that may be important in prion induced neurodegeneration is activation of the endosomal and lysosomal pathway, or autophagy, within neurons [68]. However, we did not see any gene deregulation indicative of an increase in endosomal and lysosomal compartments or the autophagic system in our LCM neuronally enriched samples. The lysosomal activation marker LAMP2 is up-regulated only at EP, and we saw no change in the mRNA levels of the autophagy marker genes BECN1 (beclin 1), ATG5 and MAP1LC3B (LC3-II). This contrasts with some studies suggesting that either the activation of these systems within neurons, or their aberrant dysfunction, is an early and pervasive mechanism instigating the degenerative process [69]–[72]. This suggests that any changes we see in endosomal/lysosomal related genes over basal levels are limited to phagocytic cells clearing debris, or else, only triggered within terminal neurons long after the synaptic alterations are apparent. However, in assessing our data we also need to consider that this study relates to a highly specific group of neurons within the hippocampal CA1 region. As previously mentioned, FluoroJade C staining revealed that these neurons seem particularly robust in comparison to other neurons affected during prion disease. Therefore, it is possible that autophagy may play a larger role in these more vulnerable neurons. In addition, autophagy is a fundamental process that is carried out in all neurons and perhaps the process itself is important in prion pathobiology, but the up-or down-regulation of these pathways at the gene level does not occur. Further studies on other neuronal subsets that display differing vulnerabilities to prions will be of great interest in delineating the responses of different neurons and perhaps identify novel or cell-type specific pro-death programs. These are very important considerations for the development of appropriate therapies.
Gene expression within the CA1 neurons is dynamic and appears to have at least two distinct phases. In this manuscript we describe the earliest detectable transcriptional alterations induced in infected mice that occur prior to any major loss of cells. At this stage, beginning between 40 and 70 DPI and ending prior to 130 DPI, we see a major cluster of altered genes and miRNAs that either return to basal levels, or alternatively undergo a direct reversal in expression profile during clinical disease. Notably, ontological analysis at each time point revealed that many of those neuronal specific genes were altered at pre-clinical stages of infection (Figure S8). Individual genes include alterations in receptors and ion channels that may lead to changes in the synaptic properties of the CA1 cells. Also altered are numerous cytoskeletal processes involved in dendrite morphology and synapse assembly. Pre-clinical changes also involve a regulatory response to stimulus and stress, including mitochondrial and spliceosomal changes. The dysregulated expression of genes involved in vesicular transport and function are evident by day 110 post-infection and it is not until day 130 that we see some expression of genes involved in the positive regulation of programmed cell death. Although we identified a number of genes involved in cell death and apoptosis, the list was quite small and does not point to a specific predetermined pathway.
Although at least 150 genes were up-regulated pre-clinically, these did not fall into many functional groups according to annotated ontologies. We looked at these in more detail and noticed that many were activity-regulated genes that are induced by various neuronal stimuli. Many, such as FOS and EGR1, are known to share binding sites for the trans-acting transcription factor CREB. Although CREB was not up-regulated at the mRNA level in the early stages of disease we were able to confirm by immunohistochemistry that phosphorylated CREB (the active form) is increased in CA1 pyramidal neurons during pre-clinical prion disease, correlating with the upregulation of potentially CREB regulated genes. Synapse dysfunction, compromising nuclear calcium signaling, is known to play a key role in multiple neurodegenerative diseases that are associated with the loss of synapses [73]–[74], of which prion disease is one example [75]–[77]. In this model, compromised synaptic signaling leads to the induction of calcium regulated genomic programs that are postulated to result in the degeneration and eventual death of neurons.
In recent years, a number of studies have shown that NMDA receptor signaling can play pivotal roles in the development of these pathologies as well as in other conditions such as ischemia, epilepsy and schizophrenia [78]–[80]. Additionally, NMDAR signaling was observed to be altered in PrPC knockout mice in a number of reports showing that there is an impaired synaptic inhibition taking place via weakened GABAA receptor-mediated fast inhibition [81] and an increased excitability of dentate gyrus hippocampal neurons from slice cultures [82]. Recent data further supports this observation by showing that PrPC knockout enhanced NMDAR activity causing heightened excitability states of neurons and enhanced glutamate excitotoxicity, both in vitro and in vivo, suggesting that normal PrPC mediates a neuroprotective role [83]. Therefore, we were very interested to discover that, in a meta-analysis of our data with other published sets of genomic data from hippocampal neurons, our gene list was highly reminiscent of that created by stimulation of synaptic NMDARs in primary hippocampal neurons [47].
NMDA receptors have distinct subunit compositions and subcellular locations that allow them to partner with different signaling pathways, creating functional diversity [84]. The NMDAR is composed of a subunit coded by three regulatory subunits whose composition can vary. The NR2 subunits reportedly differ between synaptic and extrasynaptic receptors. NR2B–containing receptors are more commonly located extrasynaptically and appear to promote cell death by inhibiting ERK1/2, inactivating CREB and stimulating excitotoxicity [85]–[87]. In turn, NR2A subunits appear to be involved in promoting cell survival [88]. It has been postulated that the ratio of NR2A and NR2B subunits also contributes to a balance between the protective and pro-cell death pathways [89], [90]. We found that GRIN2B, which is the gene coding for the NR2B subunit, was significantly down-regulated, according to microarray analysis, during early prion infection (70–90 DPI) with levels rising at later stages of disease. This finding was further validated by qRT-PCR. Interestingly, it has been previously reported that with decreased levels of NR2B subunit, the NR2A subunit increases in abundance at the synapse [91]. Therefore, our data is suggestive of an alteration in the ratio of these subunits towards a NR2A enriched composition in early stages of disease, substantiating the NMDAR neuroprotective response. Although some studies suggest that both NR2A and NR2B can activate cell death and survival [92], [93], NR2B receptor location preferentially at the extrasynaptic sites [94], [95] may reflect the pro-death phenotype. Additionally, we validated the expression of the transcription factor TRIB1 (tribbles), also found in the Zhang study [47]. The function of this gene is not well described, however, members of this family appear to be involved in the regulation of a number of fundamental signaling pathways and have been implicated in numerous human diseases [96]. This may well be an important regulatory molecule to investigate further in regards to neurodegeneration.
It has often been suggested that the lack of a proper neuroprotective response to the stress caused by prion build-up is at the heart of the neurodegenerative process. However, our analysis of temporally induced changes suggests that synaptic signaling leading to the induction of an anti-apoptotic genetic program is evident early in the course of disease. Perhaps, as disease progresses, these neuroprotective programs are gradually either shut-off or actively antagonized by a pro-death system. Excessive glutamate mediated NMDAR signaling (excitotoxicity) that initiates neurodegenerative processes has been linked to the development of a number of chronic disorders including Alzheimer's disease (AD) and perhaps prion disease [97]. It was therefore possible that the shift in gene expression patterns that occurs between 110 and 130 DPI represents a turning-point after which the over-stimulation of NMDARs switches from protection towards the induction of excitotoxicity related cell death. We looked at our data set further to investigate this possibility but we found little evidence of the induction of gene expression indicative of excitotoxicity. Of note, Zhang and colleagues identified a second group of genes whose expression was stimulated by an excessive amount of glutamate at extra-synaptic NMDARs and were involved in cell death [46][47]. The expression of these genes was detected in the CA1 samples in our analysis, however, we did not see any deregulation of these genes, including the reported ‘key’ genes DAPK1 and CLCA1 (data not shown).
Although we did not detect a clear pattern of NMDAR stimulated pro-death genes, this does not necessarily rule out extrasynaptic pro-death signaling as it is possible that either additional genes were below our threshold of detection in the LCM generated samples, or that the pathway may be induced post-transcriptionally. Indeed, the total number of pro-death genes identified in the Zhang study was much fewer, only 11 versus the 185 pro-survival genes. Although we did not find a clear pro-death genomic program we did identify some up-regulated genes that have been implicated in neuronal death in a number of acute and chronic neurodegenerative conditions. The gene FOXO1 appears to be activated in response to various stress stimuli, such as epileptic seizures and oxidative and endoplasmic reticulum stress, and its role appears to be to eliminate damaged neurons by apoptosis [98]. A FOXO target gene also found to be up-regulated was BBC3 (PUMA), a Bcl-2 homology 3 (BH3)-only member of the Bcl-2 family that is known to be involved in endoplasmic stress-induced apoptosis in neurons [99]. The role of endoplasmic reticulum (ER) stress is an avenue that should be further explored in prion diseased neurons as potential mechanisms leading to their ultimate death.
NMDA receptors are highly permeable to calcium ions and it has been postulated that disruption of Ca2+ homeostasis may be a major contributing factor in prion disease. Cellular PrP protein may even have a major direct role in Ca2+ homeostasis (for review, see [100]). This could occur via two means: PrPC misfolding during disease interrupts Ca2+ homeostasis or PrPRes oligomerization forms pores in cell membranes which are permeable to ions. The downstream effects are fluctuations in the amounts of Ca2+ found in neurons that results in the type of physiologic and genetic impacts we see in neurons during early stages of disease, such as alterations in neuronal excitation and neurite outgrowth. A number of groups have reported a reduced Ca2+ response in brains and/or cultured neurons infected by prions along with electrophysiologic and morphologic synaptic abnormalities that may be the result of perturbed Ca2+ signaling [45], [101]–[103]. We chose to validate some of the genes involved in these processes that were dysregulated in the microarray analysis at early stages of infection. We confirmed the early upregulation of CAMK2D, CAMK4 and CAMK1; Ca2+/calmodulin-dependent protein kinases that are activators of the transcription factor CREB. NMDA glutamate receptors are coupled to the activation of the Ras/Erk signaling cascade and to the maintenance of CREB transcription factor via RASGRF1 (p140 Ras-GRF1) and RASGRF2 (p130 Ras-GRF2), a family of calcium/calmodulin-regulated guanine-nucleotide exchange factors that activate the Ras GTPases [104]. Consistent with this, ischemia-induced CREB activation is reduced in the brains of adult Ras-GRF knockout mice while neuronal damage is enhanced [105]. We validated the expression of RASGRF2 which was significantly up-regulated 70 and 90 DPI providing further support for stimulation of this pathway during early disease.
We validated the early over-expression of HOMER1, another gene that is induced by synaptic activity and a novel finding in prion disease. HOMER1 is a gene belonging to a family of dendritic scaffold proteins that regulate group 1 metabotrophic glutamate receptor (mGluR) function and have roles in dendrite morphology [106]. It has also been shown to be up-regulated in neurons early in Alzheimer's disease [107]–[109]. The binding of HOMER1 to mGluR1 leads to the release of intracellular calcium and major cellular consequences such as neuronal excitability changes, enhancement of neurotransmitter release, the potentiation of the activity of NMDA or AMPA receptors and effects on dendrite spine density and morphology [110]–[113].
An additional role for Ca2+/calmodulin-dependent kinases is in chromatin remodeling by the inhibition of HDAC activity through phosphorylation [114]. Interestingly, one of the gene expression alterations we validated was the progressive downregulation of HDAC9 over the course of the disease beginning at 90 DPI. A seminal study has also linked the expression of HDAC9 with dendritic growth in neurons [115]. In this study, knockdown of HDAC9 appeared to promote dendritic growth and concomitantly, expression of the immediate early gene FOS was down-regulated independent of synaptic activity. This study suggested that chromatin remodeling and the nucleocytoplasmic translocation of HDAC9 was able to regulate activity-dependent gene expression and dendritic growth [115]. The progressive down-regulation of HDAC9 correlates well with the down-regulation of immediate early genes at later stages of disease. Deregulation of genes involved in dendrite, axon and cytoskeleton development (both down- and up-regulation) was particularly obvious throughout the course of prion disease. Expression of another related gene, DOCK1, was assessed using qRT-PCR which was found to be up-regulated at 70 and 90 while down regulated by 110 DPI and remained unchanged during clinical disease. Previous reports showed that DOCK1 is involved in cytoskeletal reorganization [116] by mediating axon outgrowth, attraction [117] and pruning [118]. Recently, additional evidence suggests that DOCK1 mediates regulation of spine morphogenesis in cultured hippocampal neurons [119], further highlighting an essential role in dendrite morphology early in disease that is reminiscent of the neuronal protective mechanisms.
Another set of genes identified as deregulated at early stages of disease were those involved in transport processes such as binding of unfolded proteins, the localization of proteins to clathrin-coated and synaptic vesicles as well as those involved in adhesion and the spliceosome. It was beyond the scope of this study to validate the genes involved in all of these processes. However, we did confirm the early increase in expression of myosin motor genes, MYO6, MYO5A and DAB2 that function in intracellular vesicle and organelle transport by clathrin-mediated endocytosis in non-neuronal cells. Interestingly, all of these proteins have been shown to be expressed in neurons, localized to synapses and enriched at the postsynaptic density [120]–[122]. MYO6 has many roles in hippocampal neurons such as the clathrin-mediated endocytosis of acid-type glutamate receptors, such as α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors (AMPARs), and the formation of synaptic structures and astrogliosis [123]. MYO6 also binds directly to DAB2 which has also been postulated to be a negative regulator of neurite outgrowth [124]. MYO5A has been shown to be pivotal for the extension of the endoplasmic reticulum into neuronal dendritic spines [125].
We were especially excited to confirm the deregulation of a number of miRNAs in CA1 neurons during early prion disease as this is novel to the field. In a number of ways we found that miRNA expression profiles mirrored gene expression patterns. In total, 17 miRNAs were identified as significantly dysregulated between 70–110 DPI and we successfully validated the expression levels of 6 of these miRNAs: miR-16-5p, miR-26a-5p, miR-29a-3p, miR-132-3p, miR-140-5p and miR-146a-5p. All of these miRNAs were highly up-regulated in infected samples as compared to controls at early stages of infection. All miRNAs, except miR-146a-5p, showed a similar pattern to many of the mRNAs that were up-regulated early in that they returned to basal levels by 130 DPI and were down-regulated in clinical disease. Therefore, we hypothesize that by association these miRNAs may well have related functions. Although published studies on the functions of many of these miRNAs are few, a number provide evidence for the function of miR-29a-3p, miR-124a-3p and miR-132-3p in neuronal synapse formation and plasticity. More specifically, the dendritic arborization and neurite outgrowth appears to be controlled, in part, by these miRNAs.
MiR-124a-3p is one of the most studied miRNAs in the nervous system and is important in the differentiation of progenitor cells into mature neurons [126], [127]. Recent evidence suggests that miR-124a-3p promotes embryonic [128], [129] and adult neurogenesis in vivo [130]. Although the exact mechanism remains unknown, multiple gene targets for miR-124a-3p have been identified. One recent target of miR-124a-3p is Ephrin-B1 [131] which functions to regulate cytoskeletal dynamics [132]. Interestingly, neurons with higher miR-124a-3p expression levels exhibited increased neurite outgrowth [132], similar to miR-124a-3p over-expression in isolated cortical progenitor cells [128]. Recent evidence has also shown that LHX2 down-regulation by miR-124a-3p is important in preventing the apoptosis of hippocampal and retinal neurons during development [55]. Perhaps, this indicates that miR-124a-3p could potentially have a general role in neuroprotection. Similarly, neuronal-specific functions have been extensively shown with miR-132-3p levels. One of the targets of miR-132-3p in hippocampal neurons is the Rho family GTPase-activating protein p250GAP [133]. The repression of this gene via miR-132-3p results in a greater dendritic complexity of hippocampal neurons [133]. More specifically, a greater number of stubby and mushroom shaped spines exhibiting longer protrusion widths have been observed [134]. In turn, miR-29a-3p was shown to target ARPC3, a component of the ARP2/3 actin nucleation complex that initiates branching in actin filaments [135], which regulates the morphology of dendritic spines [136]. The authors found that miR-29a-3p reduced the number of mushroom-shaped dendritic spines on hippocampal neurons while filopodial-like outgrowths became more frequent [136]. Interestingly, the authors suggested that miR-29a-3p may function as a homeostatic mechanism to counterbalance excessive positive cues to stimuli at the synapse [136].
The enrichment shown amongst the group of miRNAs deregulated pre-clinically in our prion model for those that play roles in dentrite and synapse formation fits very well with the ontology assignments of genes similarly deregulated. Synaptic disruption and dendritic abnormalites have long been recognized as the earliest pathologic consequence of prion replication in the brain and miRNAs up-regulated in early disease appear to be key regulatory molecules in these processes. Dendritic spines are quite dynamic throughout the lifespan of a CA1 neuron, however, over-stimulation effects by a ‘toxic’ agent may further enhance the change in dendrite and spine morphology; a function that may potentially involve miR-124a-3p, miR-132-3p and miR-29a-3p and possibly also the other miRNAs, such as miR-16-5p, miR-26a-5p, miR-140-5p and miR-146a-5p whose functions in neurons have yet to be characterized.
Recent publications have, however, revealed one interesting gene target of miR-16-5p. One of these is amyloid protein precursor (APP) and in a murine model of early-onset Alzheimer's disease the authors found a decrease in miR-16-5p levels while APP was increased [137]. Upon a short term infusion of miR-16-5p into the brains of diseased mice, the authors found a decrease in APP accumulation, further suggesting that miR-16-5p has an inhibitory effect on APP protein formation [137] and therefore, acts as a protective molecule in the disease. In prion disease, we saw an up-regulation of miR-16-5p during early disease and the expression of this miRNA decreased with disease progression. It would be an interesting area of study to determine whether miR-16-5p has a similar neuroprotective role in prion diseased mice.
Although little is known about specific functions of miR-26a-5p, miR-140-5p and miR-146a-5p within neurons, recent evidence points to their involvement within the central nervous system. More specifically, miR-26a-5p has been shown to regulate photoreceptor L-type voltage-gated calcium channel alpha1C in retinal cells [138]. MiR-140-5p targets EGR2, modulating myelination in dorsal root ganglion and Schwann cell co-cultures [139] while miR-146a-5p has been identified in the CA1 neurons of rat hippocampus using in situ hybridization [59]. Therefore, additional investigation into the neuronal function of these miRNAs may further enhance our understanding of prion-induced pathobiology.
As miRNA expression in prion disease has not been extensively reported in the literature we provide a brief discussion of those miRNAs deregulated, or only detected at EP, in infected mice. By analogy with gene expression data we expect that the majority of these genes will correspond to glial expressed miRNAs involved in an innate immune response. Noteworthy is the miR-146a-5p, a well know immune regulatory miRNA that is induced in activated microaglia and was significantly up-regulated at the end-point of disease with a fold change of 3.09±0.13 (p-value≤0.01) (Figure 5D). Increased expression of miR-146a-5p in mouse scrapie, human prion disease (CJD and GSS patients) and Alzheimer's disease has been reported and it is suggested that this miRNA may be a general biomarker of neuroinflammation [49] [52]. Accordingly, this role may potentially be applied to the other similarly expressed miRNAs we described. Interestingly, we also found that miR-146a-5p is up-regulated early in disease, suggesting a pre-clinical function during disease in addition to a role in activated glia. Although miR-146a-5p is well studied in a number of different immune cells its expression in neurons has not yet been reported. We did confirm that we can detect miR-146a-5p expression in this region by qRT-PCR as well as in primary mouse hippocampal neuronal cultures (data not shown). In addition, in situ hybridization from our (data not shown) and Aronica et al [59] suggests that hippocampal CA1 neurons express this miRNA in significant levels in contrast to other neurons. Further work on miR-146a-5p will be required to determine if it has a neuronal specific function. In immune cells, miR-146a-5p targets TRAF5 and IRAK1 to suppress the inflammatory response to stimulation and we also showed in microglia that predicted targets may be involved in microglial morphology [54]. It will be interesting to determine whether miR-146a-5p has any similar targets in CA1 neurons and whether it plays a role in neuronal degeneration.
Of note, we were unable to detect miR-494-3p or miR-342-3p to be deregulated in our samples at clinical disease, both of which were detected in scrapie infected mice [49] and in BSE-infected macaques, a model for CJD [140]. However, we determined that miR-342-3p is expressed at low levels in the hippocampal CA1 neurons while a greater expression of this miRNA was observed in cerebellar granule neurons (data not shown). Therefore, we expect that this discrepancy is most likely attributed to these miRNAs being expressed in particular brain cell types other than hippocampal CA1 neurons.
Taken together, our work provides insight into the previously unknown temporal nature of the transcriptional changes that accompany prion replication. Beyond the reduction of transcripts related to synaptic vesicle trafficking and dendrite morphology in neurons during pre-clinical infection, our data suggests that miRNAs may be involved as post-transcriptional regulators of mRNAs that are functionally associated to axonal and dendritic synaptic remodelling. Dendritic spines are specific sites of protein synthesis and it is generally thought that dendritic mRNAs are transported to that location as part of ribonucleoprotein complexes [141]. MicroRNAs are then able to modulate dendritic morphology by regulating expression of proteins involved in the actin cytoskeleton [142] [132], [143]–[145] mRNA transport [146] and neurotransmission. However, the role of post-transcriptional deregulation of genes functionally associated to axonal and dendritic synaptic remodeling in prion diseases is as yet unknown. Given the concomitant alterations in genes known to promote neuronal survival we postulate that remodeling of dendrite morphology during pre-clinical disease is a response to stress. Further investigatation of the roles of the prion-induced miRNA deregulation will provide an inroad for the study of these as yet unexplored pathomechanisms in disease. Ultimately this knowledge may contribute to the identification of therapeutic molecules targeted to specific neuronal processes that contribute to disease pathobiology. The identification of the multi-phasic nature of the neuronal response adds an extra layer of complexity to the design of therapeutics. We envision two approaches, either the maintenance of the initial genetic response to promote neuronal survival or blocking a later event that promotes neuronal cell death (Figure 7A), such as that described by Moreno et al [4]. The introduction or knock-down/out of miRNAs themselves is one potential avenue for novel drug treatment. Manipulation of miRNAs may affect a number of pathways simultaneously and may allow the modulation of a specific process, rather than a complete inhibition of a pathway for which a basal level is required by the cell. Continued investigation of transcriptional and post-transcriptional deregulation in prion neurodegeneration will undoubtedly identify the molecular basis for clinical disease and help to evaluate new drug treatments.
|
10.1371/journal.ppat.1006773 | Histone demethylase LSD1 restricts influenza A virus infection by erasing IFITM3-K88 monomethylation | The histone demethylase LSD1 has been known as a key transcriptional coactivator for DNA viruses such as herpes virus. Inhibition of LSD1 was found to block viral genome transcription and lytic replication of DNA viruses. However, RNA virus genomes do not rely on chromatin structure and histone association, and the role of demethylase activity of LSD1 in RNA virus infections is not anticipated. Here, we identify that, contrary to its role in enhancing DNA virus replication, LSD1 limits RNA virus replication by demethylating and activating IFITM3 which is a host restriction factor for many RNA viruses. We have found that LSD1 is recruited to demethylate IFITM3 at position K88 under IFNα treatment. However, infection by either Vesicular Stomatitis Virus (VSV) or Influenza A Virus (IAV) triggers methylation of IFITM3 by promoting its disassociation from LSD1. Accordingly, inhibition of the enzymatic activity of LSD1 by Trans-2-phenylcyclopropylamine hydrochloride (TCP) increases IFITM3 monomethylation which leads to more severe disease outcomes in IAV-infected mice. In summary, our findings highlight the opposite role of LSD1 in fighting RNA viruses comparing to DNA viruses infection. Our data suggest that the demethylation of IFITM3 by LSD1 is beneficial for the host to fight against RNA virus infection.
| The viral genomes of DNA viruses but not RNA viruses form chromatin structure during infection. Thus, epigenetic modulators are not expected to have crucial roles in RNA viral infection. However, here, we identify for the first time, that, opposite to its role in enhancing DNA virus replication, LSD1, a histone demethylase, limits RNA virus replication. We show that, under IFNα treatment, LSD1 is involved in the demethylation of IFITM3, a well-known host restriction factor for many RNA viruses. To counteract IFITM3 activation by demethylation, several RNA viruses, such as VSV and IAV, but not Zika virus, have developed strategy to inactive IFITM3 by promoting its dissociation from LSD1. In agreement with our findings, the inhibition of the enzymatic activity of LSD1 by small molecule leads to more severe disease outcomes in IAV-infected mice. Our data suggest that although LSD1 inhibitor is beneficial for treating DNA virus infection, it could be harmful to the host suffering from RNA virus infection. On the contrary, developing strategies to stimulate LSD1 activity to demethylate of IFITM3 is essential to fight RNA viruses.
| DNA virus genomes are encapsidated without histones, but rapidly acquire chromatin structure upon infection[1, 2]. Thus, epigenetic regulators, such as histone methylase and demethylase, are found to be able to regulate DNA virus replication. Recently, increasing studies have shown that inhibition of LSD1 (lysine-specific demethylase 1; KDM1A family), a histone demethylase, can block viral genome transcription and lytic replication of DNA viruses. For herpes virus, inhibition of LSD1 suppresses viral genome transcription and lytic replication, so that virus shedding and disease severity were reduced[3–5]. LSD1 is also found to activate HBV transcription by establishing an active hepatitis B viral chromatin state[6]. Moreover, inhibition of LSD1 activity suppresses the activation of HIV transcription in latently infected T cells by demethylation of viral Tat protein[7]. However, different from DNA viruses, RNA viruses do not rely on chromatin structure and histone capsid. As a result, the activity of LSD1 is expected to have little effect on RNA virus genome transcription via histone demethylation. Nevertheless, it is still possible that LSD1 influences RNA virus replication through demethylation of other host or viral proteins.
The interferon-inducible transmembrane (IFITM) gene family, belonging to a group of small interferon stimulated genes (ISGs)[8], has recently been identified to exhibit antiviral activities to a broad spectrum of viruses through blocking the early stages of viral life cycle[9, 10]. In humans, there are at least four functional members of IFITM proteins. IFITM1, 2 and 3 are expressed in a variety of human tissues and cell lines. IFITM5 is limited to the bone and is involved in mineralization[11]. Among these IFITM proteins, IFITM3, located mainly in endosome membrane, is well characterized for the restriction of RNA viruses such as IAV [12, 13]. IFITM3 protects the mice against morbidity and mortality due to influenza virus infection[12, 14], and in humans, disease severity in IAV-infected patients are correlated with the C allele (SNP rs12252) of human IFITM3[15]. Recently, our work showed that IFITM3 is mono-methylated on lysine residue of K88 in conserved intercellular loop (CIL) domain to reduce its antiviral activity[16, 17]. We therefore search for the corresponding demethylase to stimulate IFITM3 activity.
IFITMs reside in the cytosolic leaflet of plasma or endosome/lysosome membranes with the help of two intermembrane domains (IM1 and IM2)[18, 19]. Between IM1 and IM2 is the CIL domain, and on side of IM1 and IM2 are the cytosolic N-terminus and C-terminus respectively[19, 20]. The CIL domain is exposed to cytosol and is intensively regulated by post-translational modifications. The conserved cysteine residues in CIL domain were reported to be palmitoylated, which is required for protein clustering of IFITMs on the membrane and its antiviral function[21]. Ubiquitination on any of the four lysines in the N-terminus and CIL of IFITM3 negatively regulated its antiviral activity by accelerating degradation of IFITM3[22]. Through mass spectrometry assay, we have identified a monomethylation site at K88 located in CIL domain of IFITM3. Methylation of IFITM3 is catalyzed by SET7 to inactivate antiviral activities of IFITM3[16].
In this study, the lysine specific demethylase (LSD1) is identified to demethylate the K88me1 modification of IFITM3. The interaction strength between LSD1 and IFITM3 positively correlates with the antiviral activity of IFITM3. We further verified that IFNα treatment triggers demethylation of IFITM3 by LSD1 to be antiviral active. We showed that RNA viruses like IAV and VSV have developed strategies to counteract IFITM3 by disassociating LSD1 from IFITM3. However, surprisingly, Zika virus, which may not yet well adapt to human, could not counteract IFITM3 by methylation. Unlike herpes-virus infected animals, inhibition of LSD1 in IAV-infected mice aggravates the disease severity and increases the levels of IFITM3-K88me1, suggesting that LSD1 act differently in DNA and RNA virus infections. The action of LSD1 to demethylate IFITM3 is essential for the host to fight against RNA viruses.
LSD1 was selected to test its demethylase activity on IFITM3, because it can antagonize SET7 for protein methylation[23]. We first tested whether LSD1 is recruited to IFITM3 under IFNα treatment. For this, HA-tagged IFITM3 and FLAG-tagged LSD1 were over-expressed in HEK293T cells followed by immunoprecipitation (IP) using anti-HA antibody to pull down IFITM3. As shown in Fig 1A, there is an intrinsic interaction between LSD1 and IFITM3 at 0h of IFNα treatment. Subsequently, clearly enhanced interaction between IFITM3 and LSD1 was detected at 6h post-IFNα-treatment. This increased interaction was also observed with endogenous IFITM3 and LSD1 upon IFNα treatment (Fig 1B), indicating that LSD1 responds to IFNα simulation via binding to IFITM3. As a consequence, the levels of IFITM3-K88me1 were reduced over time under IFNα treatment when more LSD1 binds to IFITM3 (Fig 1B, Input). Moreover, we confirmed that there was a physical interaction between IFITM3 and LSD1 in vitro by MBP pull-down assay (Fig 1C) which may explain the co-localization between LSD1 and IFITM3 has shown in S1 Fig.
To validate the demethylase activity of LSD1 on IFITM3, different doses of LSD1 were co-transfected with IFITM3 to examine the IFITM3-K88me1 levels. The data showed that LSD1 could demethylate IFITM3 at K88 in a dose-dependent manner (Fig 1D). Furthermore, the demethylase activity of LSD1 was antagonized with SET7 because IFITM3-K88me1 catalyzed by SET7 was largely reduced in the present of LSD1 (Fig 1E).
To test the effect of LSD1 on IFITM3 activity towards virus infection, HEK293T cells were transfected with IFITM3 and LSD1 and were then infected with VSV. As expected, the replication of VSV (expressed by vRNA copies of VSV L gene) was distinctly reduced when LSD1 was co-expressed with IFITM3 in a dose-dependent manner (Fig 2A). To verify the effect of LSD1 on virus replication in a more physiological condition, HEK293T cells were transduced with lentivirus delivering control shRNA (shCK) or shRNA targeting either SET7 (shSET7) or LSD1 (shLSD1). After two days, cells were treated by IFNα (100U/ml) for additional 24h to induce endogenous expression of IFITM3, and were then infected with VSV. Cells were collected 12h post-infection, and the gene expressions of IFITM3, LSD1, and SET7 as well as the levels of IFITM3-K88me1 were examined (Experimental procedure shown in Fig 2B). As expected, shSET7 and shLSD1 efficiently knocked down gene expression levels of SET7 and LSD1 respectively; while the expression of IFITM3 was not affected (Fig 2C). The levels of IFITM3-K88me1 were up-regulated by shLSD1 but down-regulated by the shSET7 correspondingly (Fig 2D). Consisted with the levels of IFITM3-K88me1, shLSD1 caused augments in VSV vRNA levels while shSET7 remarkably reduced VSV vRNA levels (Fig 2E). To explore whether the effect of LSD1 and SET7 on VSV infection is dependent on IFITM3 activity, IFITM3 and LSD1 or IFITM3 and SET7 were knocked down together under the same experimental protocol (S2 Fig). The results in Fig 2F showed that, neither SET7 nor LSD1 had effects on VSV replication in the absence of IFITM3. The data implied that anti-viral activity of IFITM3 against VSV was dependent on the methylation levels of IFITM3-K88me1 and under the regulation of IFITM3 methylation modification enzymes of SET7 and LSD1.
To examine the broad-reactive activity of IFITM3 against virus infections, A549 cells were infected by influenza A virus, A/WSN/33 (H1N1) (WSN), with a similar experimental procedure (shown in Fig 3A). As expected, shSET7 and shLSD1 efficiently knocked down gene expression levels of SET7 and LSD1 respectively; while the expression of IFITM3 was not be affected (Fig 3B). Transduction of shSET7 led to the reduced levels of IFITM3-K88me1, wherein there was also less expression of viral NP proteins (Fig 3C). By contrast, the increased levels of IFITM3-K88me1 by shLSD1 resulted in elevated expression of viral NP proteins (Fig 3D). The expressions of vRNA, cRNA and mRNA of viral segment 5 (NP) were suppressed by shSET7 but stimulated by shLSD1 (Fig 3D). Again, when IFITM3 was knocked down together with LSD1 or SET7, both SET7 and LSD1 lost their effects on influenza A virus replication (S3 Fig and Fig 3E).
Altogether, the data suggested that LSD1 positively regulated the antiviral activities of IFITM3 via down-regulating IFITM3-K88me1. On the contrary, SET7 negatively regulated the antiviral activities of IFITM3 via up-regulating IFITM3-K88me1.
Currently, it remains unknown how IFITM3-K88me1 and its methylation enzymes affect disease outcome in vivo. By identifying the methyltransferase SET7 and the demethylase LSD1 of IFITM3, we were able to take the advantage of chemical inhibitors towards these enzymes to study the role of IFITM3-K88me1 in IAV-infected animals. A LSD1 inhibitor, Trans-2-phenylcyclopropylamine hydrochloride (TCP), was applied intraperitoneally to mock-infected or WSN-infected mice daily. A dose of 10000 pfu viruses were used to infect mice to guarantee a lethal infection. As shown in Fig 4A and Fig 4B, a mild body-weight loss was observed in TCP-treated mock group indicating limited toxicity of TCP. Compared with PBS-treatment, infected mice treated with TCP showed faster body weight loss (Fig 4A). The survival curve in Fig 4B clearly indicated that TCP treatment accelerated the death of IAV-infected mice. To examine the effect of TCP under the infection of a nature isolate, mice were infected again with A/Sichuan/1/2009 (H1N1) (SC09, a pandemic H1N1 stain) at a sublethal dose (300 pfu) followed by TCP or PBS treatment. The data in Fig 4C and Fig 4D showed that the PBS-treated infected mice recovered from body weight loss at day 8 and all survive, but none of the TCP-treated infected mice could survive more than 7 days. However, compared with wild-type mice, it was worth noting that there was no effect of TCP in IFITM-knockout mice in IAV infection (S4 Fig).
To access the histological data, mice were mildly infected with WSN virus (500 pfu) wherein all the groups of animals could survive till at least day 9 post-infection. Mice were admitted with 10mg/kg TCP to magnify the effect of TCP. The lungs of the animals were isolated from each group for further analysis. As shown in Fig 5A, the structures of lung in both PBS- and TCP-treated mock-infected mice kept intact, while, the lung in TCP-treated WSN-infected mice had more bleeding and swelling as compared to PBS-treated WSN-infected mice. Histological data in Fig 5B indicated that TCP treatment in mock-infected mice did not induce observable toxicity. On the contrary, TCP treatment in WSN-infected mice induced marked lung pathology with massive infiltrating cells and hemorrhage in lung.
The data in all indicated that TCP treatment accelerated the disease course leading to more severe lethality and lung damage.
The lung tissues from sublethal infected mice were further collected to detect the level of IFITM3-K88me1 through Western blot. As compared to PBS-treated groups, the levels of IFITM3-K88me1 were significantly higher under TCP treatment, and IFITM3-K88me1 gradually increased during the course of the disease from Day 1 to Day 9 (Fig 6). Moreover, to establish the relationship of TCP and IFITM3, we used IFITM-knockout mice infected with sublethal infection of SC09 pandemic virus. The results showed that IFITM-knockout mice are more susceptible to infection and dead under low infection dose of virus (S4 Fig). The TCP treatment in these mice did not show further aggravation of disease neither in body-weight lost nor in lethality as compared to PBS treatment (S4 Fig). As IFITM3 makes a substantial contribution than any other IFITMs in IfitmDel knockout mice to IAV resistance in vivo[14], these data suggested that TCP-induced disease aggravation was associated with IFITM3 activity and high levels of IFITM3-K88me1.
As we observed that the IFITM3-K88me1 levels increased gradually in WSN-infected mice even without TCP treatment, we suspected that the virus may trigger IFITM3 methylation by disassociating LSD1 from IFITM3. To test this, HEK293T cells were transfected with HA-IFITM3 and FLAG-LSD1, and then infected with WSN virus. The data in Fig 7A showed that WSN infection heavily disassociated the interaction between IFITM3 and LSD1 in a time-dependent manner. This phenomenon was also observed in VSV-infected cells (Fig 7B). At 12h post-VSV-infection, the interaction between LSD1 and IFITIM3 was reduced compared with 0h post-infection (Fig 7B). Together with our previous findings that IAV or VSV infection increased the levels of IFITM3-K88me1 gradually upon infection[16], the data in all suggested that virus may develop strategies to comprise the action of type I IFN by inhibiting the binding of LSD1 to IFITM3.
However, when we tested Zika virus, the recently re-emerged arbovirus associated with microcephaly, we found that the levels of IFITM3-K88me1 were reduced upon infection even though the expression levels of IFITM3 increased (Fig 8A). Correspondingly, increased LSD1 expression and decreased SET7 expression were detected under Zika infection (Fig 8A). In addition, IFITM3-associated LSD1 levels did not decrease as fast as in IAV and VSV infections indicating that Zika virus did not disrupt LSD1/IFITM3 interaction as efficiently as VSV and IAV (Fig 8B). Nevertheless, knocking down LSD1 by shRNAs enhanced virus replication indicating that LSD1 is still anti-viral active in Zika-infected cells (Fig 8C). The result suggested that Zika virus was not able to counteract IFITM3 antiviral activity though methylation, and therefore could be more sensitive to IFITM3 as compared to IAV or VSV.
IFITM3 is an ISG gene that presents broad-reactive anti-RNA virus activities[24]. It has been known post-translational modifications of IFITM3 would regulate its action against virus infection[25]. However, little is known about the responsible host enzymes catalyzing these post-translational modifications on IFITM3 and how important these modifications are in host protection against virus infection. In the current study, we reveal that LSD1 is the lysine demethyltransferase of IFITM3 at K88, which positively regulates the anti-RNA virus activity of IFITM3. Namely, TCP, a LSD1 inhibitor triggers more severe disease outcome in sub-lethal challenged IAV-infected mice, resulting in severe body weight loss, lethality and lung injuries. These data indicate that LSD1 contributes to control a mild infection through activating IFITM3 by K88 demethylation.
Lysine 88 on IFITM3 is located in the CIL domain which is exposed to the cytoplasma, so that it is highly accessible for cellular lysine modification machinery. We identified here and previously that K88 was methylated endogenously, which negatively regulates IFITM3 anti-viral activity[16]. Methylation on the lysine residue would neutralize the positive charge of the lysine motif, and therefore may change the protein-protein interactions. It is proposed that the IFITM3 complex network formed by polymerization between IFITM3 molecules may reduce the membrane fluidity which limits virus fusion with the host cell membrane[26]. It is possible that methylation on CIL domain may interfere with IFITM3 complex to loosen the IFITM3 network, so that fusion between cellular membrane and viral membrane is more feasible. Methylation may also change the interaction between IFITM3 and other host proteins to inhibit its anti-viral activity.
Based on our previous data and current results, we suggest that LSD1 and SET7 is a pair of enzymatic modulators to regulate K88 methylation of IFITM3. The interaction between IFITM3 and LSD1 is found to be enhanced by IFNα, which in turn triggers K88 demethylation. As mRNA levels of LSD1 were stable under IFNα treatment (S5 Fig), the decrease of IFITM3-K88me1 is a response to IFNα instead of a simply consequence of increased abundance of LSD1. On the contrary, when cells are infected by RNA virus, such as VSV and IAV, the level of IFITM3-K88me1 was increased associated with less interaction between LSD1 and IFITM3. Therefore, IFNα signals strengthen host restriction factors through demethylating K88me1 on IFITM3 by LSD1, and virus infection counteract host restriction factors by promoting methylation of IFITM3 at K88. A balance between the lysine methylation and demethylation of IFITM3 at K88 may exist to maintain the cell homeostasis, and this balance may be broken under virus infection. A schematic working model was proposed and shown in Fig 9.
Zika virus has emerged as a severe health threat with a rapidly expanding range. IFITM3 is reported to be able to inhibit Zika virus infection in the early viral replication cycle[27]. In our study, Zika virus seems more susceptible to IFITM3-mediated host restriction because it cannot neutralize the anti-viral activity of IFITM3 through methylation. It is possible that further adaptation is required for Zika virus to counteract IFITM3, and viruses like IAV and VSV which circulate for a long history in humans had already well adapted to develop anti-IFITM3 strategies to methylate IFITM3.
To date, several studies have shown that post-translational modifications on IFITM3 affect its anti-viral activity in infected cells[25]. However, no evidence has been previously shown to demonstrate the important role of these modifications in infected animals in vivo. Our previous study has unexpectedly shown that IFITM3 may not limit but rather promote the progression of DNA virus such as HCMV by facilitating the formation of the virion assembly compartment during infection in cell culture[28]. Other studies have also shown in vivo that LSD1 inhibitor could be protective for host against DNA virus such as HSV infection[3, 4]. In this study, we have found that by opposition to its effect in anti-DNA virus infection, the enzymatic activity of LSD1 is critical to restrict RNA virus infection by controlling the limited levels of IFITM3 K88me1 in infected tissues. We treated WSN-infected mice with a LSD1 inhibitor (TCP) to investigate the role of IFITM3-K88me1 in host anti-viral response, and found that TCP treatment resulted in aggravated pathology as observed by severe body weight loss and increased mortality under IAV infection. The level of methylation of IFITM3 at K88 in the lung tissues of TCP-treated mice was also much higher than that in PBS-treated infected mice. Our results suggest that methylation of IFITM3 plays an essential role in disease development in IAV-infected animals. To regulate methylation, such as to stimulate the expression and upregulation of LSD1, may provide a new strategy for anti-flu therapeutics. On the other hand, as TCP strongly inhibits IFITM3-mediated antiviral activity by targeting LSD1, it could be potentially applied as an adjuvant to increase RNA virus propagation in IFITM3-competent system so as to assist vaccine industry or infection models.
The animal experiments were approved by the Institutional Animal Care and Use Committee of the Institut Pasteur of Shanghai, Chinese Academy of Sciences (Animal protocol #A2015006). All animal care and use protocols were performed in accordance with the Regulations for the Administration of Affairs Concerning Experimental Animals approved by the State Council of People’s Republic of China.
Female 6- to 8-week-old BALB/c mice were purchased from model animal center of Shanghai Life Science. IFITM-/- mice were purchased from EMMA (IfitmDel) wherein the entire Ifitm1, 2, 3, 5, and 6 locus were deleted[29].
Human neural precursor cells (hNPC)[30] was cultured in N2B27+bFGF medium composed of DMEM/F12 and Neuralbasal in a ratio of 1:1 supplemented with 0.5% N2 (Gibco), 1% B27 (Gibco), 1%NEAA (Gibco), 1% L-Glu (Gibco), and 20ng/ml bFGF (Gibco). Other cell lines used in these studies were cultured in Dulbecco’s modified Eagle’s medium (DMEM) (HyClone) supplemented with 10% fetal bovine serum (FBS) (ExCell Bology), 1% penicillin/streptomycin (Gibco). HEK293T cells (American Type Culture Collection) were transfected with appropriate plasmids using polyethylenimine (PEI) reagent (Polysciences) according to the manufacturer’s instructions. The plasmids del8.9 and VSV-G for lentivirus production were gifts from Prof. Ke Lan (Institut Pasteur of Shanghai). The full sequences of IFITM3, SET7 and LSD1 were amplified by PCR with cDNA from Hela cells (American Type Culture Collection) or HEK293T cells. These fragments were then cloned into HA-tagged, FLAG-tagged or non-tagged (pcDNA3.1) vectors by standard procedures. The IFITM3-K88R and -K88A mutants were constructed with the Quickchange II site-directed mutagenesis kit (Stratagene) according to the manufacturer’s standard procedures. shIFITM3-1, shLSD1-1, shLSD1-2, shSET7-1, shSET7-2 and shCK were constructed in pLKO.1 using the following targeting sequence:
The antibodies used in this study were as follows: anti-HA (F-7, Santa Cruz Biotechnology), anti-FLAG (M2, Sigma-Aldrich), anti-LSD1 (Cell Signaling), anti-IFITM3 (11714-1-AP, Protein Tech Group, Inc.), anti-β-Actin (C1213, Sungene Biotech), anti-α-tubulin (DM1A, Sigma), anti-Influenza A NP (9G8, Sc-101352, Santa Cruz Biotechnology). Alexa Fluor anti-mouse 555 (Invitrogen-Molecular Probes), anti-rabbit 488 (Invitrogen-Molecular Probes) and anti-rat 633 (Invitrogen-Molecular Probes). The αK88me1 rabbit polyclonal antibody was generated by Abmart, raised toward the polypeptide SRDR (Kme1) MVGD. Protein A/G PLUS Agarose beads (A10001) were purchased from Santa Cruz Biotechnology. IFNα (recombinant human interferon α-2b) was purchased from Shanghai Hua Xin High Biotechnology Inc. The LSD1 inhibitor (TCP) was purchased from Sigma.
Cells were collected and lysed in TRIzol Reagent (Invitrogen) and RNA was isolated according to the manufacturer’s instructions. Reverse transcription was performed with the PrimeScript RT regent kit (TaKaRa). The cDNA samples were used at 10ng/well in a 384-well plate and run in triplicate. PCR reactions were set up in 10μl volumes with SYBR Premix Ex Taq regent (TaKaRa) on an ABI Prism 7500 Sequence Detection System. GAPDH was used as the reference control for the target genes. The primers were listed below:
The relative gene expression is calculated by 2−ΔΔCT, where ΔΔCT = (CT,Target − CT,Actin)EG– (CT,Target − CT,Actin)CG, EG represents experimental group and CG represents control group.
The shRNA sequences were synthesized by Shanghai Sunny Biotechnology Co.Ltd. Lentivirus package plasmids delta8.9 and VSV-G were co-transfected into HEK293T cells. Forty-eight-hour post-transfection, viral supernatants were harvested after centrifugation and filtration for downstream use.
Cells were harvested and washed with ice-cold PBS and lysed in radio immunoprecipitation assay (RIPA) buffer (50 mM Tris/HCl [pH 7.4], 150mM NaCl, 1% Triton X-100, 0.5% sodium deoxycholate, 0.1% SDS, 1mM EDTA, 10% glycerol with 1 mM PMSF, 1 mM Na3VO4, 1 mM NaF and protease inhibitor [1:100, P8340, Sigma-Aldrich]) for 30min on ice. The lysates were incubated with antibody or affinity beads as indicated overnight at 4°C. The immunoprecipitations were separated by SDS-PAGE and analyzed by immunoblotting.
Hela cells grown on glass cover-slips were fixed in 2% paraformaldehyde (in phosphate-buffered saline) for 20min at room temperature and permeabilized with 0.1% Triton X-100 (in phosphate-buffered saline) for 15min at room temperature. Subsequently, the cells were blocked with 5% FBS (in phosphate-buffered saline, PBS) for 20min and incubated with the primary antibodies indicated below for 60min at room temperature. Primary antibodies to the following (and their dilutions) were used: anti-Myc (Mouse, 1:500), anti-HA (Rabbit, 1:500), and anti-FLAG (Rat, 1:1000). Cells were then washed with PBS and stained with Alexa Fluor-labeled anti-mouse, anti-rabbit and anti-rat secondary antibodies (1:1000;) for 20min at room temperature in dark and washed five times with 0.1% Triton X-100 in PBS after each antibody treatment. To visualize the nuclei, cells were counterstained with DAPI (4′, 6-diamidino-2-phenylindole; 1:5000; Beyotime) for 10min. Finally, labeled cells were mounted on slides with Mowio reagent (4–88, Chem. Bochem) overnight. Images were captured using a Leica TCS SP5 confocal laser scanning microscope.
MBP, MBP-LSD1, and GST-IFITM3 were expressed in Rosetta/pLysS and purified by amylose resin or glutathione-Sepharose 4B (GE Healthcare). After dialysis, proteins were quantified by SDS/PAGE using BSA as the standard. MBP, GST-IFITM3, MBP-LSD1, and GST-IFITM3 were incubated for 2h at 4°C in buffer A [20mM Tris·HCl (pH 7.5), 200 mM NaCl, 5 mM β-mercaptoethanol, and 1 mM PMSF] and were then incubated with amylose resins for another 2h. Amylose resins were then washed extensively with buffer A, followed by SDS/PAGE.
VSV (Indiana serotype, American Type Culture Collection). Recombinant VSV expressing green fluorescent protein (VSV-GFP)[32] was propagated in Vero cells after infection at a MOI = 0.01. Forty-eight hours later, supernatant was collected after centrifugation and then filtered through a 0.45μm filter (Millipore) and stored at -80°C. Virus infections were performed in the absence of serum for 1h, and then replaced with fully supplemented growth medium. Endpoint dilution assay was used to quantify VSV titers with log dilutions on 100% confluent Vero cells plated in 96-well plates.
Recombinant viruses were generated by plasmid-based reverse genetics followed by two-round plaque purification and propagation in MDCK cells (American Type Culture Collection) as described previously[33]. Briefly, eight viral cDNAs (pHW2000-WSN-NP, pHW2000-WSN-PB1, pHW2000-WSN-PB2, pHW2000-WSN-pA, pHW2000-WSN-NS, pHW2000-WSN-M, pHW2000-WSN-HA, pHW2000-WSN-NA) from A/WSN/33 (H1N1) were gifts from Hans-Dieter Klenk (University of Marburg), which were co-transfected into HEK293T cells. Supernatant was then transferred onto MDCK cells 72h post-transfection. Virus multiplied from MDCK cells was further purified by a two-round plaque assay followed by multiplication. To examine the broad-reactive activity of IFITM3 against virus infections, A549 cells (American Type Culture Collection) were transfected with shRNAs followed by virus infection at MOI = 5 for 30min in PBS 0.2% BSA. After 3×washes with PBS, infected cells were then cultured in complete DMEM culture media supplemented with 2% FBS. Cell extracts were collected 8h post-infection and analyzed by Western blot and qPCR.
Zika virus SZ-WIV001 (KU963796) was kindly provided by Wuhan institute Virology CAS[34]. Zika virus was amplified in VeroE6 cells (American Type Culture Collection). Cells were infected with Zika virus in the absence of FBS for 1h. Media were then removed, and substituted by the appropriate culture medium, supplemented with 3% FBS and 20mM HEPES. Infected cell cultures were observed daily for detection of cytopathic effects (CPE, day 3–4 post-infection). Supernatants were collected when CPE occurs, stored at -80°C after centrifugation for virus stock. Human neural precursor cells (hNPC) were infected with Zika virus at MOI = 1, cell extracts are collected 24h, 48h and 72h post-infection and analyzed by Western Blot. Huh7 cells were first transfected with IFITM3 and LSD1 or transducted with lentivirus derived shRNAs and then infected with Zika virus at MOI = 1. Cell supernatants were further collected to titrate virus RNA copies by real-time PCR.
Female 6- to 8-week-old BALB/c mice were randomly divided into 4 groups: PBS-mock group, TCP-mock group, PBS-WSN treated group, and TCP-WSN treated group.
A/WSN/33 of 10000 pfu (lethal dose) was given in 50μl PBS per mouse intranasal. The body weights of mice were monitored throughout the infection time course. After sublethal infection (500 pfu), mice would be sacrificed at Day 1, Day 3, Day 6, and Day 9 respectively to collect lung tissues for the detection of IFITM3-K88me1 through Western blotting method and of residue virus titers through plaque titration on MDCK cells. At Day 9, the lung tissues were further fixed in neutral formalin, embedded in paraffin, sectioned, and stained with hematoxylin and eosin (H&E).
A/Sichuan/1/2009 (H1N1) (SC09) virus (kindly provided by Prof. Yuelong Shu in China CDC) were also used to infect the mice at 300 pfu, and the body weights of mice were monitored throughout the infection time course.
Data were analyzed by Student’s t test. A P value of < 0.05 was considered to be statistically significant difference. ns, p >0.05; *, p <0.05; **, p <0.01; ***, p <0.001. Statistical analyses were done using GraphPad Prism. Error bars represent standard deviation (s.d.).
|
10.1371/journal.pgen.1000015 | DNA Damage Activates the SAC in an ATM/ATR-Dependent Manner, Independently of the Kinetochore | The DNA damage checkpoint and the spindle assembly checkpoint (SAC) are two important regulatory mechanisms that respond to different lesions. The DNA damage checkpoint detects DNA damage, initiates protein kinase cascades, and inhibits the cell cycle. The SAC relies on kinetochore-dependent assembly of protein complexes to inhibit mitosis when chromosomes are detached from the spindle. The two checkpoints are thought to function independently. Here we show that yeast cells lacking the DNA damage checkpoint arrest prior to anaphase in response to low doses of the DNA damaging agent methyl methane sulfonate (MMS). The arrest requires the SAC proteins Mad1, Mad2, Mad3, Bub1, and Bub3 and works through Cdc20 and Pds1 but unlike the normal SAC, does not require a functional kinetochore. Mec1 (ATR) and Tel1 (ATM) are also required, independently of Chk1 and Rad53, suggesting that Mec1 and Tel1 inhibit anaphase in response to DNA damage by utilizing SAC proteins. Our results demonstrate cross-talk between the two checkpoints and suggest that assembling inhibitory complexes of SAC proteins at unattached kinetochores is not obligatory for their inhibitory activity. Furthermore, our results suggest that there are novel, important targets of ATM and ATR for cell cycle regulation.
| Genome integrity is assured, in part, by regulatory systems called “checkpoints” that assure that cells do not improperly progress through the cell cycle. The DNA damage checkpoint assesses the status of DNA replication and inhibits cell cycle progression when the cell makes mistakes in DNA replication or when the cell has been assaulted by a DNA damaging agent from the environment. The checkpoint allows the cell time to repair the DNA and then permits the cell cycle to resume. There is a separate “spindle checkpoint” that monitors whether chromosomes are properly attached to the spindle and if so, allows cells to proceed through mitosis. The DNA damage checkpoint and the spindle checkpoint assure that daughter cells receive the correct number of chromosomes that are identical in DNA sequence. Here we show that the two checkpoints are not independent but that they cooperate to restrict mitotic progression in the face of DNA damage. We show that the spindle checkpoint can be induced by DNA damage and that there is a novel kinetochore independent mechanism to activate the spindle checkpoint proteins. In addition, we implicate the ATM and ATR kinases as kinetochore-independent activators of the spindle checkpoint.
| Two evolutionarily conserved checkpoints, the DNA damage checkpoint and the spindle assembly checkpoint (SAC), control the fidelity of chromosome segregation. The DNA damage checkpoint responds to a variety of DNA lesions and controls entry into S phase, completion of S phase and entry into mitosis [1],[2]. The DNA damage checkpoint is a signal transduction network consisting of sensors, signal transducers and downstream effectors. Central to the signal transduction network in budding yeast are two phosphotidylinositol 3’ kinase-like kinases (PIKKs), Mec1 (the yeast homolog of ATM and Rad3-related protein, abbreviated ATR) and Tel1 (the yeast homolog of the ataxia-telangiectasia-mutated protein abbreviated ATM) [1],[3],[4]. Mec1 and Tel1 activate the protein kinase transducers Rad53, Chk1 and Dun1 leading to cell cycle arrest and induction of DNA repair genes [5]–[9].
The SAC responds to chromosomes that are either unattached from the spindle or are not under tension and delays the metaphase to anaphase transition [10]. The kinetochore has an integral role in the SAC and a popular model is that the kinetochore initiates checkpoint signaling by being the site of assembly of inhibitory complexes of SAC proteins that inhibit mitosis [10],[11]. The inhibitory complexes are made up of combinations of the evolutionarily conserved proteins Bub1 Bub3, Mad1, Mad2, and Mad3 (BubR1 in higher cells) but the exact details of their assembly and inhibitory activities are unknown [12]–[15]. The two checkpoints share a common target to regulate mitosis. Pds1 (securin in higher organisms) is an anaphase inhibitor that is stabilized by two different mechanisms when the two checkpoints are activated. Pds1 is phosphorylated and thereby stabilized by the DNA damage checkpoint [16]. The SAC stabilizes Pds1 by inhibiting Cdc20, the specificity factor for an E3-ubiquitin ligase called the anaphase-promoting complex (APC) that is responsible for the proteolysis of Pds1 [17],[18].
There are indications, from yeast to humans, that the DNA damage checkpoint and the SAC have overlapping functions. Laser microbeam-induced DNA damage during late prophase in some human cell lines delays progress through metaphase in a P53-independent manner and the delay is abrogated by inhibiting Mad2 [19]. Cells derived from a mouse mutant, heterozygous for a deletion of BubR1, are defective in the response to genotoxic agents suggesting that BubR1 is limiting in the DNA damage response [20]. Drosophila grapes mutants (grp), lacking the homolog of Chk1, delay anaphase after X-irradiation and the delay is dependent on BubR1 [21]. Camptothecin induces a mitotic delay in fission yeast cells lacking the DNA damage checkpoint and the delay requires Mad2 [22]. In addition, fission yeast Mad2 plays a minor role in the mitotic delay imposed by growing cells in the presence of the ribonucleotide reductase inhibitor hydroxyurea (HU) but Mad1, Bub1 and Mad3 do not play a role [23]. Budding yeast cells lacking the DNA damage checkpoint (rad9 rad24 double mutants) and compromised for DNA replication by mutations in cdc2-1, pol1-17, mcm2-1,or mcm3-1 delay in mitosis in a Mad2-dependent fashion [24]. Compromising both DNA replication and the DNA damage checkpoint in orc1-161 rad53-11 cells causes a delay in mitosis in a Mad2 and Bub1-dependent manner [25]. The DNA alkylating agent, methyl methane sulfonate (MMS), HU, and ultraviolet light also induces a mitotic delays in cells lacking the DNA damage checkpoint and the delays require Mad1 and Mad2 [24],[26]. Models to explain why such diverse mutants and treatments cause a SAC-dependent mitotic delay propose that kinetochores may be damaged or poorly assembled due to aberrant centromere DNA replication or defects in sister chromatid cohesion may result in a loss of tension across sister kinetochores [23]–[27]. These models are in accord with the proposition that the SAC signal is generated at kinetochores that are either detached from the mitotic spindle or from kinetochores that are on chromatids lacking tension, as would be caused by defective cohesion [10], [11], [28]–[31]. However, explanations invoking a role for the kinetochore in a DNA damage response are harder to reconcile with observations that double strand DNA breaks near telomeres in yKu70Δ cells or a single double strand break induced by HO at URA3 induces a mitotic delay in cells lacking the DNA damage checkpoint [32],[33]. It was proposed that telomere proximal double strand breaks in cells lacking Yku70 results in dicentric chromosomes that are known to activate the SAC, presumably by altering tension at kinetochores [32]. The single double strand break introduced at URA3 causes a delay in the second cell cycle after HO induction which may also reflect the formation of dicentric chromosomes as the source of the SAC signal [33].
In this study we test the model that the kinetochore is required to activate the SAC proteins in response to DNA damage. We show that cells arrest prior to anaphase when grown in the presence of MMS and that the arrest requires the SAC proteins Mad1, Mad2, Mad3, Bub1 and Bub3. Surprisingly, temperature-sensitive ndc10-1 cells that are devoid of kinetochores also arrest in response to MMS suggesting that the kinetochore is not required to convert the SAC proteins into inhibitors under these conditions. We show that the downstream effectors of the SAC (Cdc20 and Pds1) are required for the arrest suggesting that the inhibition by the checkpoint proteins works through the canonical SAC. Furthermore, we show that the SAC is capable of restraining anaphase in response to MMS in cells lacking the DNA damage checkpoint and that the yeast homologs of ATM (Tel1) and ATR (Mec1) are required for the SAC-dependent arrest suggesting that the PIKKs are required to activate both the DNA damage checkpoint and the SAC. These studies reveal an intimate relationship between the DNA damage and SAC pathways and highlight the importance of preventing anaphase in cells with damaged chromosomes.
We applied several different assays to measure the mitotic delay in cells treated with MMS. Cells were arrested in G1 by growth in the presence of α-factor and then released to the cell cycle in the presence and absence of 0.01% MMS [24]. We monitored cell cycle progression by a combination of flow cytometry, cell morphology and Pds1 (securin) stability. Cells from four isogenic strains cycled normally in the absence of MMS as judged by DNA flow cytometry (Figure 1A, upper panels), cellular morphology (Figure 1B) and Pds1 stability (Figure 1C). MMS treated wild type and mad2 cells delayed progress though S phase, as determined by flow cytometry and arrested with a G2/M content of DNA (Figure 1A, lower panels), prior to anaphase (Figure 1B) with high levels of Pds1 (Figure 1C) due to activation of the DNA damage checkpoint. rad9 rad24 cells, lacking the DNA damage checkpoint, also delayed with a G2/M content of DNA when grown in the presence of MMS (Figure 1A, lower panel), failed to complete anaphase and accumulated as large budded cells with a single undivided nucleus (Figure 1B and Figure S2) and stabilized Pds1 (Figure 1C). The MMS-dependent mitotic delay was abrogated in rad9 rad24 mad2 cells that failed to accumulate with a G2/M content of DNA (Figure 1A, lower panel), progressed into anaphase (Figure 1B and Figure S2) and failed to stabilize Pds1 (Figure 1C). We measured reproducibility of the response by analysis of multiple flow cytometry profiles (Figure S1A–S1D). Each experiment was performed between 2–6 times and duplicates for each of the flow cytometry experiments are shown including the mean percentage of cells with the G2/M content of DNA determined from the flow cytometry profiles along with the variance in those data. The range of measurements is shown for experiments performed twice and the standard deviation was calculated and is indicated as error bars at each time point for experiments done more than twice. These data confirm that MMS treatment of rad9 rad24 cells lacking the DNA damage checkpoint cause a pre-anaphase delay that is dependent on Mad2 [24].
Haploid rad9 rad24 cells delayed with a G2/M content of DNA suggesting that they had arrested after S phase. We used Clamped Homogeneous Electric Field (CHEF) gels to analyze whole chromosomes in cells treated with MMS to determine if the synchronized cells completed DNA replication in response to MMS treatment. CHEF gels are used to separate large (yeast chromosome-sized) fragments of DNA by electrophoresis and are useful for karyotyping yeast cells [34]. In addition, they can be used to determine if DNA replication is complete as chromosomes from cells with unreplicated DNA either do not enter the gel and therefore bands are not present or the DNA appears as faintly staining bands with smeared appearances [35]–[37]. Untreated wild type, rad9 rad24 and rad9 rad24 mad2 cells had normal CHEF karyotypes with clearly identified chromosomes (Figure 1D). Wild type cells treated with the ribonucleotide reductase inhibitor hydroxyurea (HU) do not complete DNA replication and chromosomes do not enter the gel and were not detected (Figure 1D). Chromosome staining in cells grown in the presence of MMS was weak in both rad9 rad24 cells and rad9 rad24 mad2 cells and was similar to wild type cells grown in the presence of HU (Figure 1D). We detected some chromosomal staining with a smeared appearance in wild type cells grown in the presence of MMS (Figure 1D). We conclude that cells grown under our conditions of 0.01% MMS and that delayed with a G2/M content of DNA had completed the bulk of DNA replication but accumulated with lesions, most likely stalled or collapsed replication forks.
We assayed cell cycle progression in other SAC mutants to determine if all SAC proteins were required for the delay in response to MMS. Cells lacking the DNA damage checkpoint and either mad1 or mad3 proceeded normally through the cell cycle in the absence of MMS (Figure 2A, upper panels). The same cells did not accumulate with a G2/M content of DNA when grown in the presence of MMS (Figure 2A, lower panels) and reproducibility of the flow cytometry, as per Figure 1, is shown in Figure S3A and S3B. The rad9 rad24 mad1 and rad9 rad24 mad3 cells did not delay anaphase and completed nuclear division in the presence of MMS (Figure 2B and Figure S4). bub1 cells delayed with a G2/M content of DNA in the presence and absence of MMS (Figure 2A). However, bub1 cells failed to restrain anaphase and completed nuclear division slowly perhaps suggesting that they partially retain the delay (Figure 2A, 2B, and Figure S3C). Reproducibility for the flow cytometry of the bub1 cells is shown in Figure S3C. It was difficult to determine the response of rad9 rad24 bub3 cells by the same assay because of a high degree of inviability in the strain which made flow cytometry difficult to interpret. We assayed cell cycle progression by arresting cells in G1 with α-factor and allowed sufficient time for the viable cells to form mating projections. We released the cells and monitored the progression of only the cells with mating projections that subsequently budded and determined whether they completed nuclear division. Both treated and untreated cells completed nuclear division although MMS treated bub3 cells slowly entered into anaphase (Figure 2C). We conclude that bub3, like bub1, abrogates the delay.
The kinetochore is required for the SAC and is thought to act as a platform that recruits checkpoint proteins when microtubules are unattached and assembles them into novel complexes that inhibit mitosis [10],[11]. Temperature sensitive ndc10-1 cells are unable to assemble kinetochores and are unable to arrest in mitosis in response to nocodazole, a benzimidazole drug that depolymerizes microtubules [11],[38],[39]. Therefore ndc10-1 cells lack the SAC at the restrictive temperature. We synchronized haploid rad9 rad24 ndc10-1 cells with α-factor at 23°C, incubated the cells at 35°C for 1 hour to inactivate Ndc10 and then released the cells to allow them to progress through the cell cycle at the restrictive temperature. Chromosomes lacking kinetochores are unable to be segregated at mitosis and remain in the mother cell. DNA replication in the next cell cycle causes an increase in ploidy. ndc10-1 cells, untreated with MMS, completed S phase and had a 2C content of DNA and then proceeded to the next cell cycle and increased the ploidy producing cells with a 4C content of DNA (Figure 2A, upper panel, reproducibility shown in Figure S3D). Wild type cells cycled normally in the absence of MMS at 35°C and did not produce cells with a 4C content of DNA (not shown). Therefore, the ndc10-1 cells with a 4C content of DNA are the result of inactivating the kinetochore during the 1 hour incubation at 35°C. The same ndc10-1 cells delayed in the first mitosis when grown in the presence of MMS (Figure 2A, lower panel and Figure S3D). Therefore kinetochores are not required for SAC-dependent inhibition of anaphase in response to MMS.
The SAC prevents the metaphase-to-anaphase transition by inhibiting the ubiquitylation and degradation of Pds1 by the APC. The target of the SAC is the APC regulatory subunit Cdc20 [18],. We determined if MMS inhibits anaphase through APCCdc20 inhibition using CDC20-127; a dominant checkpoint-defective allele that produces a protein unable to bind Mad2 [40]. We generated CDC20-127 (CDC20Y205N) by site directed mutagenesis, confirmed it by DNA sequencing and replaced the endogenous allele by a one-step gene replacement. CDC20-127 and CDC20-127 rad9 rad24 cells were delayed with a G2/M content of DNA in the absence of MMS (Figures S5A and S5B, upper panels) and cells completed nuclear division (Figure 2D). Reproducibility is shown in Supplementary Figure S5. CDC20-127 cells delayed with a G2/M content of DNA when grown in the presence of MMS and delayed entry into anaphase (Figure S5A and Figure 2D, upper panel). In contrast, CDC20-127 rad9 rad24 cells, grown in the presence of MMS, did not delay with a G2/M content of DNA, failed to restrain anaphase (Figure S5B and Figure 2D, lower panel) and did not stabilize Pds1 (Figure S5C). We conclude that CDC20-127 abrogated the delay in response to MMS in rad9 rad24 cells. Therefore, MMS induces a delay in rad9 rad24 cells by promoting Mad2 binding to Cdc20 and inhibiting APCCdc20.
A hypomorphic top2-B44 mutant, with reduced activity of type 2 topisomerase, delays the onset of anaphase using SAC proteins independently of Pds1 suggesting a novel mitotic topoisomerase II checkpoint [42]. We assayed pds1 cells using the assay described above for bub3 cells to determine if rad9 rad24 cells treated with MMS utilize this novel pathway. Growth in the presence of MMS delayed anaphase of rad9 rad24 cells but not rad9 rad24 mad2 and rad9 rad24 pds1 cells (Figure 2E). Therefore the delay in response to MMS works through Cdc20 and Pds1 and is different from the one reported for partial topoisomerase inhibition.
The lack of a kinetochore requirement for Mad1, Mad2 and Mad3-dependent APCCdc20 inhibition was surprising because kinetochores are believed to be the source of the signal that activates the SAC [43]–[46]. One possibility for how the SAC proteins respond to DNA damage, independently of the kinetochore, is that they become activated in a DNA damage-dependent manner. We analyzed mec1 and tel1 mutants to determine if there was a role of either protein in transducing the signal from DNA damage to the SAC proteins. MEC1 encodes a PIKK that is homologous to the human ATR and is a central transducer of the checkpoint response in yeast [1],[3]. TEL1 encodes the related PIKK homologue ATM and plays a lesser role in the DNA damage checkpoint in yeast. mec1-1 cells, grown in the presence of MMS, arrested with a G2/M content of DNA (Figure 3A). Similarly, rad9 rad24 tel1 cells delayed with a G2/M content of DNA in response to MMS (Figure 3A) suggesting that the delay is independent of Mec1 and Tel1. We constructed a mec1 tel1 double mutant to determine if the kinases contributed redundantly in activating the SAC. Only 60% of the mec1 tel1 cells were viable which precluded analysis by flow cytometry. We used the same assay as described above for bub3 and pds1 cells to determine the effect of MMS in mec1 tel1 cells. Wild type and mec1 cells arrested prior to anaphase when grown in the presence of MMS but mec1 mad2 cells completed nuclear division (data not shown). Therefore mec1 cells, like rad9 rad24 cells, arrest in mitosis in a Mad2-dependent fashion in response to MMS. Interestingly, mec1 tel1 cells were unable to arrest and completed nuclear division when grown in the presence of MMS (Figure 3B). Together, these data suggest that Mec1 and Tel1 act redundantly to activate the SAC proteins and inhibit APCCdc20 in response to MMS.
It is possible that the effects of Mec1 and Tel1 on the SAC were indirect. The single mutants lacking either Mec1 or Tel1 may retain sufficient PIKK activity to activate the downstream effector kinases Rad53 and Chk1 and contribute to the pre-anaphase G2/M delay. Perhaps cells lacking both Mec1 and Tel1 do not activate Rad53 and Chk1 and in their absence the SAC is unable to restrain anaphase. This is an important distinction because it would affect the interpretation that the SAC is activated in a Mec1 and Tel1-dependent fashion. The MEC1 gene is essential and mec1-1 cells are viable in the presence of a second mutation, sml1, that suppresses the mec1-1 lethality but does not suppress the DNA damage checkpoint phenotype. We used the same assay as described above for bub3, pds1 and mec1 tel1 cells to determine if there was a an effect of MMS on mitotic progression in a set of isogenic strains lacking Sml1 and proteins of the DNA damage checkpoint and the SAC. The sml1 cells, treated with MMS, behaved like wild type cells (Figure 1A) and arrested in mitosis prior to anaphase in contrast to the mec1 tel1 sml1 cells described above (Figure 3B). rad9 mrc1 sml1 cells that lack the S-phase checkpoint delayed prior to anaphase when grown in the presence of MMS (Figure 3B). rad53 chk1sml1 cells also delayed prior to anaphase when grown in the presence of MMS although a small percentage of cells entered into anaphase. However, the delay in rad53 chk1sml1 cells was abrogated by deleting MAD2 (rad53 chk1 mad2 sml1) as shown in Figure 3B. Therefore a partially activated DNA damage checkpoint is not sufficient to explain the entire pre-anaphase delay in MMS treated rad9 rad24 cells. We conclude that the SAC is sufficient to delay anaphase in the absence of the DNA damage checkpoint and that the SAC is activated in a Mec1 and Tel1 dependent fashion. An important study has recently shown that there is PIKK-dependent phosphorylation of SAC proteins in response to DNA damage in human cells suggesting that SAC proteins are substrates of ATM and ATR in response to DNA damage [47]. Together the data suggest that there may be an evolutionarily conserved response of cells to DNA damage that involves ATM and ATR-dependent phosphorylation of SAC proteins that helps to enforce a mitotic arrest in response to DNA damage.
Our data extends the previous observation that the SAC mediates a mitotic delay in response to multiple lesions affecting DNA replication [22]–[25],[48]. Two previous studies have shown that the SAC contributes to survival of cells lacking the DNA damage checkpoint when cells are treated with MMS or when compromised for DNA replication [24],[25]. Our data extend these previous studies in two important ways. We have shown that the SAC inhibits APCCdc20 when cells are grown in the presence of MMS and SAC-dependent inhibition does not require a functional kinetochore. In addition, we have shown that the SAC depends on the PIKKs Mec1 and Tel1. Our data are summarized in a model in Figure 4. Tel1 and Mec1, in response to MMS (and other mutations and treatments), activate both the DNA damage checkpoint and the SAC. The DNA damage checkpoint and the SAC converge on Pds1, by independent mechanisms, to restrain anaphase. One possible reason is that the DNA damage checkpoint recruits the SAC as a backup to assure that cells do not enter anaphase. MMS treatment causes stalled replication forks [49]. Cells will activate the DNA damage checkpoint only after they surpass a threshold of stalled replication forks, presumably because stalled and active forks are similar in structure [50],[51]. This threshold assures that the DNA damage checkpoint does not interfere with normal replication. A cell that enters into mitosis with stalled replication forks, below the threshold, could initiate a catastrophic mitosis. If cells arrest because of some threshold of stalled replication forks, then this would constitute a new checkpoint for the completion of DNA replication. Such a checkpoint is controversial [52] but the exciting possibility that Mec1 and Tel1 activates the SAC to achieve a cell cycle arrest warrants further investigation.
All strains were derivatives of W303 (MATa or MATα ade2-1 trp1-1 can1-100 leu2-3,112 his3-11,15 ura3-1) and are listed in Table S1. Cells were arrested using the mating pheromone, α-factor at 5 µM for BAR1 strains and 0.1 µM for bar1 strains. Cells were released from α-factor by washing in water for three times and released into fresh pre-warmed medium. The temperature sensitive strain, rad9 rad24 ndc10-1, was grown at 23°C (permissive) and tested at 35°C (restrictive).
Standard yeast genetics techniques and media were used [53]. Cells were grown in YPD medium (1% yeast extract, 2% Bacto Peptone, 2% Glucose, 40 mg of adenine per liter). Yeast transformations were by the lithium acetate method [54]. Epitope-tagged alleles PDS1-13MYC-HIS were constructed by PCR-mediated one-step gene replacements [55].
The ndc10-1 mutant was obtained as by double fusion PCR [53]. Deletion of MAD1, MAD3, BUB1, and BUB3 genes were generated by PCR and transformation for each coding region was replaced by the KanMx4 or ClonNAT (NAT) genes by one-step gene replacement. The CDC20-127 dominant allele was made from PCR and transformed to wild type and rad9 rad24 strains [40]. Other mutants were made by standard tetrad genetics.
Cells were grown to O.D. of 2.0 overnight in YPD medium. For synchrony, cells were diluted to O.D. of 0.2 in YPD medium for bar1 deletion strains or acidic YPD (pH 3.5) medium (BAR1 strains) with α-factor. Cells were monitored under microscope to arrest 85–100% as unbudded cells typically after 2.5–3 hours. Cells were washed with water and resuspended in fresh medium under experimental conditions. Methylmethane sulfonate (MMS, Sigma M-4016) concentration was 0.01% V/V. For experiments with the temperature sensitive strain rad9 rad24 ndc10-1, wild type and mutant cells were grown and arrested with α-factor at 23°C. They were shifted to 35°C for 1 hour to inactivate Ndc10 and then released in fresh medium at 35°C with or without MMS. At each time point and for each strain, cells were taken for DAPI staining or FACScan (flow cytometry) using Sytox Green (Molecular Probes, Inc.) and western blot analysis. Western blots were with mouse monoclonal anti-Myc antibody 9E10, or rabbit anti-Tub2 antibody FY124, a generous gift from Frank Solomon (MIT), for tubulin loading controls. Flow cytometry was as previously described [56] and performed at the University of Virginia core fluorescence-activated cell sorting facility. Every strain was tested independently at least twice and up to six times by flow cytometry. Nuclear division for cells stained with Sytox green or DAPI was determined using a Nikon E600 microscope equipped with epifluorescence. At least 100 cells were counted for each time point.
Cells were arrested with α-factor and after 3 hrs at 23°C they were washed and released in fresh media with or without 0.01% MMS. The cells arrested in S phase were treated with 0.1 M Hydroxyurea (HU, Sigma H-862). Samples were taken in every hour. Plugs for CHEF gels were prepared as soon as the cells were sampled according to manufacturer’s instructions (BioRad). Samples were subjected to CHEF; 120° field angle, 6 V/cm, initial switch time of 60 s, final switch time of 120 s for 21 h at 11°C.
|
10.1371/journal.pcbi.1003474 | A Network Characteristic That Correlates Environmental and Genetic Robustness | As scientific advances in perturbing biological systems and technological advances in data acquisition allow the large-scale quantitative analysis of biological function, the robustness of organisms to both transient environmental stresses and inter-generational genetic changes is a fundamental impediment to the identifiability of mathematical models of these functions. An approach to overcoming this impediment is to reduce the space of possible models to take into account both types of robustness. However, the relationship between the two is still controversial. This work uncovers a network characteristic, transient responsiveness, for a specific function that correlates environmental imperturbability and genetic robustness. We test this characteristic extensively for dynamic networks of ordinary differential equations ranging up to 30 interacting nodes and find that there is a power-law relating environmental imperturbability and genetic robustness that tends to linearity as the number of nodes increases. Using our methods, we refine the classification of known 3-node motifs in terms of their environmental and genetic robustness. We demonstrate our approach by applying it to the chemotaxis signaling network. In particular, we investigate plausible models for the role of CheV protein in biochemical adaptation via a phosphorylation pathway, testing modifications that could improve the robustness of the system to environmental and/or genetic perturbation.
| Advances in the ways that living systems can be perturbed in order to study how they function and sharp reductions in the cost of computer resources have allowed the collection of large amounts of data. The aim of biological system modeling is to analyze this data in order to pin down the precise interactions of molecules that underlie the observed functions. This is made difficult due to two features of biological systems: (1) Living things do not show an appreciable loss of function across large ranges of environmental factors. (2) Their function is inherited from parent to child more or less unchanged in spite of random mutations in genetic sequences. We find that these two features are more correlated in a specific subset of networks and show how to use this observation to find networks in which these two features appear together. Working within this smaller space of networks may make it easier to find suitable underlying models from data.
| Biological systems in general show various types and degrees of robustness to environmental changes, meaning that they continue to function even when changes in the environment occur. This imperturbability is often accompanied by robustness to genetic perturbations, meaning that progeny function even though their genotype is not identical to the parent genotype [1]–[4]. Both features play an important role in evolutionary biology. While the former is a direct outcome of selection, the relationship between evolution and genetic robustness is likely to be indirect for low functional mutation rates [5]–[7] since selection acts only on the phenotype of an organism and not its genotype [8].
It has been argued that the ability of an organism to withstand genetic mutations improves its ability to evolve [8]–[11]. However, the rationale for selection for genetic robustness is still controversial [5]–[8], [12]–[14]. A correlation between the evolution of environmental and genetic robustness has been proposed [1], [8], [15], [16] based on examples observed in many biological systems such as in yeast [1], bacterial sncRNAs [2], segment polarity in the fruit-fly [3], bacterial chemotaxis [4], [17]–[24], heat-shock proteins [25], [26], and miRNA stem-loop structures in various species [27] and based on numerical models of evolution under varying fitness conditions [15], [16]. Similarly, it has been shown that metabolic networks evolving under fluctuating environments acquire robustness to the loss of certain genes as well, while those evolving under stable environments do not [28]. However, there is no general mathematical proof for this correlation [8].
In this study, we develop a computational experiment to investigate the plausibility of this hypothesis, that there is a general correlation between environmental and genetic robustness, and provide a quantitative measure of the degree of correlation, if any. In more detail, we shall show that the presence of a specific dynamic network characteristic in networks is associated with a better correlation between genetic and environmental robustness than found in networks where it is absent. Rather than focusing on a particular system in a specific organism, we choose one function of interest: The ability to attain steady state output for constant input. If a network capable of carrying out this function is robust to external environmental perturbations, what is the probability that it is also robust to internal (e.g., genetic) disruption? To be specific, we define environmental robustness of a biological network as the ability to maintain an output in the face of input perturbations. Genetic robustness is defined as the ability of a biochemical system to maintain the same output in the face of genetic mutations represented as rate constant changes in the equations representing it. This representation of a mutation as a jump from one set of parameters to another is a standard assumption [29].
For mathematical convenience, we restrict our discussion to Michaelis-Menten type networks as they are likely to reach a steady state under constant inputs relative to general networks without sigmoidal saturation. Such networks were also used in the analysis of three node biochemically adaptable networks by Ma et al. [30]. The sensitivity of biochemical kinetic models to parameter perturbations has been intensively investigated [29], [31]–[35] as a mathematical model of a biological system should be able to reproduce the function of interest or fit experimental data with a minimal need for parameter fine-tuning [35], [36]. Systems of biochemical adaptation [30], [37]–[40] have been of interest in particular.
Defining a topology to be a graph of interactions independent of parameter values, we test a large number of random N-node topologies for networks capable of reaching a steady state both under constant input concentrations and after a persistent step change in these input concentrations. We define a network as a topology with a specific set of parameters. Each network is given a numerical value for its level of robustness to input and parameter perturbations. The level of robustness of the topology is determined by averaging over this value obtained from its corresponding networks. In particular, we differentiate between networks that show a transient response to a step change in input and those that do not. We find that there is a statistically significant model II regression between the level of robustness to input of a topology and its level of robustness to parameter perturbations that has a steeper slope in networks with a transient response. Our results may be relevant to the discussion about the relationship between the need to survive in a constantly changing environment and the evolution of genetic robustness.
There is a large literature on functional motifs that are necessary for a biological system to carry out specific tasks [30], [41]–[49]. Here, we test all possible 3-node topologies to find the particular motifs that are of use in achieving both robustness to input and parameters. Having established the correlation between environmental and genetic robustness, we ask if there are topologies sharing certain sets of motifs/architectures that show stronger correlations than others. Ma et al [30] computationally explored all possible topologies of 3-node Michaelis-Menten enzymatic networks for motifs that can best accomplish biochemical adaptation. Using our results on this correlation between different sets of architectures we refine the list of motifs of biochemical adaptations previously published [30].
Our approach can be used to select/reject plausible/improbable models of a system of interest. We demonstrate this via a comparative study of bacterial chemotaxis signaling systems. Chemotaxis is a process generally used by bacteria to sense changes in their chemical environment [4], [17]–[24]. Chemotactic signaling is a well-studied system, but most of the focus has been on the chemotaxis network of the Escherichia coli (E. coli) bacterium [4], [17]–[20] despite the fact that chemotactic signaling pathways differ between species [21]–[24]. For instance, CheV is a chemotaxis protein found in many bacteria but not in E. coli. In many species, it was shown that CheV, or a variant of it, plays a role in biochemical adaptation during chemotaxis via its phosphorylatable receiver domain [24], [50], [51]. However, the exact mechanism is still not known [24]. Here, we compare the coarse-grained network of E. coli chemotaxis with several others involving CheV phosphorylation. We draw conclusions based on the resultant values of robustness to both input and parameter perturbations and the correlation between them.
In summary, we provide extensive evidence for a mathematical principle stating that, statistically speaking, dynamical systems that are biochemically adaptable are also genetically robust. We apply this knowledge to search for topological categories and subcategories within 3-node networks that show a particularly strong correlation and a linear relationship between their robustness to input and to parameter perturbations, and to shed more light on the chemotactic signaling pathways in bacteria. This method of searching for motifs can be extended to other functions and to bigger networks in order to find motifs that combine more complex functions necessitating larger numbers of nodes.
In the current work, we sample over 50,000 topologies each of 5-node, 10-node, 15-node, and 30-node networks, and over all 39 possible topologies of 3-node networks. For each topology , we average over a large number of randomly chosen parameter sets. The parameters are chosen from a uniform distribution within fixed ranges as described in the Methods section. For each network defined by topology and parameter set , we compute two values: which is a measure of the robustness of the network to a persistent step change in input, and which is a measure of the robustness of the network to perturbations in its set of parameters . We take the geometric averages of and over the whole parameter space as a quantitative evaluation of the robustness of topology to a step change in input and to parameter perturbations respectively (Fig. 1).
In previous work on biochemical adaptability, it was assumed that networks that quickly respond to input change are better adapted than those with slower response [30], [37]. In this work, we take a qualitative approach to avoid a bias towards larger or faster transients. A biochemically adaptable network is defined as one that is both robust to input perturbations and has a transient response to a persistent step change in input, independent of the magnitude of the transient. A step change in input (Fig. 2A) induces three possible responses from the output dynamics (assuming a steady state can be reached): No response (Fig. 2B), a monotonic response (Fig. 2C), or a transient response (Fig. 2D). Due to possible computational noise in the time-course of the output concentration, we need an objective way to distinguish between a network with a small transient and a non-responsive or monotonically responsive one (e.g., Fig. 2E,F). To this end, we evaluate the Pearson shape correlation between the network's time-course and two model time-courses representing the dynamics of a network characterized by perfect biochemical adaptation (the red time-courses in Fig. 2E–H) and that of a monotonically responsive one (the green time-courses in Fig. 2E–H). Here, the time-course represents the dynamics of the concentration of the output node from one steady state (before the change in input concentration) to a new steady state (after a persistent change in input concentration). In summary, we use the term “transiently-responsive” (TR) for a network that responds to a persistent step change in input and then returns to a new steady state different than the peak response regardless of whether it is also robust (Fig. 2H) or not (Fig. 2G). Any network that does not pass the Pearson test, whether it shows no response or a monotonic one, is termed Non-Pearson (NP). A perfectly biochemically adaptable network is one that is both transiently-responsive and perfectly robust to input perturbations (Fig. 2H).
We define and derive (see Methods) two quantitative measures of input and parameter robustness for each topology : , , , and . and are the values of input and parameter robustness of TR networks (networks that passed the Pearson test) while and are the values of input and parameter robustness of NP networks (networks that did not pass the Pearson test). A topology is perfectly robust to input perturbations if is very small, and similarly is perfectly robust to parameter perturbations if is very small (i.e., has a very large negative value).
For topologies with more than 3 nodes we sample over at least 50000 different ones of each size (5, 10, 15, and 30-node topologies) while 3-node topologies are exhaustively sampled. The different topologies (that have more than 3-nodes) are sampled randomly as described in Selection Criteria in the Methods section. We reject topologies with a low fraction of TR networks (, where is the ratio of the number of TR networks to the total number of networks) and exclude them from any further analysis. We chose 2.3% to be the cutoff on as it is the minimal value of that removes clusters and outliers (Fig. S1). With this, we are left with 2445, 7847, 18300, 19264, and 16589 3-node, 5-node, 10-node, 15-node, and 30-node topologies respectively. The networks sampled from each of these topologies are qualified as TR (Fig. 3) or NP (Fig. 4) and separated accordingly.
We find that over the parameter space of a topology , and can span a wide range of values. Within both TR and NP networks, we find a significant (p≅0.0) linear correlation between and . A comparison of the slope of the linear regression (using model II regression, in particular the ordinary least square bisector method described in [52]) shows a clear and systematic pattern between topologies of different sizes and TR and NP networks of the same size. We find that the slope increases as the size of the network increases: slope = 0.62±0.09 for 3-node (Fig. 3A), 0.64±0.05 for 5-node (Fig. 3B), 0.76±0.04 for 10-node (Fig. 3C) 0.80±0.05 for 15-node (Fig. 3D), and 0.96±0.13 for 30-node topologies (Fig. 3E). The marginal error is taken as the 95% confidence interval where the variance of the slope is calculated using its estimate for OLS bisector regression derived by Isobe et al [52]. Similarly, for NP networks we obtain: slope = 0.50±0.04 for 3-node (Fig. 4A), 0.44±0.02 for 5-node (Fig. 4B), 0.54±0.02 for 10-node (Fig. 4C), 0.59±0.02 for 15-node (Fig. 4D), and 0.70±0.03 for 30-node topologies (Fig. 4E). As above, the marginal error here is taken as the 95% confidence interval. The confidence intervals show that, for all sizes, the values of slopes for TR networks are consistently higher than that for NP networks of the same size and that the difference between the two slopes is significant.
The values of the Pearson correlations within TR and NP networks show no clear pattern. This is mainly due to the variability introduced by parameters whose robustness stays invariant and reducible topologies within N>3 N-node networks (see Text S1 and Discussion). Due to these caveats we are cautious about drawing conclusions based on the values of the Pearson correlation.
Sampling over all 39 possible topologies, our results show only 4153 topologies have associated TR networks. Within these topologies we find a significant linear correlation before (Fig. S2 A) and after (Fig. 3 A) introducing the cutoff, as discussed in the previous section. In what follows we show how we can extract motifs (i.e., basic topologies that may be more likely to appear in biological systems) by examining the slope of the linear regression between and . Here, we show that motifs can be extracted from topologies representing the basic backbones shared by a set of topologies showing a stronger relation between environmental and genetic robustness as follows.
We first consider two known motifs, the incoherent feedforward motif (IFF) and the negative feedback loop motif (NFL) and examine their corresponding relations. IFF (Fig. 5A) is a topology wherein the output node is affected by the input receiving node via two paths, one direct and the other indirect, such that, collectively, one path is activating and the other is deactivating. This implies four subcategories denoted IFF1–IFF4. NFL (Fig. 5A) is a topology wherein a node is activated/deactivated by another node , and node is deactivated/activated back by node either directly (NFL1, NFL2) or indirectly (NFL3–NFL10). We find that the majority of TR topologies have IFF, NFL, or both IFF and NFL motifs. Only a few have neither IFF nor NFL; these are, however, robust to neither input nor parameter perturbations (Fig. 6A) and they all have low fractions of successful trials, indicating that TR networks generated from these topologies are sparse. All 4 subcategories of IFF are fairly robust to both input and parameter perturbations (results not shown). Though NFL only topologies are generally less robust than those containing IFF, a small group of them (green cluster at the bottom left of Fig. 6A) have low numbers of successful trials but are highly robust within their small TR space. When not coexisting with other robust motifs, only 4 out of the 10 categories of NFL (NFL1, NFL2, NFL4, and NFL6) are robust to both input and parameter perturbations (Fig. 6B). Seeing how NFL1 topologies show separate groups in Fig. 6B, we examine the distribution of all topologies containing NFL1 according to its 8 types (Fig. 5B). We find that NFL1 type1 topologies are strongly correlated (r = 0.92, p≅0) while NFL1 type2 show separate clustering (Fig. 6C). Thus, we further divide NFL1 type2 into two subtypes (Fig. 5C), type2a (the output node deactivates itself) and type2b (all others). While both subtypes show strong correlation between their and values (Fig. 6D, type2a: r = 0.97 and p = 10−28, type2b: r = 0.97 and p = 10−51), type2b shows a much steeper slope (1.12 for type2b, 0.33 for type2a, ttest = 3.8 and p = 0.0002). This steeper slope may be advantageous for specific biological functions, though both types show strong correlation between the two types of robustness. In the presence of IFF, the two types show no correlation (p = 0.14 and 0.20 for type2a and type2b respectively).
In this section, we answer the following questions: (1) What is the reason for the large variation around the regression lines in Figs. 3 and 4? (2) How does the distribution of and values and their correlation relate to correlations in and values of the networks within the individual topologies?
To answer the first question, we speculated that since clearly each of the parameters in a topology will have different robustness values, we might be able to separate the parameters into different categories such that the regression along each category leads to different slope values. If we show this to be true, then as the number of possible categories increases, one expects larger variation in the value of for a given value. If, in addition, the number of categories is proportional to the number of nodes, then the observed variation would increase for a bigger network, as evident in Fig. 3. In what follows, we investigate this possibility within 3-node networks.
Consistent with the 5-, 10- 15-, 30-node analysis above, we remove topologies with a low fraction of TR networks () and are left with 2534 topologies to work with.. Next, we separate the parameters of each topology into 7 categories (Fig. 7). Parameters belonging to categories 1 or 2 are those associated with links affecting (i.e., directed towards) the input receiving node, node 1. Those belonging to categories 3 or 5 are associated with links affecting the buffer node, node 2. The rest (in categories 4, 6, 7) are associated with links affecting the output node, node 3. Then, for each category j of a network , we evaluate the value which takes into consideration only robustness to perturbations in parameters belonging to category j (Eq. 26 in Methods). The corresponding value for the topology , is (Eq. 28 in Methods). We find that indeed the regression on each of the 7 categories results in a different slope and different correlation strengths. The results of the overall linear regression are shown in Fig. 8A. For the separate categories, we find that the strongest correlation is between and robustness to perturbations in parameters belonging to category 1, (Fig. 8B, slope = 0.97, r = 0.97, p = 0), followed by category 2 (Fig. 8C, slope = 1.01, r = 0.78, p = 0). Conversely, and show no correlation (Fig. 8H, slope = 1.0, r = 0.01, p = 0.81). In fact, the strength of the correlation between and decreases in the following order: j = 1 (r = 0.97), 2 (r = 0.78), 3 (r = 0.44), 5 (r = 0.42), 4 (r = 0.32), 6 (r = 0.12), and 7 (r = 0.01).
The second question is related to whether within each topology the parameter subspace corresponding to input robustness is positively correlated with that corresponding to parameter robustness. If they are not correlated, then the two subspaces could be disjoint and the collective/coarse-grained correlation (i.e., the correlation between the and ) does not support our hypothesis.
We follow the same procedure as above and separate the parameters into the 7 categories depicted in Fig. 7. The aim is to be able to compare the results with those in Fig. 8. For each topology , we perform a linear regression on the relationship between and for each category j. The results of the correlation strength and slopes are represented by their corresponding square of the Pearson correlations and slopes , for . The relationship between and is shown in Fig. 9 while that between and is shown in Fig. 10. As above, the strongest correlations and steepest slopes are found between and of parameters belonging to category 1, . For all the topologies, remains ≥0.9 (Fig. 9A) and ≥0.8 (Fig. 10A). A weaker fine-grained correlation indicates a less collective robustness as indicated by the increase in (i.e., decrease in parameter robustness) as decreases, for (Fig. 9A,B,C,E). This pattern does not appear for (Fig. 9D,F,G), which is consistent with the results in Fig. 8 where categories 4, 6, and 7 show the weakest correlations between and compared to the other categories. In particular, most of the values are very small, less than 0.2, which is consistent with the results in Fig. 8H where no correlation is found (as indicated by the high p value). Furthermore, one can map the different clusters appearing in Fig. 8B–H into the clusters that appear in Fig. 10A–G. For example, the set of topologies showing a can be mapped to the cluster in Fig. 9G with ranging between 0 and 0.4 and that in Fig. 10G with ranging between 1.0 and 1.5. Similarly, in Fig. 8D, the separate two sets of topologies showing a low parameter robustness value ( between −0.2 and 0) can be mapped to the two clusters in Fig. 9C on the top left side with ranging between 0 and 0.3 for one, and between 0.2 and 0.4 for the other, and the two clusters in Fig. 10C with negative values of . Further investigation of the set of topologies corresponding to the different clusters goes beyond the scope of the work presented here.
The main proteins/receptors involved in E. coli chemotaxis are CheA, CheW, CheB, CheR, CheZ, and CheY. E. coli uses an anticlockwise rotation of its flagella to move forward. A decrease or increase in the concentration of nutrients (chemo-attractants) or harmful chemicals (chemo-repellents), respectively, provokes a change to a clockwise rotation which causes the E. coli to tumble and thus change direction. This signal to the flagella is controlled by the chemotaxis protein CheY. A stimulus (i.e., a change in the chemical concentration in the environment) is sensed by periplasmic binding proteins which couple to CheA in the inner membrane with the help of CheW. An increase in chemo-attractant concentrations inhibits the phosphorylation of the receptor complex CheA-CheW (RC-P) (Fig. 11A) while a chemo-repellent enhances it (Fig. 11B). RC-P gives its phosphate group to both CheY and CheB (CheY-P, CheB-P). CheB-P demethylates glutamate residues while CheR enhances methylation. In turn, methylated glutamate (M) enhances the phosphorylation of the receptor complex. The chemotaxis protein CheZ helps speeding the autodephosphorylation of CheY-P [19]–[21] (Fig. 11A–B). For simplicity, we further coarse-grain this network such that M and RC-P interact via a negative feedback loop (Fig. 11C–D). In the supplementary material (Fig. S3), we demonstrate that there is no significant difference in the results between the topologies shown in Fig. 11A and its coarse-grained equivalent shown in Fig. 11C (slope = 0.79 and 0.77, respectively), though coarse-graining improves the Pearson correlation as it removes the redundant link leading to additional variability. The topology under the influence of a chemo-repellent has a much lower fraction of TR networks and shows no correlation (r = −0.01, p = 0.85) in its un-coarse-grained form (Fig. 11B). It was important to remove the redundancy to obtain a significant correlation (Fig. 11D, slope = 0.37, r = 0.45, p = 10−14).
Chemotaxis in many other bacteria is more complex and involves more proteins. One such protein is CheV which generally contains a phosphorylatable domain [24]. Here we consider all possible coarse-grained interactions between phosphorylated CheV (CheV-P), RC-P, and M. The only assumption we make is that RC-P gives its phosphate group to CheV in addition to CheB and CheY (Fig. 11E–F). With this, we obtain 33 possible sets of signed directed edges as listed in table 1, where we are considering all 3 possibilities (i.e., activation, deactivation, or no link) for the 3 suggested links.
For each of the 27 topologies, we compute the and values (Figs. 12, 13) and the slopes of the regression between and values of their corresponding TR networks (Figs. 12B, 13B). We compare the results with that of the E. coli topology both under positive (Fig. 12) and negative (Fig. 13) stimuli. Topologies 1–3, 5–7, 10–12, 16, and 19 are highly improbable as they have no significant number of TR networks within the sampled parameter space when chemo-repellents are the stimulus (Fig. 13). Topologies 4, 8, 18, 25–27 are also eliminated as they either show either a negative or no correlation (Fig. 12C–D, 13C–D) under either a chemo-attractant or a chemo-repellent. Topologies 9, 14, 18, 21, 23 are less likely than the rest (13, 15, 20, 22, 24) as they have a weaker correlation between and than E. coli as deduced from the lower p values (Fig. 12C, 13C). Topologies 20 and 22 are less robust to input perturbation than Ecoli when chemo-repellents are the stimulus (Fig. 13), and 24 has a significantly smaller slope. Finally 13 is more robust to parameter perturbations than 15. The distributions of , for each topology are shown in Figs. S4, S5, S6.
In this work, we demonstrated that there is a general positive power-law correlation between environmental and genetic robustness in TR networks, and a statistically significant trend to a directly proportional linear relationship between the two in the limit of large networks. Conversely, monotonically responsive and non-responsive (NP) networks show a weaker relationship than TR ones. Furthermore, this distinction between the two classes becomes more prominent as the size of the networks increases. Therefore, this relationship associated with TR may be relevant to the evolution of biochemical networks. While other factors have played a role in the evolution of genetic robustness, our results show that, for TR networks, as the system evolves to withstand external environmental perturbations, it will, with high probability, concomitantly become robust to certain genetic perturbations.
We speculated that the inverse of the slope is proportional to where is the number of nodes. We performed the corresponding regression and obtained for and (r = 0.9439, p = 0.008 (1-tailed), p = 0.016 (2-tailed)). To confirm our results, we performed a Bayesian analysis for the model with a uninformative flat prior on the parameters and obtained and from the second moments of the posterior. Thus the Bayesian analysis confirms the linear regression. For NP networks, the same regression gives for and (r = 0.7608, p = 0.07 (1-tailed), p = 0.13 (2-tailed)). The value of for TR networks in the limit of N large is thus 0.99±0.08 while that of NP networks is 0.68±0.21. While the latter's regression is not significant at the p = 0.05 level, the value of the intercept did not significantly change for different power values (we tried and ). The statistically significant regression for TR networks implies that as a network evolves to be more robust to input perturbations it will also evolve to be robust to parameter perturbation (and vice versa) at a faster rate. Most importantly, as the size of TR networks becomes larger, the linear relationship between the logarithms quantifying robustness to input and that to parameter perturbations implies that for larger TR networks, is, with statistical significance, and within the computed uncertainty, proportional to while for larger NP networks, tends to be proportional to . As standard in the analysis of power-law relationships, we computed the regression using logarithms. For specific biological situations, it may be conceptually more appropriate to compute a direct fit, but for general random networks, we know of no such principle. An exponential fit between and for different numbers of nodes would be difficult to interpret as the power-law is changing with the number of nodes, tending to a constant only as the number of nodes becomes large.
A drawback of our method is that the random generation of large networks does not account for reducible topologies which can introduce more variability and thus more error and a lower correlation between the two robustness measures. This makes a comparison between the correlation coefficients of topologies of different sizes a trifle problematic. However, the space of topologies grows so rapidly with the number of nodes that the likelihood of randomly selecting a reducible network decreases precipitously. Similarly, the averaging method does not distinguish between links contributing to the robustness of either input or parameters and those that do not. A method that could pinpoint such links would be useful in this context.
Our results on the adaptability of 3-node motifs differ somewhat from Ref [30] due to our use of a qualitative test, the Pearson shape correlation, for assessing the transient response property of a network. We are not aware of a biologically plausible rationale for an explicit cutoff on the size or speed of a response as biological examples can exhibit both extremes of size or duration of transients. The general motifs shown in the literature [30] need further qualification to be deemed biochemically adaptable. For example, many topologies containing NFL are nonresponsive. Conversely, we show that a subcategory of NFL, NFL1 type2b is particularly robust and exhibits a strong correlation between robustness to input and parameter perturbations (Fig. 6D).
Our results are consistent with biological networks described in the literature. For example, we show that the coarse-grained network topology of E. coli chemotaxis, as described in the literature [17]–[21], is NFL1 type2b (Fig. S7C), as follows. When the receptor complex is activated, it causes the phosphorylation of the response regulator CheY leading to increased probability of tumbling. An increase in the chemo-attractant level (I) suppresses the activity of the complex and, in turn, the phosphorylation of CheY (Fig. S7A). If I is the input (which we set to always activate the input-receiving node in our computations, for consistency), then we can define the concentration of the input-receiving node as that of the deactivated complex, X1 (i.e., the activated complex represent X1 in its deactivated form). In this case, X1 deactivates CheB which inhibits methylation (M). M activates the complex which is equivalent to deactivating X1. The latter inhibits the phosphorylation of CheY (the output) and thus decreases the probability of tumbling (Fig. S7B).
An example of IFF is the Ras model of MAPK cascades discussed in Ref [53]. The input simultaneously activates two factors, SOS and RasGAP which activate and deactivate Ras, respectively and simultaneously. The model is shown [53] to be responsive only when the activation of SOS is faster than that of RasGAP. Thus, one can further coarse-grain it by removing the intermediate node between Ras and the input node (Fig. S8). This reduces to an IFF1 topology.
In Ref [30], all NFL topologies wherein the output node directly affects the input receiving node were found to be not robust or transiently responsive. While consistent with our results showing that NFL7–NFL10 are not robust, note that when the negative feedback loop has a direct and an indirect path, the outgoing and incoming links of the input receiving node must have the same sign for adaptability and parameter robustness to be achieved (see NFL4 and NFL6 as opposed to NFL3 and NFL5 in Fig. 6B). Our work goes beyond pointing out general motifs. We refine subcategories within these motifs and show that, in fact, they do vary in their biochemical adaptation properties.
Traditionally, network motifs represent subgraph topologies that appear in biological networks much more often than one would expect in a randomly constructed network [49], and specific functions were assigned to different types of motifs [41], [46]–[48]. The validity of this approach has been questioned as the frequency of occurrence of these motifs was not statistically significant when compared with corresponding (i.e. same degree) randomly constructed networks [54]. It was argued that one cannot analyze subgraphs independently of the rest of the network as interactions will drastically change the functions assigned to the particular topology [55]. In our work, a motif does not represent a subgraph, rather the topology of the backbone of (possibly much) bigger networks.
We use our approach to differentiate between plausible models of the role of the CheV-P protein in bacterial chemotaxis. We find that there are only a few possible ways that CheV-P can be linked to RC-P and M. We suggest that while there are at most 9 possible topologies, the most plausible one has M enhancing the phosphorylation of both CheV and the receptor complex.
Some specific network features have been associated with robustness to environmental variation in bacterial gene expression. Insulating gene expression by different modes of control, from activation to repression depending on the required high or low activity, has been suggested as a general control feature [56].
Our approach to motif discovery can be extended to networks with backbones with more than 3 nodes. While exhaustive enumeration of small motifs with desired functions is fascinating [30], [41]–[43], it is neither immediately evident nor has it been demonstrated in any context that such motifs could be put together to make systems with multiple functions while preserving the robustness or responsiveness properties of the separate motifs. To get to the point where we can plausibly discuss architectural principles in biology, it seems necessary to find general characteristics of classes of networks of all sizes that could perform functions of biological interest. Our work is a step towards this goal.
Following the same initial setup as in Ref [30], a biochemical network is represented with a directed signed graph wherein the nodes of the network represent the enzymes. The latter can either be active or inactive and are able to interconvert between the two states. Thus, the elements of the corresponding adjacency matrix can take the values , implying that node deactivates node , has no effect on , or activates , respectively. No parallel links going in the same direction are allowed, i.e., cannot be >1. We divide the nodes into two types, varying nodes and fixed nodes. The latter correspond to inputs and basal enzymes which are added to each network to ensure that each node has at least one activating and one deactivating link. Thus, for an N-node network with inputs and basal enzymes, is an matrix where . These concentration values are represented by an vector(1)where is the concentration of the active form of the enzymes at time ,(2) and are the time-independent concentrations of the inputs and basal enzymes, respectively.
Assuming that the enzymes are non-cooperative and hence that they obey the Michaelis-Menten kinetics, the rate equations governing the dynamics of the network take the following compact form(3)where is a unit step function defined as(4) and are the catalytic and Michaelis-Menten rate constants for the regulation of enzyme by enzyme , for and .
In equation (3), the total concentration of each enzyme is kept constant and normalized (i.e., the concentration of the active form of an enzyme plus that of its inactive form is always equal to one). Thus, for . For all simulations presented here we use only one input, . This particular choice of input concentration should not have a significant effect on our qualitative results, as we have checked explicitly while formulating our hypothesis. Networks are allowed to reach steady state before the concentration of the input is perturbed. We are only concerned with the relative change in steady state concentrations.
N-node networks are identified with directed signed graphs representing their topology and a set of parameters , for , where is the total number of sampled topologies, excluding those wherein one or more nodes have a total degree of zero or the output node cannot be reached from the input receiving node (Fig. 1). Each topology, in turn, is sampled over a large number of random networks, i.e., a large number of randomly chosen parameter sets , for , where is the total number of sampled networks (sets of parameters) for topology . The total number of parameters in each set (i.e., length of the vector ) varies depending on the topology. The order of magnitude of increases exponentially with the size of the networks. For example, values for are around 20, around 40 for , and 500 for . Similarly, the number of iterations (i.e., number of sampled networks ) required (see Selection Criterion below) also increases with . For example, for values range between 104 and 105, while for , values range between 106 and 107. Typically, an iteration takes less than 10−3 seconds of CPU time for small networks (), thus 1 to 2 minutes to test each topology. On the other hand, for large networks (), an iteration typically takes around 0.04 seconds of CPU time, 3 to 5 days for testing each topology.
We define a TR network as one whose output dynamics has a non-monotonic transient between two steady states as a response to input change (i.e., the steady state values before input perturbation and that after input perturbation). We find the transition time (i.e., the time at which the concentration is maximal/minimal before it starts decreasing/increasing again) and enzyme concentrations, and (i.e., concentrations at ), by solving for the turning point(5)where is the concentration of node k at steady state.
We use a Pearson test to determine if a given network is TR.
First, we define two functions and as model functions of perfect adaptability and non-adaptability (a monotonically changing network), respectively (Fig. 2):(6)(7)Define corresponding Pearson shape correlations and as(8)(9)where , and are the mean values of , and . With this, a network is deemed TR if . Comparing the absolute values or and instead of the actual values is necessary. Even though our definitions of and will most likely lead to positive values, this is not always the case. The reason is that eq. (6) and (7) assume the perfect case where the differences in the concentrations from the initial steady state have always the same sign (as in Fig. 2 E and F). If instead the difference in concentrations at the transition point is smaller than that at the final steady state (i.e., post-perturbation steady state), then and/or will have negative values. However, that does not matter since we are only interested in the shape of the time-course (see Fig. 2G, for example).
Note that the mean values are taken as the average over all the discretized time-steps; for example, for time steps. The size of the time-step, , is the same for all networks (), but this is not the case for the number of time steps, , as the length of the time-course of each network varies depending on how long the network needs to reach a new steady-state (i.e., the rate equations in eq. (3) for all nodes reach zero again after input perturbation. Of course, computationally, the run will stop when the rate equation for all nodes is less than 10−10). For example, in Fig. 14A and 14B we show the time-courses (in blue) of two different networks. The network in Fig. 14A needed around 130 seconds () to reach a steady state, while that shown in Fig. 14B needed around 150 seconds (). Simulations that take too long to reach a steady state () are thrown away and not considered in the analysis (i.e., are thrown away without performing the Pearson test). This cutoff on the maximal number of time-steps allowed is chosen for computational efficiency. Preliminary results showed that for most networks, if a steady state was not reached within 2000 time-steps, it is unlikely it will be reached for a long time. Since we are only interested in the statistical results and since networks are chosen randomly, there is no reason to insist on including a network that takes a lot computational time to reach a steady state. We chose because we were looking for the largest time-step (to improve computational time) that does not change the statistical results. In preliminary runs, we compared the results of 3-node networks when using and . The finer time-step allowed more topologies to pass as TR. However, these topologies had very low fraction of TR networks and were removed after the cutoff. Moreover, the statistical results were the same both before and after the cutoff. As mentioned in Experimental Setup above, large networks (30-node) take 3 to 5 days of CPU time for each topology. Using would increase this simulation time to over a month for each topology which is impractical.
We test the robustness of the Pearson test described above by comparing the results from the 3-node simulations to those employing instead the Spearman correlation using the same definition of and (Fig. S9A). Both are also compared to simulations using a different definition, and as follows:(10)(11)where is chosen here to be . This new definition allows and to get to the transition concentration, , at a slower rate, then after the transition point, , coincides with while relaxes back to the pre-perturbation steady state, at a much slower rate (Fig. 14A,B). In all cases we find no significant difference between the results for 3-node simulations (Fig. S9). This does not mean that there are no variations within individual networks. For example, in Fig. 14, we show the and values of the Pearson test for all the networks corresponding to a typical 3-node topology using both definitions (Fig. 14C). We find that there are 90 out of 21579 networks that were deemed TR in one but NP in the other. A typical time-course where the outcome of the Pearson tests differ or agree are shown in Fig. 14A and Fig. 14B, respectively. In general, most networks do not fall into this category where the values of and are very close such that different definitions of and lead to different outcomes. In Fig. S9, we verify that this change does not affect any statistical observations.
To quantify the degree of robustness to input and parameter perturbations of a particular network, we calculate the relative change in the steady state concentrations of the output node due to perturbing the input and parameter values, respectively. Let and be the average of the sensitivity of the steady state concentration of the output to each input and each parameter, respectively. Then,(12)(13)where the node is the output node, is the set of and corresponding to each reaction/link (i.e, the non-zero values), , is the total number of links, and is the set of steady state concentrations of the output node for input and parameter set .
Defining the degree of input and parameter robustness of a network as inversely proportional to the values of and ensures all the inputs and parameters of the network are taken into consideration. Analyzing the rate functions of equation (3) (see Steady State Analysis in the Text S1 for the detailed derivation), we obtain(14)(15)where , , and are the Jacobians with respect to the node concentrations, input, and parameters, respectively.
A robust topology is one that gives rise to robust networks with a higher probability when tested with a large number of parameter sets. Quantitatively, the degree of robustness to input perturbations of a given topology is taken to be the geometric average of over all . Similarly, the degree of parameter robustness is the geometric average of . A TR topology is one that has a statistically significant number of TR networks. Topologies that do not have enough TR networks are rejected and excluded from any further analysis. With this, we are left with topologies out of the sampled ones. For each topology , we define two quantitative measures each for input ( and ) and parameter ( and ) robustness. and are the values of input and parameter robustness of TR networks(16)(17)(18)where if network passes the Pearson test and zero otherwise.
and are the values of input and parameter robustness of NP networks (networks that did not pass the Pearson test)(19)(20)(21)where if network reaches a steady state (see Selection Criterion below) but it does not pass the Pearson test and zero otherwise.
We choose the geometric average as more suitable than the arithmetic average as a conservative approach to detecting a possible correlation, as the latter gives too much weight to much larger outliers.
A trial is rejected if it takes too long to reach equilibrium, or its corresponding Jacobian with respect to the node concentrations is singular (i.e., is noninvertible). With this, we obtain matrices for the relative errors and for and . and thus are determined when , , and the fraction of successful trials, , reach equilibrium values. We reject a topology if is obviously too small to be statistically significant or takes too long to reach equilibrium (see below), indicating that the parameter space leading to TR networks for that topology is too small.
We sample over 50,000 different topologies for each and all possible 3-node topologies (19683), and for each we randomly sample over a large number of parameter sets from a uniform distribution within the ranges and (whenever a link exists between vertices and ). For the topologies were randomly generated such that the value of each in the corresponding adjacency matrix can take the values −1 or 1 with probability each, and a value 0 with probability . We generated different set of topologies with . We found no significant difference in the distributions of and values depending on . The results shown here represent the collection of all the sets.
We automatically reject trials whereinWe investigate the effect of the choice of the ranges above by running two separate 3-node simulation. In the first, the parameters are chosen from a uniform distribution in the ranges and , while in the second, the parameters are chosen from a uniform distribution in the ranges and . For both ranges we find a significant correlation between robustness to input and parameter perturbations (Fig. S10). For range1 and range2 we obtain the respective values 0.38 and 0.54 for the slopes and 0.73 and 0.68 for the Pearson correlation (Fig. S10A). Furthermore the difference in the slopes becomes insignificant when only networks appearing in both ranges are taken into consideration (Fig. S10B).
As discussed above, the degree of input and parameter robustness is seen as inversely proportional to the average of the sensitivity of the steady state concentration of the output to each input and each parameter, respectively. Then,(22)(23)Inserting (22) in (23), we obtain(24)Similarly, for parameter perturbations, and(25)
In this section we analyze the parameter robustness of different types of parameters. Thus, the parameters of a topology are now divided into categories. Their corresponding measures of robustness are now defined as(26)(27)for .
Thus, the measure of robustness of a topology to perturbations in its parameters of category j takes the form(28)Note that a topology does not have to have parameters belonging to all the defined categories.
Next, to obtain an idea about how robustness to input and parameter perturbations correlate within the networks of each individual topology, we calculate the value which is the value of the slope obtained from the linear regression on vs for topology and category .
|
10.1371/journal.pgen.1005056 | A Systems-Level Interrogation Identifies Regulators of Drosophila Blood Cell Number and Survival | In multicellular organisms, cell number is typically determined by a balance of intracellular signals that positively and negatively regulate cell survival and proliferation. Dissecting these signaling networks facilitates the understanding of normal development and tumorigenesis. Here, we study signaling by the Drosophila PDGF/VEGF Receptor (Pvr) in embryonic blood cells (hemocytes) and in the related cell line Kc as a model for the requirement of PDGF/VEGF receptors in vertebrate cell survival and proliferation. The system allows the investigation of downstream and parallel signaling networks, based on the ability of Pvr to activate Ras/Erk, Akt/TOR, and yet-uncharacterized signaling pathway/s, which redundantly mediate cell survival and contribute to proliferation. Using Kc cells, we performed a genome wide RNAi screen for regulators of cell number in a sensitized, Pvr deficient background. We identified the receptor tyrosine kinase (RTK) Insulin-like receptor (InR) as a major Pvr Enhancer, and the nuclear hormone receptors Ecdysone receptor (EcR) and ultraspiracle (usp), corresponding to mammalian Retinoid X Receptor (RXR), as Pvr Suppressors. In vivo analysis in the Drosophila embryo revealed a previously unrecognized role for EcR to promote apoptotic death of embryonic blood cells, which is balanced with pro-survival signaling by Pvr and InR. Phosphoproteomic analysis demonstrates distinct modes of cell number regulation by EcR and RTK signaling. We define common phosphorylation targets of Pvr and InR that include regulators of cell survival, and unique targets responsible for specialized receptor functions. Interestingly, our analysis reveals that the selection of phosphorylation targets by signaling receptors shows qualitative changes depending on the signaling status of the cell, which may have wide-reaching implications for other cell regulatory systems.
| Signaling networks that drive cell survival and proliferation regulate cell number in development and disease. We use a simple Drosophila model of cell number control, which centers on PDGF/VEGF receptor signaling. Performing a genome-wide RNAi screen under Pvr-sensitized conditions, we identify regulators of cell number that have not been found in conventional screens. Validation by in vivo genetics reveals previously unrecognized roles for EcR and InR in the balance of cell survival in the Drosophila embryo. Phosphoproteomic analysis demonstrates distinct mechanisms of cell survival regulation by EcR and receptor tyrosine kinase signaling. It further identifies common phosphorylation targets of Pvr and InR including regulators of cell survival, and receptor-specific phosphorylation targets mediating unique functions of Pvr and InR. Importantly, the study provides precedence that the selection of phosphorylation targets by signaling receptors can change with the signaling status of the cell, which may have wide-reaching implications for other cell regulatory systems.
| The regulation of cell number varies greatly and typically depends on developmental and environmental stimuli that determine the intracellular balance of pro- and anti-death, and proliferative signals [1–3]. Proto-oncogenes and tumor suppressors play roles as regulators of cell number and the pathological extension of cell survival is a major hallmark of tumorigenesis [4]. Accordingly, understanding the complex signaling networks that regulate cell survival is an important yet incompletely accomplished goal [4,5], which can be facilitated by studying a simple model organism.
Blood cells in the fruitfly Drosophila melanogaster have been instrumental in the discovery of fundamental concepts in immunity, hematopoiesis and wound healing [6–11], but they are also a convenient model to study mechanisms that regulate cell number. In particular, the Drosophila PDGF/VEGF Receptor (Pvr), a member of the Receptor Tyrosine Kinase (RTK) family, controls anti-apoptotic survival signaling in Drosophila blood cells (hemocytes) in vivo and in the embryonic cell line Kc in culture [12]. In other instances, Pvr has been reported to regulate cell proliferation [13,14], differentiation [15,16], cell size [17,18], cytoskeletal architecture [19] and cell migration [20–22]. Drosophila Pvr therefore parallels roles of the vertebrate family of PDGF/VEGF Receptors in development and disease [12,21,23–26].
Here, we took advantage of the role of Pvr in embryonic blood cell survival and performed a systematic RNAi screen to identify regulators of cell number, using the Drosophila cell line Kc under sensitized conditions of Pvr knockdown. The screen identified enhancers and suppressors of the Pvr RNAi phenotype, many of which were not found in conventional RNAi screens examining cell growth and viability. In particular, we found that knockdown of InR enhanced the Pvr RNAi phenotype while knockdown of the Ecdysone receptor (EcR) [27] and its co-receptor ultraspiracle (usp) [28] suppressed the Pvr RNAi phenotype. We confirmed functional roles for these genes related to Pvr both in cell culture and in vivo. Phosphoproteomic analyses revealed major differences in the signaling signature of Pvr deficient cells rescued by activation of InR as compared to inactivation of EcR. Further, our analysis identified distinct sets of phosphorylation targets, common to both Pvr and InR, and unique to each receptor. Most importantly, we provide precedence that the selection of phosphorylation targets by signaling receptors can depend on the signaling status of the cell, which may have wide-reaching implications for cell regulatory systems in animal development, disease, and the experimental and therapeutic manipulation of signaling pathways.
Previously, we demonstrated that the Drosophila PDGF/VEGF Receptor, Pvr, is essential for anti-apoptotic survival in embryonic hemocytes and in the related cell line Kc, which maintains autocrine Pvr signaling [12,29]. Taking advantage of these systems, we sought to examine the signaling networks that mediate anti-apoptotic survival and regulate cell number.
First, we confirmed that RNAi-mediated knockdown of Pvr induces apoptotic cell death in Kc cells. RNAi silencing of the Drosophila inhibitor of apoptosis DIAP1, or thread (th), served as positive control (Fig. 1A). Expression of the baculovirus inhibitor of apoptosis p35 [30] rescued hemocyte survival, leading us to establish a selected pool of Kcp35 cells (Kcp35 cells, Fig. 1A). Immunoblotting confirmed that Pvr knockdown was equally efficient in Kc and Kcp35 cells (S1 Fig). Closer examination by incorporation of the thymidine nucleoside analog EdU (5-ethynyl-2’deoxyuridine) in Kc versus Kcp35 cells revealed that Pvr also moderately contributes toward cell proliferation in this system (Fig. 1B), an effect that could not be distinguished in a previous study employing cell cycle profiling [12]. Reduction in proliferation was also suggested by immunoblotting, where lysates of equal numbers of cells showed a decrease in the proliferation marker phospho-histone H3 (pHH3) in Pvr knockdown samples (S1 Fig).
Using Kc cells, we queried signaling pathways that might be involved in Pvr-dependent cell survival and proliferation. Examining activity of the Akt/TOR and Mek/Erk pathways by using antibodies to phosphorylated forms of S6Kinase (S6K, an Akt pathway target), Mek and Erk, we found that both pathways are active in Kc cells. Pvr RNAi led to a significant reduction in the phosphorylation levels of these proteins, indicating that Pvr is a major activator of these pathways in Kc cells. Single signaling mediator knockdowns of Akt, the TOR-associated Raptor, S6K, Mek and Erk served as controls (Fig. 1C). Phosphorylation signals were also quantified and displayed as a ratio with the amount of unphosphorylated signaling mediator (Fig. 1C). These findings suggest that Pvr triggers activation of the Akt/TOR and Mek/Erk and pathways, thereby supporting anti-apoptotic cell survival and proliferation.
Next, we asked whether silencing of either or both of these pathways is sufficient to affect cell viability and mimic loss of Pvr function. Combining dsRNAs targeting various mediators of the Akt/Tor and Mek/Erk pathways, we found that, despite efficient knockdown of the genes (S2 Fig), neither single nor simultaneous inhibition of both pathways caused a significant reduction of cell numbers, as quantified by CellTiterGlo assay based on ATP content (Fig. 1D), and cell counting (Fig. 1E). In contrast, Pvr RNAi, showed significant decreases in cell number (Fig. 1D, E). This predicted the presence of one or more additional, redundant cell survival/proliferation pathway(s) downstream of Pvr (‘X’, Fig. 1F), and/or parallel signaling pathways that contribute to the overall survival and proliferation of the cell (‘Y’, Fig. 1F).
Based on our prediction, we sought to identify other signaling pathways that contribute to the anti-apoptotic survival of Kc cells. We hypothesized that re-activation of just one survival or proliferation pathway would be sufficient to rescue cell numbers in Pvr deficient cells (Fig. 1F). Indeed, silencing of negative regulators of the Akt/Tor and Erk pathways rescued the Pvr RNAi phenotype, validating our screening approach. For these experiments, we ruled out that silencing of downstream signaling mediators such as Akt would result in upregulation of Pvf2 expression, the major Pvr ligand in Kc cells that mediates autocrine signaling (S3 Fig). Expanding our approach, we screened the DRSC Genome-Wide RNAi library 1.0 (Drosophila RNAi Screening Center, Harvard Medical School) for modifiers of cell number, specifically under conditions of Pvr RNAi-mediated silencing compared to a control background (Fig. 2A). The DRSC 1.0 set targets 22,914 distinct amplicons based on Flybase release 5.51 of the Drosophila genome, corresponding to 13,777 unique genes [31], 6944 of which are expressed in Kc cells [29]. Screening was performed in 384-well format, quantifying ATP content as a readout of cell number (CellTiterGlo). To determine an increase or decrease over the average value of ATP content, Z scores were calculated for each well. Focusing on those dsRNAs that show differential effects in Pvr knockdown (Pvr RNAi) versus control cells (GFP RNAi), we calculated the difference of each of the Z scores (ZDiff = Z[Pvr]-Z[GFP]), and selected amplicons with ZDiff> = 2 and ZDiff< = -2 as primary screen hits (S1 Table). Cluster analysis of the values Z[Pvr], Z[GFP], and ZDiff for each amplicon revealed three distinct classes of signatures, i.e. Pvr Suppressors, Pvr Enhancers, and Pvr ‘Upstream Genes’ (Fig. 2B). By our cutoff criteria, 64 amplicons scored as suppressors of the Pvr knockdown phenotype, rescuing cell numbers more effectively in the Pvr RNAi background compared to control cells. 65 amplicons scored as Pvr Enhancers, exacerbating the Pvr knockdown phenotype. We classified 290 amplicons as Pvr ‘Upstream Genes’, reducing cell numbers in control cells, but having rather minor effects in the Pvr silenced background. Among this group we found amplicons targeting Pvr itself and many ribosomal proteins, suggesting that many of the targeted genes play a role in the production or activation of Pvr (S1 Table).
Subsequent secondary testing of screen hits was carried out for Pvr Suppressors and Pvr Enhancers. We selected 47 suppressor genes and 47 enhancer genes based on a cutoff of ZDiff> = 2.2 and ZDiff< = -2.2 (S2 Table) and synthesized non-overlapping alternative amplicons that were free of 19bp or larger overlaps with other genes, in order to avoid off-target effects [32,33]. As in the primary screen, amplicons were tested for their ability to modify cell number, specifically comparing Pvr knockdown cells relative to control cells (S2 Table). To identify promising ‘high confidence candidates’ for further analysis, we calculated the average of the ZDiff scores among all amplicons of a gene from the primary and secondary screens (ZDiffFinal) (S3 Table). Based on ZDiffFinal values of > = 1.6 and < = -1.2, we report 30 high-confidence Pvr Suppressors and 14 high-confidence Pvr Enhancers (S3 Table). Z value cutoffs were guided by the scores of predicted genes within the set, such as members of the Akt/Tor and Mek/Erk pathways. Candidates of specific interest were confirmed by live/dead cell counting, omitting genes with obvious roles in RNA interference, such as AGO2 (S4 Fig).
Relatively few genes scored as Pvr Enhancers. Among those, we identified the RTK InR [34], and cropped (crp) encoding the helix-loop-helix transcription factor that is a homolog of the mammalian transcription factor AP-4 [35]. The screen also identified tonalli (tna), encoding a protein similar to mammalian ZMIZ1 and ZMIZ2 involved in sumoylation [36] that interacts genetically with the Brahma ATP-dependent chromatin remodeling complex in Drosophila [37].
Among the Pvr Suppressors, the screen yielded all known tumor suppressors and negative regulators of the Akt/TOR pathway, including Phosphatase and Tensin Homolog (Pten), Tuberous Sclerosis Protein 1 (Tsc1), gigas (gig)/Tuberous Sclerosis Protein 2 (Tsc2), SNF4A–a and -γ, also known as AMP-Activated Protein Kinase subunits a and γ(AMPK–α and AMPK–γ), Forkhead Box Protein (foxo), and Lobe (L), a protein with similarities to the vertebrate Proline-rich Akt substrate of 40 kDa (PRAS40) [38–41]. We further identified negative regulators of the Ras/Erk pathway Mitogen-activated protein kinase phosphatase 3 (Mkp3), and microtubule star (mts) and widerborst (wdb), which encode components of the protein phosphatase PP2A complex [42–44]. We calculated which protein complexes were over-represented with respect to the frequency of their components among the high confidence hits in the RNAi screen, and found, besides the PP2A complex, two other major protein complexes among the high confidence hits in the RNAi screen (Fig. 2C): the ecdysone receptor complex, consisting of the nuclear hormone receptors EcR and usp [27,45], and the Brahma SWI2/SNF2 family ATPase chromatin-remodeling complex, comprising osa and dalao [46,47]. Other Pvr Suppressors were CG6182, an ortholog of mammalian TBC1 domain member 7 (TBC7), and GckIII, and CG31635, an ortholog of mammalian LRRC68. Given the reported interplay between ecdysone and insulin signaling during development [48], we wanted to dissect whether common and/or distinct downstream mechanisms mediate Pvr suppression, and therefore chose InR and EcR /usp for in vivo validation.
Using Kc cells, we examined the functional roles of InR and EcR/Usp in more detail. EcR and Usp form a heterodimer and are induced by binding of the steroid hormone 20-hydroxyecdysone (20E) [45,49]. Signaling by the EcR complex plays a major role during molting and metamorphosis [50], yet a role in embryonic cell death and cell number control has not been established [51]. We confirmed the effects of silencing or stimulating InR, or silencing EcR or usp, on Pvr RNAi-induced apoptosis using TUNEL assays, and we quantified effects on proliferation using EdU incorporation in Kcp35 cells (Fig. 3A-C). Consistent with the results from the screen, we found that, in combination with Pvr knockdown, silencing of InR exacerbated apoptosis. Further, silencing of EcR or usp, or stimulation of InR with insulin rescued apoptosis (Fig. 3A). In contrast, when examining proliferation, only insulin stimulation or a Tsc2/gigas (gig) RNAi Akt pathway control significantly suppressed proliferation defects, suggesting that EcR and Usp mainly function in the regulation of cell death, rather than proliferation (Fig. 3B, C). InR knockdown seemed to enhance the reduction of EdU incorporation in Pvr knockdown cells, but differences were not statistically significant based on three independent biological replicate experiments (Fig. 3C).
Next we examined the effects of ecdysone stimulation. Anti-proliferative effects of ecdysone in Kc cells have been reported previously [52–54], but whether ecdysone also has direct pro-apoptotic effects in embryonic cells not been determined. To test this, we stimulated Kc cells with 20E at concentrations close to physiological levels (0.01ug/ml) [54,55]. Overall, 20E induced a marked reduction in cell number at stimulation times of >48h (Fig. 3D). As expected, it resulted in a reduction of cell proliferation as measured by EdU incorporation, both in Kc and in apoptosis-resistant Kcp35 cells (Fig. 3E). However, TUNEL analysis showed a substantial increase in apoptotic cells upon 20E stimulation, which was largely suppressed in Kcp35 cells (Fig. 3F). 20E did not cause a decrease of Pvr protein levels (S5 Fig), suggesting that molecular mechanisms other than Pvr downregulation account for the observed increase in apoptosis. During metamorphosis-associated programmed cell death (PCD), several genes have been described as ecdysone-induced pro-death targets, in particular Ecdysone-induced protein 93F (E93), broad (br), Ecdysone-induced protein 74EF (E74A), and reaper (rpr) [56–58]. When we examined the expression levels of these genes during ecdysone stimulation of Kc cells we found that, indeed, rpr and E93 levels increased from the first day of 20E stimulation (Fig. 3G), consistent with the induction of apoptosis. We also examined whether Pvr knockdown would have an effect on the expression of rpr and E93 but found no significant difference relative to controls (S6 Fig).
In summary, we conclude that the EcR complex has pro-apoptotic functions in the cell line Kc, which become apparent under sensitized conditions of Pvr loss of function, or experimental addition of 20E.
To test the role of InR and EcR in the suppression and enhancement of apoptosis in vivo, we examined the function of these genes in the survival of hemocytes in the Drosophila embryo. Drosophila embryos typically show a developmentally fixed number of ~600 hemocytes post stage 11 until early stage 17, and loss of Pvr signaling causes a rapid decline in hemocytes due to their apoptotic death and phagocytic clearance by the small number of remaining live hemocytes [12]. Based on our findings in Kc cells, we predicted that inhibition of EcR would rescue, and inhibition of InR would enhance, Pvr loss-of-function in embryonic blood cells [12]. Indeed, hemocyte-specific suppression of EcR signaling by expression of dominant-negative forms of EcR [59,60] partially rescued hemocyte counts in Pvr1 mutant embryos (Fig. 4), resembling rescue by the baculovirus inhibitor of apoptosis, p35 [12] (see also Fig. 4A). Conversely, expression of dominant-negative InR in hemocytes enhanced the Pvr phenotype, further reducing embryonic hemocyte numbers (Fig. 4). Consistently, we previously demonstrated that activated PI3K, a positive mediator of the Akt/TOR pathway downstream of InR, can partially rescue the Pvr mutant in vivo phenotype [12]. To confirm hemocyte autonomous effects of EcR and InR, we induced embryonic hemocyte death by hemocyte-specific expression of dominant-negative PvrΔC [12], and examined the effects of co-expressed dominant-negative versions of EcR or InR. Again, we found that dominant-negative EcR rescued apoptotic loss of hemocytes, while dominant-negative InR exacerbated the cell death phenotype (Fig. 4). Expression of the transgenes in the wild type background had no significant effects (Fig. 4A).
Intrigued by the mild increase of hemocyte numbers upon overexpression of dominant-negative EcRdn (Fig. 4A), we asked whether blocking EcR signaling alone would have a positive effect on hemocyte numbers at a later point during development, for example at the transition from the embryo to the larval stage. A time course of total hemocyte counts in live animals illustrates that, compared to stage 16 embryos, hemocyte numbers in young 1st instar larvae decline to about 60%, suggesting a putative connection with the embryonic ecdysone peak in mid-embryogenesis [61] (S7 Fig). However, comparing live hemocyte counts of controls to animals with hemocyte-specific expression of EcRdn, we did not see a significant rescue in the total number of hemocytes, despite a mild increase in EcRdn overexpressing larvae (S7 Fig).
Taken together, our findings suggest that EcR signaling accounts for a basic level of pro-death signaling in embryonic hemocytes, which however is revealed only under sensitized conditions such as Pvr loss of function. Conversely, signaling by InR contributes to the trophic survival of embryonic hemocytes, which acts redundantly with Pvr signaling, and therefore again is only evident in conjunction with loss of Pvr signaling (Fig. 4G).
Based on our findings, we sought to further dissect the relationship between Pvr, InR and EcR signaling. First, we asked whether signaling by the EcR complex acts epistatically or in parallel with RTK-triggered signaling pathways such as Akt/Tor. When comparing the effects of silencing of the EcR/Usp and Akt/Tor pathways separately and in combination, we found that simultaneous knockdown of genes from both pathways resulted in increased cell number rescue (e.g. EcR and Pten), which in many cases was significant when compared to knockdown of two genes from the same pathway (i.e. EcR and usp, or Pten and gig). This suggested a parallel, rather than epistatic relationship (Fig. 5A). Biochemically, insulin stimulation of Pvr deficient cells restored, albeit to distinctive levels, phosphorylation of downstream signaling mediators of the Akt/Tor and Mek/Erk pathways, while EcR knockdown did not show such effects (Fig. 5B). This suggested similar but not identical signaling profiles for the RTKs Pvr and InR, and distinct mechanisms for the EcR complex.
To compare the signaling profiles of Pvr, InR and EcR in a more systematic manner, we chose a phosphoproteomics approach. We utilized mass spectrometry and an isobaric labeling strategy that enables multiplexing and relative quantification between samples [62,63]. For this analysis, we surveyed the phosphoproteome by formally comparing conditions of (1) ‘high Pvr’ signaling (+ control dsRNA; taking advantage of the high endocrine Pvr activity in Kc cells); (2) ‘low Pvr’ signaling (+ Pvr dsRNA); (3) ‘high InR’ signaling (+ insulin, to stimulate endogenous InR in Kc cells); (4) ‘low InR’ signaling (+ control dsRNA; taking advantage of the low InR activity in Kc cells under standard culture conditions presumably due to low levels of dIlp expression [29], see also Fig. 5B); (5) ‘high EcR’ (endogenous EcR in Kc cells); and (6) ‘low EcR’ (+ EcR dsRNA).
First, we assessed which phosphoproteins were up- or downregulated in the rescue of Pvr silenced Kc cells. We surveyed the phosphoproteome under conditions of high and low Pvr activity (Fig. 6A and B, respectively), and analyzed separately for the two conditions the effects of EcR silencing or InR activation, assessing biological duplicates (S4 Table and S5 Table). Under ‘high Pvr’ conditions, approximately 10% of the detected phosphorylation was altered more than 1.5-fold under conditions of InR stimulation, which we refer to as the ‘InR-specific set’ (Fig. 6A and S4 Table). This percentage nearly doubled under ‘low Pvr’ conditions (Fig. 6B and S5 Table). Although some of these phosphorylations could be attributed to the fact the InR may phosphorylate Pvr targets in the absence of Pvr, this finding also suggested the emergence of new sets of up and down-regulated phosphosites that were not observed upon InR activation under ‘high Pvr’ conditions (below).
Secondly, we sought to directly measure the degree to which phosphosites were altered by InR under conditions of ‘high’ versus ‘low’ Pvr signaling, hypothesizing a ‘sensitization’ of InR signaling by the absence of Pvr. We repeated our phosphoproteomic analysis, this time directly comparing the six experimental signaling conditions among each other (Fig. 7A and S6 Table). While nearly three-quarters of the ‘InR-specific set’ of phosphopeptides remained upregulated following InR activation in the absence of Pvr, the ‘InR-specific set’ showed qualitative differences in the absence and presence of Pvr signaling. For instance, InR stimulation elevated levels of phosphorylation of fifteen phosphoproteins specifically under ‘low Pvr’ activity as compared to ‘high Pvr’ signaling. These included Chromosome-associated protein (Cap), lava lamp (lva), Enhancer of decapping 3 (Edc3), Bicaudal D (BicD), lethal(2)03709, eukaryotic translation Initiation Factor 2α (eIF-2 α) and several uncharacterized gene products. InR activation restored phosphorylation to nearly all sites downregulated in Pvr deficient cells, (Fig. 7C). These phosphorylations likely account for the ability of insulin to rescue Pvr deficiency. EcR knockdown, meanwhile, had very little effect on the phosphoproteome, both in low and high Pvr conditions (Fig. 6A, B), despite efficient knockdown (S8 Fig). Similar findings were made from the comparative analysis of all six experimental conditions (Fig. 7A, B). This is consistent with an alternative mode of action, such as the transcriptional modulation of EcR/Usp target genes (see Fig. 3G).
Lastly, we identified a pool of common phosphoproteins induced by both Pvr and InR, which comprise signaling mediators for common functions in cell survival and proliferation. At the same time, we distinguished Pvr- or InR-associated targets that may mediate receptor-specific functions. A common set of phosphorylation targets for Pvr and InR, either direct or indirect, can be inferred from the reciprocal effects of Pvr knockdown and InR stimulation, comparing ‘low Pvr’ and ‘high InR’ conditions (Fig. 7B; 153 phosphosites: S7 Table). Examples include phosphorylation of Structure specific recognition protein (Ssrp), La related protein (Larp), eukaryotic translation initiation factor 4G (eIF4G), Lamin, NAT1, Claspin, Gartenzwerg (Garz), Nedd4, Nopp140, Lk6, Yorkie (Yki), Stat92E, and Moleskin (Msk). Many of these common signaling mediators function in cell survival and cell proliferation. For example, the transcription factor Yki coordinates cell proliferation and apoptosis by directing the expression of cell cycle and cell death regulators [64]. Stat92E loss-of-function has been reported to inhibit hemocyte proliferation [65,66], while the importin Msk localizes MAP kinase to the nucleus to promote cell proliferation and survival [67]. We found enrichment for the regulation of phosphorylation of components of specific complexes by both InR and Pvr, including the Chs5p/Arf1-binding protein complex, the chromatin remodeling FACT complex, the translation initiation factor 2 complex, the cohesion-Sa complex, TRAPP complex and splicing associated factor complex (Fig. 7D, E). While we do not expect that all components of an individual complex require an alteration in phosphorylation in order for complex activity to change, more confidence for implication of that complex downstream of Pvr or InR is gleaned from multiple components exhibiting altered phosphorylation. As such, we expect that these complexes play key roles downstream of both InR and Pvr.
To distinguish Pvr- or InR-specific targets that may mediate receptor-specific functions we compared phosphoproteomes under ‘high Pvr, low InR’ and ‘low Pvr, high InR’ conditions (S8 Table). Among the Pvr-specific phosphorylations, we identified phosphoproteins involved in cell migration, cytoskeleton, and regulation of cell shape such as CIN85 and CD2AP ortholog (Cindr), Tenascin major (Ten-m), Vacuolar protein sorting 4 (Vps4), Rab7, Rho GTPase activating protein at 15B (RhoGAP15B), and Sprouty (Sty). Cindr is a recognized component of the CIN85 complex, one of three complexes for which multiple components exhibited a dependence on Pvr for specific phosphorylation (Fig. 7F). With respect to InR-specific phosphorylations, we detected phosphoproteins associated with the Gene Ontology Consortium terms growth regulation (i.e. Gp150, Foxo, L, Chico), glycogen metabolism (i.e. Glycogen Synthase), and the innate immune response (i.e. G protein-coupled receptor kinase interacting ArfGAP and Mustard). These differential phosphorylations likely provide receptor specificity and function to modulate the activity of specific complexes such as those over-represented in terms of the number of components modulated by InR activity (Fig. 7G).
Here, we present a genome- and proteome-wide survey in Drosophila to identify signaling networks and cellular regulators that control cell survival and cell number. Starting from a genome-wide RNAi screen for modifiers of cell number under Pvr sensitized conditions, we established a new proapoptotic role for the EcR complex, and an anti-apoptotic function for InR, in the balance of blood cell number in the Drosophila embryo. Phosphoproteomic analyses of Pvr deficient cells under low and high InR signaling states enabled the identification of common Pvr and InR phosphorylation targets regulating cell survival, and receptor-specific phosphorylation targets that mediate unique functions of Pvr and InR (model, Fig. 8). Our study further highlights the ability of signaling receptors to modulate their targets depending on cellular context, in our case, specifically based on the activity of other RTKs. These observations are important in light of mechanisms of acquired RTK inhibitor resistance that were recently described [68].
Previously, we demonstrated that Pvr mediates cell survival in the Drosophila embryonic hematopoietic system and in Drosophila Kc cells in culture [12]. Similar roles for Pvr in other cell populations such as glia were subsequently reported [69]. Here, we find that Pvr also contributes to the proliferation of Kc cells, which is revealed when Pvr-dependent cell death is suppressed. These Pvr functions are well conserved with mammalian systems, where PDGF/VEGF Receptors mediate cell survival and proliferation during normal development [70,71] and in pathologies such as leukemias and other forms of cancer [26,72,73].
Our findings encompassing the role of Pvr in the activation of the Mek/Erk and Akt/Tor pathways are consistent with previous reports of Pvr-dependent phosphorylation of Erk [21,23], the activation of the TOR1 Complex and Erk by Pvr [14], and the physical interaction of Pvr with PVRAP, Grb2, Shc, and the regulatory subunit of PI3K in cell culture [14,74]. Since our screen was designed to eliminate general regulators of cell number and instead focus on those genes that show differential effects under sensitized conditions, it predominantly revealed genes with tumor suppressor-like activities (Pvr Suppressors), many of which were not detected in conventional RNAi screens for cell proliferation or survival previously [17,75–78]. Several of the identified Pvr Enhancers (5/14) and half of the Pvr Suppressors scored as hits in other genome-wide RNAi screens examining RTK signaling, specifically InR and EGFR signaling using the same screening platform and dsRNA libraries [79]; see S3 Table for specific overlap).
Many genes identified in the screen regulate redundant pro-survival pathways downstream of Pvr (Fig. 8), as was predicted by our initial screening hypothesis (Fig. 1F), and which is also supported by others [14]. However, some regulators identified in the screen instead act in pathways parallel to Pvr signaling, as we demonstrated for InR and EcR signaling. Among the RNAi screen hits, we distinguished three major classes of modifiers. First, we identified a large group of ‘Upstream Genes’ that specifically affect cell number only in signaling competent, but not Pvr depleted cells. Among these, we found a large number of ribosomal protein genes. Interestingly, a recent Drosophila in vivo study identified ribosomal protein RpS8 as functional upstream regulator of Pvr in hemocytes of the lymph gland, proposing it may exert its function by interaction with Bip1 (bric à brac interacting protein 1), which shows similar phenotypes [16]. While our screen did not identify Bip1, it revealed RpS8 as putative Pvr ‘Upstream Gene’. Ribosomal subunits may promote Pvr expression also as part of the general translation machinery, or may play more specialized roles in translation regulation, according to previous reports on target-specific ribosomal activities that may influence the cellular signaling makeup in development and tumorigenesis [80–82]
Second, inherent to our system, our screen yielded relatively few Pvr Enhancers. From this group we chose InR for verification analysis by in vivo genetics, which we further complemented with a phosphoproteomic survey that illuminated synergy between Pvr and InR. Analogous synergistic relationships between InR with other RTKs have been reported in Drosophila development [69,83,84], and vertebrate signaling [85]. The specificity of redundant RTK signaling pathways is of major interest is the fields of cell signaling and cancer research and subject of ongoing intense study [68,86].
Third, the screen yielded a group of Pvr Suppressors, which function as tumor suppressor-like genes whose loss rescues cell survival under sensitized conditions. This group contains all negative regulators of the Akt/Tor pathway, many of which are known tumor suppressors in mammalian systems [39–41], and several negative regulators of the Mek/Erk pathway such as mts and wdb, encoding for components of the PP2a complex [42,43], and Mkp3, which encodes for a phosphatase known to negatively regulate Erk [44]. As expected, several genes identified in the Pvr modifier screen also scored in previous screens for signaling mediators of the Pvr, Akt/Tor and RTK/Erk pathways [14,83,87]. The screen also revealed novel, or only recently characterized, genes. CG6182 is an ortholog of the mammalian TBC7, that interacts physically with Tsc1 [88]; GckIII is a counterpart of mammalian Serine Threonine Kinase 25 (STK25), also known as SOK1, that localizes to the Golgi [89] and induces cell death upon overexpression in mammalian cell culture [90]. Some of the identified genes have been characterized in Drosophila, yet no role in cell number control in the embryo has been described. For example, we identified multiple members of the Brahma SWI2/SNF2 family ATPase chromatin-remodeling complex [46,47], with osa and dalao scoring as Pvr Suppressors, and Brahma associated protein 60kD (Bap60) and moira (mor) scoring in mixed categories. Two of the strongest hits among the Pvr Suppressors were genes encoding the nuclear hormone receptors EcR and Usp [27,45], which we followed up with subsequent analyses.
EcR and Usp have previously been studied for their roles in proliferation, differentiation and cell death during larval molting and metamorphosis [50,61,91]. In Kc cells, the EcR/Usp ligand ecdysone has been known to arrest the cell cycle and trigger a cell differentiation program [52–54]. However, neither in the embryo nor in Kc cells has ecdysone signaling been previously associated with cell death [51,92]. Here, we describe a role for ecdysone signaling in embryonic cell death, a function revealed only under sensitized conditions or when directly stimulating EcR pathway activity. When treating Kc cells with 20E, we find that the EcR targets E93 and rpr are transcriptionally upregulated, consistent with previous reports describing these genes as transcriptional targets of EcR [93]. E93 and Rpr drive apoptosis [57,94–96] and are required for ecdysone-induced death of the larval midgut and salivary glands during metamorphosis and in the larval cell line l(2)mbn [56,57,97–100]. As for the mechanism of cell death rescue by EcR silencing, we were unable to detected measurable levels of Halloween gene expression, which is required for biosynthetic maturation of 20E [101]. We therefore propose that Kc cultures produce low levels of 20E through low-level expression of Halloween genes, or the EcR complex may have residual pro-death functions even in the absence of ligand. Previous publications have suggested that the unligated EcR complex has an active role and can bind to ecdysone response elements [102,103].
Vertebrate counterparts of EcR and Usp are the liver X receptors (LXRs), and retinoid X receptor (RXR), respectively [104,105]. RXR plays central roles in cell proliferation, apoptosis, and differentiation [106–108] during development and in pathologies such as cancer and metabolic disease [109,110]. Lack of activation of the RXR/Retinoic acid receptor (RAR) pathways causes Acute Promyelocytic Leukemia (APL) and other malignancies due to impairments in cell differentiation and increased cell survival [109,111,112] and treatment with synthetic retinoids or rexinoids has proven promising in reverting malignant phenotypes [109,111]. Interestingly, dependence on additional anti-apoptotic pathways has been reported in RxR-dependent APL. In particular, Akt/Tor signaling contributes to the increased cell survival in APL, and, consequently, dual therapy with PI3K inhibitors and retinoids has shown great therapeutic promise [113].
Drosophila InR and Akt/TOR signaling were recently reported in several studies for their multifaceted roles in the regulation of lymph gland hemocytes, an independent blood cell lineage in Drosophila [114–116]. Drosophila InR further promotes the trophic survival of germline stem cells in the Drosophila ovary [117], which relies on its downstream mediator Tor [118]. Besides this, Drosophila InR is known mostly for its role in cell growth and regulation of body and organ size [119]. In contrast, the related mammalian Insulin-like Growth Factors 1 and 2 (IGF1 and IGF2) play important roles in the trophic survival of various cell types [120]. Our study now demonstrates that Drosophila InR also exerts trophic function in embryonic blood cells, which is revealed once redundant receptor activity such as Pvr activity is suppressed.
Several mammalian orthologs of InR phosphorylation targets uncovered from our phosphoproteomic analyses have been previously reported to be regulated by insulin, such as Ssrp, Larp, eIF4G, Lamin, NAT1, Claspin, Garz, Nedd4, Nopp140, Yki and Lk6 [121,122]. These phosphoproteins are also regulated by Pvr and contribute to the roster of common RTK targets that likely account for Pvr/InR-induced cell survival. An additional example of a commonly targeted phosphosite is the activating phosphorylation of Stat92E [123], a proposed target of the insulin receptor [124]. These examples highlight the success of our phosphoproteomic approach to uncover bona fide targets shared by InR and PVR. The approach also unveiled novel downstream effectors. For example, the requirement of Pvr for Ssrp phosphorylation hints to a relationship between Pvr and the chromatin remodeling FACT complex, a heterodimer comprised of Ssrp and Dre4 [125]. This hypothesis is reinforced by 1) the suppression of Pvr deficient cell proliferation by dre4 knockdown; and 2) a reported two-hybrid interaction between Pvr and Spt6 [126], a component of an elongation complex that includes FACT [127]. The possibility that FACT functions downstream of these RTKs to regulate transcriptional initiation and elongation as a cell survival mechanism will be an interesting area of future investigation.
Our phosphoproteomics experiments additionally uncovered InR-specific phosphorylations: e.g. phosphosites on Foxo, Unkempt (Unk), Chico, Tsc1, Spaghetti (Spag), L, Ajuba (Jub), and Git. Notably, many of these proteins were identified by affinity purification and mass spectrometry as components of an InR/Tor protein interaction network [128], supporting our proposition that indeed these phosphoproteins serve InR-specific functions. Remarkably, four of the thirty high confidence Pvr Suppressors and two of the fourteen high confidence Pvr Enhancers exhibited altered phosphorylation specifically under InR activation indicating these localized phosphorylation events may be critical for the rescue of Pvr deficient cells provided by InR stimulation.
Our analysis identified very few EcR-dependent phosphoproteins, however, we cannot rule out that these few may indeed regulate cell number. For example, phosphorylation of the methionine sulfoxide reductase Ecdysone-induced protein 28/29kD (Eip71CD) was upregulated by EcR knockdown. Eip71CD confers protection to oxidative stress, increases cell size and number, and promotes longevity [129,130]. Additionally, phosphorylation of the diacylglycerol O-acyltransferase Midway (Mdy) was upregulated upon EcR knockdown. mdy mutant egg chambers exhibit premature nurse cell death and degeneration during mid-oogenesis [131] comparable to EcR and Eip75B germline clones [132]. We observed an upregulation of phosphorylated Transforming acidic coiled-coil (Tacc) in Pvr deficient cells subjected to either EcR knockdown. Vertebrate TACC proteins interact with RxRβ to regulate specific gene expression [133]. Phosphorylation could potentially influence Tacc interaction with Usp, the Drosophila ortholog of RxR, and consequently impact Usp-dependent gene expression, thereby permitting cell survival. We cannot exclude the possibility that, due to incomplete coverage, our phosphoproteomics analyses may have failed to capture critical phosphorylation changes induced by EcR knockdown that account for Pvr deficient cell survival. Our analyses did, however, generate a list of candidates for future study and highlight the substantially different responses by insulin and EcR knockdown to rescue Pvr loss.
The dual dependence of Drosophila Kc cells and embryonic hemocytes on the Akt/TOR and Mek/Erk pathways (this study and [12,14] echoes the dependence of many mammalian cells, in particular tumor cells, on these two signaling pathways. In addition to concomitant activation by upstream receptors, Akt/Tor and Mek/Erk signaling further show a substantial degree of crosstalk between each other [134]. Dual inhibition of these two pathways has therefore become a promising approach in targeted cancer therapies [135]. However, many molecularly targeted approaches remain challenging due to the plasticity of signaling, the involvement of additional undefined redundant signaling pathways, and the variation of signaling networks downstream of even closely related receptors [68,135]. Findings from the Drosophila model provide insight into the pools of common and unique signaling targets of the related RTKs Pvr and InR. Further, this study suggests that the qualitative signaling specificity of receptors can be switched in response to the signaling status of the cell. This notion may be of wide-reaching consequences for many cellular processes, and requires careful consideration when aiming for the experimental or therapeutic manipulation of signaling systems.
Fly lines used were: Pvr1/CyO [12], srpHemoGAL4 [12], Pxn-GAL4 [136], UAS-PvrΔC [12], UAS-p35 [30], UAS-srcEGFP (E. Spana), UAS-lacZnls (E. Spana), UAS-mCD8::GFP [137], UAS-Stinger [138], UAS-EcRA, UAS-EcRB1, UAS-EcRB2 [57], UAS-EcRB1 W650A (dominant-negative) [59] and UAS-EcR A W650A (dominant-negative) [60], UAS-InR-dn [139].
For in vivo quantification of hemocytes in Drosophila embryos, the srp-Hemo-GAL4 driver was used to express UAS-lacZnls and UAS-srcEGFP in hemocytes. Genotypes of Pvr1 mutant rescue/enhancement experiments were: Pvr1,UAS-srcEGFP/ Pvr1,srpHemoGAL4; UAS-p35/UAS-lacZnls and Pvr1,UAS-EcR (A or B1) W650A or UAS-InR-dn/Pvr1,srpHemoGAL4; UAS-mCD8::GFP/UAS-lacZnls. Genotype for the alternative rescue/enhancement of Pvr dominant-negative expressing hemocytes were: srpHemoGAL4, UAS-srcEGFP/+; UAS-PvrΔC, UAS-EcRA W650A or UAS-InR-dn/ UAS-lacZnls.
Embryos were collected on apple juice agar plates and fixed and stained as described previously [12]. Antibodies used were goat anti-GFP (1:1500) (Molecular Probes) and mouse anti-β-Gal (1:750) (Promega), and Alexa Fluor secondary antibodies (Invitrogen) Imaging was done on Leica DMI 4000B and Leica SP5 microscopes. Hemocyte counts were conducted under fluorescent microscopy at 40X, assessing 10 independent embryos per genotype and stage. Standard deviations and p values by Student’s t-test were calculated.
To examine hemocyte numbers at the embryo-larva transition, hemocytes were marked by Pxn-GAL4 driven expression of UAS-Stinger. The transgenic driver UAS-Stinger; Pxn-GAL4 was crossed to w1118 (control), or UAS-EcRA dn, respectively. Dechorionated embryos or larvae from 2 hour timed collections were mounted under glass slides and subjected to visual/manual counting under a fluorescence microscope. At least ten embryos or larvae per time point and genotype were assessed. Standard deviations and p values by Student’s t-test were calculated.
Kc167 cells [140], here labeled Kc, were cultured in Schneider’s Drosophila Medium (Millipore, Gibco) supplemented with 10% Fetal Bovine Serum (FBS) and 1000 units/ml Penicillin and 1000mcg/ml Streptomycin. Insulin was supplemented to a final concentration of 5 μg/ml for InR stimulation experiments.
20E experiments: 20-Hydroxyecdysone (Sigma-Aldrich), 20E, was dissolved in ethanol to make a 5mg/ml stock. A subsequent stock of 1μg/ml stock was made by diluting in water. 1x105 cells were seeded into each well of a 24 well plate and 20E was added to achieve the indicated final concentrations.
All cell experiments were based on Kc167 cells, in short Kc. Effectene Transfection Reagent (Qiagen) was utilized for transfection. Kc cells were co-transfected with driver Actin-GAL4, UAS-puromycin, UAS-GFP and UAS-p35 plasmid constructs. Three days after transfection, cells were selected with puromycin 10ug/ml. After 2 weeks, surviving cells were harvested and sorted by Fluorescence-activated cell sorting (FACS) to isolate the highest 20 percentile of GFP-expressing cells. To further select cells that are resistant to apoptosis, thread RNAi knockdown was used to eliminate cells with weak resistance to caspase-dependent apoptosis. The surviving cells were expanded for experimental use. The presence of p35 transgene in the p35 stable cell pool was confirmed by PCR verification.
RNAi knockdown was performed as described previously [141]. Briefly, Kc167 cells were re-suspended and diluted in serum-free medium before seeding. dsRNAs targeting each specific gene were added and incubated for 45 minutes before supplementing with complete medium with FBS to adjust to a final concentration of 10% FBS.
We screened a set of 62 384-well plates that were pre-arrayed with dsRNAs, corresponding to 22,914 distinct amplicons based on Flybase release 5.51 of the Drosophila genome corresponding to 13,777 unique genes [31], and 7463 Sanger predictions [142] (DRSC). To determine differential effects between Pvr silenced and control cells, we screened each plate under two conditions, dsRNA-mediated knockdown of Pvr, or knockdown of a control (GFP). All experiments were performed in duplicate. Each well contained 0.25ug of pre-arrayed dsRNA. Before seeding, Kc cell suspensions were pre-mixed with Pvr or control (GFP) dsRNAs in batch, corresponding to a final concentration of 0.3ug per well. Cells were seeded at a density of 7,000 cells/well and incubated for 4 days. CellTiter-Glo assay (Promega) was performed according to the instructions of the manufacturer, and luminescence was read using Analyst GT or SpectraMax plate readers (Molecular Devices). Liquid handling was performed using WellMate (Matrix), MicroFill (BioTek), or MultiDrop (Thermo), high-throughput dispensers. Z scores [z = (χ-μ)/σ] were calculated as follows: μ = Mean of readings from controls wells (i.e. wells without pre-arrayed candidate dsRNAs), σ = Standard deviation from readings of the control wells. χ = Reading of candidate gene well. Z score for Pvr knockdown condition (Z[Pvr]) and for control knockdown condition (Z[GFP]) were generated and the differential effects in Pvr knockdown condition and control knockdown were calculated by the difference of each Z scores (i.e. Zdiff = Z[Pvr]- Z[GFP]). Cluster analysis of primary screen data was performed of amplicons scoring ZDiff> = 2.0 and ZDiff< = -2.0, using Z[Pvr], Z[GFP], and Zdiff values for each amplicon. Analysis included hierarchical clustering by centered correlation, and complete linkage, and results were displayed using TreeView [143]. For verification screening, genes were selected at cutoffs of ZDiff> = 2.2 and ZDiff< = -2.2, and one or two amplicons per gene, non-overlapping with the primary screen amplicons, and devoid of 19bp off-target overlaps, were utilized (DRSC). Secondary screening involved differential screening of Pvr RNAi and GFP RNAi cells as outlined above. Assays were performed in replicate and repeated in two independent duplicates. Final, ‘high confidence’ Pvr modifiers were determined by calculation of final average ZDiff scores determined from the averaged ZDiff scores of all amplicons targeting specific genes that were evaluated in both primary and secondary screening. Regarding the error rate of the screen, we generated false positive and false negative rates as follows: To evaluate false positives we i) assembled a list of 355 protein-coding genes that are not expressed across Drosophila tissue/stage/cell lines based on both modEncode RNA-Seq data as well as FlyAtlas data; ii) compared this list with genes scoring in the primary screen (there is only 1 gene overlapping and the relevant amplicon has >5 predicted off targets); and iii) estimated a false positive rate: 1/355 = <1%. To evaluate false negatives we i) assembled a list of 38 high confidence genes based on secondary screening hits, which are the genes that scored with at least 2 independent amplicons and each amplicon was consistently scored among replicates; ii) identified all amplicons relevant to these 38 genes from the genome library and found 80 of them. 41 scored in the primary screen while 39 failed to score; and iii) used these numbers to calculate a false negative rate: 39/80 = 49%.
To obtain cell counts, 1x105 cells were seeded into each well of a 24-well plate followed by treatment with dsRNAs or 20E. 3.3ug of dsRNA targeting each specific gene knockdown was added. After culturing for the indicated period of time, cells were re-suspended and diluted 1:1 with 0.4% Trypan Blue. Numbers of viable/dead cells were assessed by hemocytometer counting based on Trypan Blue exclusion/staining.
For EdU and TUNEL assays, 20,000 Kc cells were seeded into each well of 96-well black clear bottom plate and immediately treated with dsRNAs or 20E. 0.825ug of dsRNA was used to target each specific gene knockdown.
To assess cell proliferation, cells were incubated for 4 hours with 10uM of Click-iT EdU (Invitrogen) one or several days after the dsRNAs or 20E treatment. The Click-iT EdU cell proliferation assay was conducted according to manufacturer's instructions.
For assessing cell death, cells were processed by TUNEL assay according to manufacturer's instructions (Invitrogen).
Stained cells were counted visually/manually and by ImageJ automated cell quantification. In brief, for ImageJ analysis, still images were converted to 8 bit images and cells were selected by setting a threshold against the background. Highlighted cells were then counted by the ‘Analyze Particles’ function. At least three still images for each sample were taken at random sites using a 40X objective. Percentages of EdU or TUNEL positive cells were calculated as follows: (# of EdU or TUNEL positive cells/ total # of cells) * 100. Cell culture figures show compilations of three independent biological replicate experiments. Error bars indicate standard deviation. Student’s t-test as indicated. * for p < 0.05; ** for p < 0.01; *** for p < 0.001; NS for not significant.
In most cases, dsRNA amplicon sequences were selected by the Drosophila RNAi Screening Center (DRSC), as indicated by DRSC amplicon numbers. Primers used for generating the amplicon template contained a 5' T7 RNA polymerase-binding site (TAATACGACTCACTATAGG) following by the amplicon specific sequences. dsRNAs were generated by in vitro transcription using Megascript T7 transcription kit (Ambion). dsRNAs were purified with a RNeasy Mini Kit (Qiagen) and product size confirmed by agarose gel electrophoresis. dsRNA concentrations were measured with a Nanodrop 2000C spectrophotometer (Thermo Scientific).
Total RNA was extracted using a RNeasy mini kit (Qiagen), according to manufacturer's instructions. Total purified RNA were measured with a Nanodrop 2000C spectrophotometer (Thermo Scientific). 1ug—0.1ug of purified RNA was reversed transcribed into cDNA using an iScript cDNA synthesis kit (Biorad). Real time PCR reactions were carried out using iQ SYBR Green Supermix (Bio-Rad) on a Bio-Rad CFX96 Real Time System and gene expression levels were analyzed with CFX Manager Software (Bio-Rad). Primers for real time PCR assays were designed using web-based software ProbeFinder (Roche Applied Science Universal ProbeLibrary Assay Design Center) or by the author. Primer sequences for real-time PCR assessment will be made available upon request.
Kc cells were lysed using Triton lysis buffer (50mM Tris-HCl (pH 7.5), 150mM NaCl, 1% Trition X-100, 30mM NaF) freshly supplemented with 1mM Na3VO4 and protease inhibitors (Complete, Roche) and immunoblot analysis was performed as described previously [12]. Primary antibodies were obtained from Cell Signaling Technology except monoclonal anti-β-tubulin (Sigma T5168), anti-Pvr [12], anti-EcR (Developmental Studies Hybridoma Bank, DSHB) and anti-histone H3 (Abcam 39950); signal was detected by HRP conjugated secondary antibodies (Amersham NA934V/NXA931 and Jackson ImmunoResearch 706–035–148) and ECL.
Kc cells were serum starved for 1 hr, incubated with dsRNA for 30 minutes and then diluted either in Schneider’s Drosophila Medium (Gibco) supplemented with Fetal Bovine Serum (FBS) (final concentration of 10%), Penicillin (50 units/ml final concentration), and Streptomycin (50 ug/ml final concentration), with or without insulin (5ug/ml final concentration). After two days cells were lysed in: 8M urea, 75mM NaCl, 50mM Tris-HCl pH 8.2, 1mM NaF, 1mM β-glycerophosphate, 1mM sodium orthovanadate, 10mM sodium pyrophosphate, 1mM PMSF, EDTA-free Protease Inhibitor Cocktail Tablet (Roche). One milligram of protein from each sample was reduced with 5mM dithiothreitol at 56°C for 25 minutes. Cysteines were alkylated with 14mM iodoacetamide for 30 minutes at room temperature in the dark. Unreacted iodoacetamide was quenched by incubation with additional dithiothreitol to 5mM for 15 minutes at room temperature in the dark. Lysates were diluted 1:5 with 25mM Tris-HCl, pH 8.2 and CaCl2 added to 1mM. Digestion with 5ug sequencing grade trypsin (Promega) was overnight at 37°C with agitation. Peptides were acidified with 10% trifluoroacetic acid and desalted using 1cc Sep-Pak tC18 solid-phase extraction cartridges (Waters). Eluted peptides were lyophilized, resuspended in 200mM Na-HEPES pH8.2, and labeled with TMT reagent (Thermo Scientific) in anhydrous acetonitrile (2mg TMT reagent per sample) for 1 hour at room temperature. TMT labeling was as follows:
Experiment 1. control dsRNA, biological replicate #1: 126 (high Pvr, low InR); control dsRNA, biological replicate #2: 127 (high Pvr, low InR); control dsRNA + insulin, biological replicate #1: 128 (high Pvr, high InR); control dsRNA + insulin, biological replicate #2: 129 (high Pvr, high InR); control dsRNA + EcR dsRNA, biological replicate #1: 130 (high Pvr, low EcR); control dsRNA + EcR dsRNA, biological replicate #2: 131 (high Pvr, low EcR).
Experiment 2. Pvr dsRNA + control dsRNA, biological replicate #1: 126 (low Pvr, low InR); Pvr dsRNA + control dsRNA, biological replicate #2: 127 (low Pvr, low InR); Pvr dsRNA + control dsRNA + insulin, biological replicate #1: 128 (low Pvr, high InR); Pvr dsRNA + control dsRNA + insulin, biological replicate #2: 129 (low Pvr, high InR); Pvr dsRNA + EcR dsRNA, biological replicate #1: 130 (low Pvr, low EcR); Pvr dsRNA + EcR dsRNA, biological replicate #2: 131 (low Pvr, low EcR).
Experiment 3. Pvr dsRNA + control dsRNA: 126 (low Pvr, low InR); Pvr dsRNA + control dsRNA + insulin: 127 (low Pvr, high InR); Pvr dsRNA + EcR dsRNA: 128 (low Pvr, low EcR); control dsRNA: 129 (high Pvr, low InR); control dsRNA + insulin: 130 (high Pvr, high InR); control dsRNA + EcR dsRNA: 131 (high Pvr, low EcR).
Reactions were quenched by the addition of hydroxylamine to 0.3% and incubation at room temperature for 15 min. Labeled peptides were combined, lyophilized, and stored at -80°C until further processing. Samples were acidified with 10% trifluoroacetic acid and desalted using a 3cc Sep-Pak tC18 solid-phase extraction cartridge (Waters). Phosphopeptides were enriched by strong cation exchange chromatography (SCX; [144]). Lyophilized peptides were resuspended in 400 ul SCX buffer A (7 mM KH2PO4, pH 2.65, 30% acetonitrile) and injected onto a SCX column (Polysulfoethyl aspartamide, 9.4 mm×250mm, 5 uM particle size, 200 Ǻ pore size, PolyLC). A gradient was developed over 35 min from 0% to 30% buffer B (7 mM KH2PO4, pH 2.65, 30% acetonitrile, 350 mM KCl) at a flow rate of 2.5 ml/min. 12 fractions were collected and lyophilized. Peptides were then desalted with 1cc Waters Sep-Pak tC18 solid-phase extraction cartridges and subjected to TiO2 based phosphopeptide enrichment [145] using 0.5mg titanium dioxide microspheres per mg protein. Eluates were further desalted using STAGE tips [146] and lyophilized. Samples were reconstituted in 4ul 5% formic acid / 5% acetonitrile.
In most signaling systems, the major gatekeeper of signal transduction is protein phosphorylation, which can be adjusted rapidly according to the needs of a cell. A caveat of solely measuring phosphorylation is that a change in phosphopeptide levels for any particular peptide can result from a change in phosphorylation of the peptide, or from a change in levels of that protein. We expect that roughly a quarter of altered phosphorylation we observe would be explained by the latter mechanism, based on previous reports (Bodenmiller et al. Science Signaling 2010 and Wu et al. Mol. Cell Proteomics, 2011; Sopko et al. Dev Cell 2014). Given that coverage of the Drosophila proteome is not yet comprehensive in a single mass spec run, normalization of phosphorylation to protein amounts can only be estimated. Further, lowly expressed proteins are often missed. For these reasons, we chose to focus exclusively on phosphorylation in our study. Samples were subjected to LC-MS/MS with an Orbitrap Velos Pro mass spectrometer (Thermo Scientific) using higher-energy collision dissociation (HCD; [147]) and a top ten method [148]. MS/MS spectra were searched against a composite database of Drosophila melanogaster proteins derived from Flybase version 5.23 in both the forward and reverse orientation using the Sequest algorithm [149]. Search parameters included: a precursor mass tolerance of 20 ppm; up to two missed cleavages; static modification of TMT tags on lysine residues and peptide N termini (+229.162932 Da) and +57.021464 Da accounting for carbamidomethylation on Cys; dynamic modification of phosphorylation (+79.966330 Da) on Ser, Thr and Tyr and oxidation (+15.994915 Da) on Met. A target-decoy database search strategy [150] enabled thresholding of the false discovery rate (FDR) for MS/MS spectral assignment at 1%. Correct spectral matches were distinguished from incorrect matches using linear discriminant analysis based on parameters including Xcorr, ΔCn, precursor mass error, peptide length, and charge state [151]. The localizations of individual phosphorylations were assigned using the probability-based AScore algorithm [152] and only phosphosites with AScores greater than 13 (p < 0.05) were considered in our analysis. Moreover, only phosphopeptides with isolation specificity greater than 0.75 were considered for further analysis. Further filtering of the dataset resulted in a final protein FDR of ~2% and a peptide FDR near 0.15%. TMT labeling was >98% efficient. For TMT reporter ion quantification, a 0.03 Da window centered on the expected mass of each reporter ion was monitored and the intensity of the signal closest to the expected mass was recorded. Reporter ion signals were further adjusted to correct for impurities associated with each TMT label, as described elsewhere [153]. Raw TMT reporter ion intensities for individual phosphopeptides were normalized to the summed reporter ion intensity for each TMT label. Adjusted reporter ion intensities were averaged between replicates. Peptides for which only one replicate TMT labeled sample generated detectable reporter ions were excluded from further analysis.
Complete information on Pvr modifier screen data, and DRSC library dsRNA amplicons can be accessed at http://www.flyrnai.org/.
|
10.1371/journal.pgen.1007100 | UNC-16/JIP3 regulates early events in synaptic vesicle protein trafficking via LRK-1/LRRK2 and AP complexes | JIP3/UNC-16/dSYD is a MAPK-scaffolding protein with roles in protein trafficking. We show that it is present on the Golgi and is necessary for the polarized distribution of synaptic vesicle proteins (SVPs) and dendritic proteins in neurons. UNC-16 excludes Golgi enzymes from SVP transport carriers and facilitates inclusion of specific SVPs into the same transport carrier. The SVP trafficking roles of UNC-16 are mediated through LRK-1, whose localization to the Golgi is reduced in unc-16 animals. UNC-16, through LRK-1, also enables Golgi-localization of the μ-subunit of the AP-1 complex. AP1 regulates the size but not the composition of SVP transport carriers. Additionally, UNC-16 and LRK-1 through the AP-3 complex regulates the composition but not the size of the SVP transport carrier. These early biogenesis steps are essential for dependence on the synaptic vesicle motor, UNC-104 for axonal transport. Our results show that UNC-16 and its downstream effectors, LRK-1 and the AP complexes function at the Golgi and/or post-Golgi compartments to control early steps of SV biogenesis. The UNC-16 dependent steps of exclusion, inclusion and motor recruitment are critical for polarized distribution of neuronal cargo.
| Synaptic vesicles (SVs) have a defined composition and size at the synapse. The multiple synaptic vesicle proteins (SVPs) found on these vesicle membranes are synthesized at and trafficked out of the cell body in distinct transport carriers. However, we do not yet understand how different SVPs are sorted and trafficked to the synapse. We show that UNC-16/JIP3 plays a critical role, in a series of essential steps, to ensure proper membrane composition and size of the ensuing SVP carrier exiting the cell body. These processes are “exclusion” of resident Golgi enzymes followed by the “inclusion” of synaptic vesicle proteins in the same transport carrier. Regulation of composition and size seems to occur independently of each other and depends on two distinct AP complexes acting downstream to LRK-1. Our study further indicates that the composition of the transport carrier formed is important for the recruitment of motors and consequently for the polarized localization of SVPs.
| The secretory pathway in a cell involves the synthesis and trafficking of proteins through the ER-Golgi network and their subsequent targeting to different sub-cellular compartments. Generation of a defined transport carrier, with a characteristic protein and lipid composition, along the trafficking pathways is known to involve at least three steps (a-c, see below), several occurring at the trans-Golgi network (TGN). (a) Protein sorting where the secretory cargo is segregated away from Golgi resident proteins [1–4]. For example, segregation of different Regulated Secretory Proteins (RSP) such as POMC occurs via receptor-mediated sorting [5,6]. (b) A post-sorting step where clustered cargo undergoes budding and separation to form a vesicular compartment from the donor membrane. In part, the adaptor protein (AP) complexes regulate such steps by recognizing signal sequences on proteins and ensuring that they are sorted into appropriate compartments [7,8]. The AP-1 complex recruits proteins like Clathrin, which causes membrane deformation followed by budding and scission from the TGN and post-Golgi compartments [9,10]. (c) A third step is the recruitment of specific motors, dependent on the characteristic membrane composition, constituting proteins and/or lipids, of the newly formed cargo. For example, AP-1 interacts with the Kinesin-3 motor KIF13A to coordinate endosomal sorting during melanosome biogenesis [11]. These events ensure the formation of a defined transport carrier that gets targeted to a specific sub-cellular compartment. Protein sorting occurs at post-Golgi compartments as well. For example, during the multi-step maturation of secretory granules, sorting of proteins also occur post-Golgi at intermediate compartments known as the immature secretory granules (ISGs) [12]. However, the genes that regulate such processes remain to be well understood.
Synaptic vesicle proteins (SVPs) are essential for neurotransmission and synaptic vesicles at the synapse are known to have a defined composition [13]. Several SVPs have transmembrane domains and are trafficked out of the TGN through the regulated secretory pathway to the synapse. Each synaptic vesicle found at the pre-synaptic bouton has an assortment of proteins important for processes such as neurotransmitter filling (VGLUT1 [14]), docking and neurotransmitter release (SNB-1, SNT-1 [15–17], and fusion of synaptic vesicles with the plasma membrane (UNC-13 [18]). It is not clear whether all of these proteins are found on a single SVP transport carrier as it exits the cell body. An earlier study indicates that different SVPs, for example Synaptophysin and SV2 are associated with distinct pools of membranous organelles [19]. This could imply that different SVPs might travel in separate compartments before they come together in a mature synaptic vesicle that is found at the synapse. However, another study suggests that most or all SVPs, including Synaptobrevin and SV2, are transported in a single transport carrier to the presynaptic active zone [20]. Studies using PC12 cell lines and in vivo studies in C. elegans have shown that the AP-3 complex is required for synaptic vesicle biogenesis, potentially directly from the Golgi, and for the axonal targeting of SVPs respectively [21–23]. SVPs are also known to require the molecular motor KIF1A for their exit from the cell body and transport to the synapse in multiple systems [24,25]. Thus, although proteins such as AP-3 and KIF1A have been implicated in SVP transport carrier biogenesis and trafficking, several aspects of these early steps remain unclear. For example, it is not fully understood how the AP complexes confer specificity to the sorting of different SVPs. Mechanisms that ensure separation of SVPs and/or lipids of the cargo membrane from those integral to the Golgi membrane is also poorly understood. Further, we do not know in what order these multiple steps are carried out during SVP trafficking and transport carrier biogenesis
JIP3/UNC-16/dSYD, a JNK-signaling scaffold protein present at the Golgi in Drosophila, is thought to have roles in protein trafficking [26–28]. The C. elegans unc-16 mutants also show mis-trafficking and mis-accumulation of multiple neuronal cargo such as SVPs, dendritic receptors, lysosomes and early endosomes [28,29]. It is unclear how UNC-16, a molecule known to interact with and scaffold multiple kinases, including MAPK family members, carries out its trafficking roles. Among the UNC-16-interacting proteins, LRK-1, the C. elegans homolog of LRRK2, has been previously implicated in the polarized trafficking of SVPs and is also found at the Golgi, like UNC-16 [30,31]. Thus, UNC-16 and LRK-1 could play roles in the early steps of SVP protein trafficking.
In this study, we show that the SVP transport carriers formed in unc-16 and lrk-1 mutants have an altered composition and size. These mutants also show defects in polarized distribution of SVPs, which mis-localize to the dendrites [29,30]. We show that UNC-16 regulates the composition via LRK-1 by excluding Golgi enzymes from the SVP transport carrier and by increasing the incidence of co-transport of SNB-1 and RAB-3 in the same transport carrier. We show that size of the SVP transport carriers formed is determined by UNC-16 via LRK-1-dependent localization of the μ-component (UNC-101) of the AP-1 complex at the Golgi. We also show that the UNC-16 through LRK-1 and the AP-3 complex ensures that specific SVPs are included in the same transport carrier. Thus, based on genetic and biochemical evidence, we propose a novel role of UNC-16 in synaptic vesicle protein trafficking wherein it functions via LRK-1 to regulate exclusion and inclusion of proteins and consequently motor recruitment on the carrier. The AP-1 and AP-3 complexes likely function at the Golgi and/or post-Golgi intermediate compartments downstream to UNC-16 and LRK-1, to regulate respectively the size and composition of the SVP transport carrier formed in the neuronal cell body.
UNC-16/JIP3/dSYD has been implicated in the trafficking of multiple proteins such as those associated with synaptic vesicles and lysosomes [28,29,32]. To investigate the role it plays in regulating protein trafficking, we first examined the localization of UNC-16. Using two independent Golgi markers, Mannosidase-II (Man-II) and RUND-1 [33], we show that UNC-16 localizes to the Golgi in C. elegans neuronal cell bodies (Fig 1a). A similar localization of mammalian dSYD/JIP3 has been reported in CV-1 epithelial cells [26]. We isolated a new allele of unc-16 (tb109), with an early stop codon at amino acid 423, that caused mis-trafficking of the trans-membrane VAMP Synaptobrevin-1 (SNB-1) to the dendrite of the amphid sensory neurons (Methods, S1a and S1b Fig, S1 Table). This phenotype is identical to those reported in other unc-16 alleles [29]. The dendritic mis-localization of SNB-1 in tb109 was rescued by the transgenic expression of wild type UNC-16 (S1b Fig, S1 Table). Upon examination of a dendritic receptor ODR-10 in the AWB neuron we found that, unlike in wild type animals, ODR-10 is ectopically localized to the axonal compartment in unc-16 mutants (S1c Fig).
Loss of UNC-16 thus leads to the loss of polarized distribution of cargo in neurons. This mis-trafficking is not dependent on the orientation of microtubules, which is similar in both unc-16 and wild type (S1d Fig). The defects in axonal and dendritic targeting along with the reported trafficking defects in dendritic, endosomal and lysosomal proteins suggests that UNC-16 acts as a general regulator of early events in the trafficking pathway [27–29,34].
To understand the nature of early defects in unc-16, potentially occurring at the Golgi, we examined both—the localization of Golgi enzymes and whether cargo such as SVP transport carriers had altered composition.
Unlike wild type, in unc-16 mutants, the Man-II enzyme mis-localizes to dendritic tips of the ASI neuron (S2g Fig, S1 Table) and both Man-II and Sialyl transferase (ST) are present as discrete compartments throughout the touch receptor neuron (TRN) process, up to the synapse (Fig 1b, S2b Fig), similar to reported observations [28,34]. Comparable to wild type, in unc-16 mutants Golgi resident enzyme Man-II and ST continue to localize as 1–3 large puncta in the cell body (Fig 1b, S2b and S2d Fig). Other Golgi markers, such as RUND-1 and RAB-6.2, show a punctate distribution in the neuronal cell body of unc-16 mutants, very similar to those observed in wild type animals, with no gross changes in the number or position of Golgi puncta (S2d Fig). This suggests that potentially only a subset of Golgi markers is mis-trafficked into the axons of unc-16 animals.
In order to check if the mis-trafficking of Golgi proteins seen in unc-16 alters SVP transport carriers, we carried out dual colour imaging of both Man-II and RAB-3. About 86% of RAB-3 containing compartments emerging from the cell body carry the Golgi enzyme Man-II, unlike in wild type where only ~ 5% of RAB-3 marked compartments co-transport Man-II (Fig 1d, S1 and S2 Movies). Similar observations were also made with the Golgi enzyme ST (S2a Fig). The mis-trafficking of Man-II along with RAB-3 into the neuronal process in unc-16 mutants is rescued by the transgenic expression of wild type UNC-16 (Fig 1d).
Thus, UNC-16 is necessary to restrict Golgi resident proteins to the cell body and to exclude them from SVP transport carriers. The defects seen in unc-16 are likely due to disruption of retention/retrieval mechanisms, leading to mis-trafficking of Golgi enzymes.
To assess the membrane composition of the atypical SVP transport carriers formed in unc-16, we examined the co-transport of two synaptic vesicle proteins, RAB-3 and SNB-1, in TRNs. In wild type animals, the incidence of co-transport of RAB-3 and SNB-1 is ~ 39% of the mobile SVP transport carriers. In unc-16 however, the frequency of co-transport of RAB-3 and SNB-1 reduces by half to ~19% (Fig 1c and 1e). This was corroborated by the reduction in co-localization of endogenous transmembrane Synaptotagmin (SNT-1) and RAB-3 at non-synaptic regions of the sub-lateral cord from ~95% in wild type to ~60% in unc-16 animals (S1e Fig). Unlike in wild type, these erroneous SVP transport carriers exit the cell body as long tubular compartments. These long compartments show a ~1.5-fold increase in average length compared to those found in wild type and are three times more frequent in multiple alleles of unc-16 across several neuronal cell types (Fig 2a, S3a Fig, S2 Table, S3–S5 Movies). Electron micrograph analyses showed an increase in the width of vesicles in non-synaptic regions of the dorsal and ventral nerve cord in unc-16 compared to wild type animals (Fig 2b). Such larger vesicular profiles have been previously seen at the synapses of unc-16 animals [34], [35]. The longer SVP transport carriers we see in our live imaging could contribute to these wider vesicular profiles observed in our electron micrographs (S4 Movie). We also verified that mutants in known interactors of UNC-16 such as jnk-1, unc-116 and dhc-1 [29,32,36] do not show these large moving tubular profiles carrying either SNB-1 or RAB-3, nor do they mis-localize the Golgi enzyme ST into the TRNs (S2 Table, S2b Fig).
Thus, UNC-16 facilitates “exclusion” of Golgi enzymes (see above) from SVP transport carriers, “inclusion” of specific SVPs into the same transport carrier and regulates the size of such compartments exiting the cell body.
To identify other genes in the UNC-16-mediated SVP trafficking pathway, we examined mutants in LRRK2/LRK-1, known to be present on the Golgi [30]. LRRK2/LRK-1 regulates the trafficking and distribution of synaptic vesicles in presynaptic boutons [37] and lrk-1 mutants show mis-trafficking of SNB-1 into dendrites of C. elegans chemosensory neurons, similar to phenotypes seen in unc-16 animals [30].
In lrk-1 animals, Golgi enzymes Man-II and ST occasionally exit the cell body (Fig 3b and S2a Fig). However, unlike in unc-16, Man-II does not mis-localize to the dendrite in lrk-1 animals (S1 Table, S2c Fig). In lrk-1 animals the incidence of compartments carrying both Golgi enzyme and SVP is closer to wild type, with only ~18% of the mobile compartments co-transporting Man-II and RAB-3 (Fig 3b, S2a Fig). The incidence of co-transport of RAB-3 and SNB-1 in the same compartment reduces in lrk-1 animals, with the severity of the phenotype similar to that observed in unc-16 (Fig 3c). Additionally, there is a ~5-fold increase in the frequency of longer SVP transport compartments seen exiting the cell body in both alleles of lrk-1 examined (Fig 3d and 3e, S3b Fig). As reported earlier, we also found that the dendritic marker ODR-10 shows a wild type-like polarized distribution in lrk-1 animals [32].
Taken together, our observations suggest that LRK-1 plays a critical role in regulating the composition as well as the size of the SVP transport carriers formed at the cell body, potentially at the Golgi. However, lrk-1 mutants lack the aberrant distribution of dendritic and Golgi proteins seen in unc-16.
Since lrk-1 has trafficking defects that are similar to unc-16, we tested whether UNC-16 and LRK-1 genetically function in the same pathway. We built double mutants and firstly assessed phenotypes present in unc-16 but absent in lrk-1. The Golgi enzyme Man-II is mis-trafficked into RAB-3-containing compartments similar to unc-16 single mutants (Fig 3b). Thus, the lrk-1; unc-16 double mutants are similar to unc-16 single mutants alone. Further, comparable to unc-16 and lrk-1 single mutants, the lrk-1; unc-16 double mutants show loss of polarized distribution of SNB-1 (Fig 3a), reduced co-transport of RAB-3 and SNB-1 in the TRN process (Fig 3c) and an increased frequency of long moving compartments carrying RAB-3 (Fig 3e).
We next determined whether UNC-16 acts upstream of LRK-1 in the trafficking of SVPs by overexpressing transgenic LRK-1 in unc-16 animals. Overexpression of LRK-1 greatly reduces the dendritic mis-localization of SNB-1 in unc-16 animals (Fig 4a, S2 Table). Over-expression of LRK-1 was also sufficient to exclude Golgi enzyme Man-II from RAB-3 containing SVP transport carriers (Fig 4b), to restore incidence of co-transport of RAB-3 and SNB-1 to wild type (Fig 4c) and to reduce the frequency of long transport carriers to frequencies seen in wild type (Fig 4d and 4e, S3c Fig). Interestingly, the overexpression of LRK-1 in unc-16 animals does not suppress the mis-trafficking of the dendritic marker ODR-10 into the axon (S1c Fig) nor does it completely suppress the exit of Man-II into the neuronal process (although these no longer travel in RAB-3 containing transport carriers). This suggests that transgenic LRK-1 only ameliorates the SVP-specific trafficking defects observed in unc-16.
As LRK-1 and UNC-16 are both present on the Golgi, we examined the localization of each protein and found that in unc-16 animals the punctate localization of LRK-1 on the Golgi is reduced in neuronal cell bodies (Fig 4f). In unc-16, a 10-fold increase was seen in the number of cell bodies showing a completely diffuse localization of LRK-1, compared to wild type animals (S4d Fig). In addition, LRK-1 is also mis-localized into the dendrites of sensory neurons in unc-16 animals (S4c Fig). On the other hand, UNC-16 localization remains unaffected in lrk-1 animals (S4e Fig). We further examined if both these proteins were part of the same complex. Immunoprecipitation experiments from C. elegans expressing LRK-1::FLAG and UNC-16::GFP show that LRK-1 and UNC-16 are present together in a complex in vivo (S4a and S4b Fig).
Given that overexpression of LRK-1 in unc-16 is sufficient to restore the processes of exclusion, inclusion and size regulation, our data suggest that unc-16 functions genetically upstream of lrk-1. This, along with the observation that lrk-1 mutants have modest exclusion defects but inclusion defects similar in severity to unc-16, suggests that exclusion may precede inclusion during SV biogenesis. The presence of UNC-16 likely facilitates the localization of LRK-1 on the Golgi and both together, possibly in a complex, regulate the trafficking of multiple SVPs.
Dendritic trafficking of proteins is known to require the AP-1 complex [10,23,38]. The mis-trafficking of SVPs into dendrites in lrk-1 mutants is dependent on UNC-101, the μ-chain of the Adaptor protein-1 (AP-1) complex in C. elegans [30]. Therefore, we tested if UNC-101 is involved in the UNC-16 and LRK-1-mediated regulation of the composition and size of the SVP transport carrier.
In unc-101 animals, like in wild type TRNs, Man-II and ST are restricted to the cell body and the incidence of co-transport of SVPs RAB-3 and SNB-1 is about 40% (Fig 5b and 5c, S2a Fig). Thus UNC-101, unlike UNC-16 and LRK-1 does not appear to play a role in regulating the composition of SVP transport carriers. However, nearly 40% of the RAB-3 containing vesicles in unc-101 animals were longer in the PLM neuron (Fig 5d and 5e, S3d Fig, S6 Movie), similar to unc-16 (Fig 2a, S4 and S5 Movies) and lrk-1 (Fig 3d and 3e) mutants. These longer transport carriers were seen in several other neuronal types (S7 Movie) as well. Thus, UNC-101 is involved in regulating the size of the SVP transport carrier leaving the cell body. In order to genetically position UNC-101 relative to UNC-16 and LRK-1, we built double mutants with both unc-16 and lrk-1. The mis-trafficking of the Golgi enzyme Man-II in the double mutants unc-101; unc-16 and unc-101; lrk-1 animals were similar to unc-16 or lrk-1 single mutants respectively (Fig 5b). Further, the rescue of size defects in unc-16 mutants by LRK-1 over-expression was found to require UNC-101 (Fig 5e, S3d Fig). Thus, UNC-101 acts downstream of both UNC-16 and LRK-1 in regulating size of the transport carrier but does not appear to influence sorting of SVPs into the carrier.
Previous studies have shown that the aberrant trafficking of SNB-1 into the dendrite of the sensory neuron in lrk-1 animals is dependent on UNC-101 [30]. Consistent with this, we observed that in unc-101, lrk-1 animals SNB-1 is excluded from the dendrite (S2 Table). In unc-101; unc-16 double mutants, SNB-1 continues to be mis-localized to the dendrite in ~ 20% of the animals, unlike in unc-16 animals where mis-localization is seen in 100% of the animals (Fig 5a, S2 Table). The MAN-II mis-localization into the dendrites in unc-16 and lrk-1 also requires UNC-101 (S2c Fig). Thus, even in absence of LRK-1 or UNC-16, UNC-101 continues to regulate the trafficking of proteins into dendrites. Our data, thus, suggests two roles for UNC-101 –(i) regulation of dendritic trafficking and (ii) regulating the size of axonally trafficked SVP transport carrier.
The AP-3 complex is known to function downstream to LRRK2 in the trafficking of lysosomal membrane proteins and to sort axonal proteins away from dendritic proteins [23,39]. Thus, we examined SVP trafficking in apb-3 mutants defective in the AP-3 β-subunit.
Unlike unc-101 animals, apb-3 mutants do not have longer SVP transport carriers, suggesting that AP-3 is not involved in size regulation (Fig 5e). Further, similar to wild type and unc-101 mutants, the Golgi protein Man-II largely stays restricted to the cell body in apb-3 mutants with little or no co-transport of Man-II with RAB-3 (Fig 5b). However, the apb-3 mutants show inclusion defects wherein incidence of co-transport of RAB-3 and SNB-1 is reduced to ~14% (Fig 5c), similar to unc-16 or lrk-1 single mutants. Both apb-3; unc-16 and apb-3, lrk-1 double mutants show reduced co-transport of RAB-3 and SNB-1, similar in severity to that seen in unc-16 or lrk-1 single mutants alone (Fig 5c). Additionally, over-expression of LRK-1 in unc-16 mutants is unable to rescue the inclusion defects in absence of APB-3 (Fig 5c). Thus, the AP-3 complex acts downstream of LRK-1 and may have roles in UNC-16 and LRK-1-mediated regulation of composition of the SVP transport carrier formed.
As the phenotype of long moving compartments observed in unc-16, lrk-1 and unc-101 is similar, we investigated whether there were changes in reported Golgi localization of UNC-101 in unc-16 and lrk-1 [23]. UNC-101 is not present as a defined puncta in unc-16 neuronal cell bodies while in lrk-1 animals, the puncta are reduced in both number and intensity as compared to wild type (Fig 6a, 6b and 6c). The overexpression of transgenic LRK-1 in unc-16 animals restores the localization of UNC-101 on the Golgi (Fig 6a, 6b and 6c). Since overexpression of LRK-1 rescues the size defects seen in unc-16 but not in an unc-101; unc-16 background, the LRK-1 mediated Golgi localization of UNC-101 is required for maintaining the size of the SVP transport carriers Figs 4e and 5e). Thus, the altered size of the synaptic vesicle precursors observed in unc-16 and lrk-1 depends on the presence of UNC-101 (or the AP-1 complex) on the Golgi.
SVP transport carriers in unc-16 are thought to recruit non-canonical motors [29,36], which may arise as a consequence of altered membrane composition. The molecular motor UNC-104 is known to transport synaptic vesicles protein transport carriers in C. elegans [24,39]. In concordance with this role, in unc-104 mutants, the SVPs RAB-3 or SNB-1 are found to be trapped in the cell body and absent from synapses (Fig 7bii) [15,40–42]. A previous study has shown that SVPs formed in unc-16 are transported independently of the UNC-104 motor in the DD and VD motor neurons [29]. This loss of motor-cargo specificity is also observed in TRN neurons such that SVP transport carriers travel up to the synapse in unc-16; unc-104 animals (Fig 7biii and 7biv). In lrk-1; unc-104 and in apb-3; unc-104 double mutants RAB-3 was observed in the proximal portion of the neuronal process but not at synapses, suggesting that the SVP carriers formed in lrk-1 and apb-3 are only partially dependent on UNC-104 (Fig 7bv, 7bvi and 7bx). The RAB-3 containing vesicles in unc-101; unc-104 were restricted to the cell body of the TRNs like in unc-104 alone (Fig 7bix) suggesting that unlike in unc-16 or lrk-1 mutants, the SVP transport carriers in unc-101 mutants were completely dependent on the UNC-104 motor.
We tested if the overexpression of LRK-1 was sufficient to restore the dependence of the SVP transport carrier on UNC-104 in unc-16 mutants and found that in these animals the SVP transport carriers are partially dependent on UNC-104 and are unable to reach the synapse (Fig 7bviii). Conversely, transgenic expression of UNC-16 does not significantly affect the altered dependence of the SVP transport carrier on UNC-104 in lrk-1 animals (Fig 7bvii). These observations demonstrate that the overexpression of LRK-1 in unc-16 sufficiently changes the composition of the SVP transport carrier such that it is now largely dependent on UNC-104. This further suggests that composition of the SVP transport carrier is critical to allow sufficient UNC-104 motor recruitment or to potentially exclude other motors along with excluding other proteins.
The mechanisms by which synaptic vesicle proteins (SVPs) are sorted into transport carriers and trafficked out of the cell body are not yet clearly understood. It is thought that like other membranous cargo such as secretory granules and dense core vesicles, SVPs are sorted at the Golgi and post-Golgi compartments and the transport carrier that is formed moves out of the cell body by recruiting specific motors, such as KIF1A/UNC-104 [1,4,5,43–45]. Using the model system C. elegans, we have uncovered a novel role for UNC-16/JIP3 in the trafficking and biogenesis of SVP transport carriers. Given that UNC-16/JIP3 localizes to the Golgi in C. elegans (Fig 1a) as well as in mammalian epithelial cells [26], it likely has conserved functions at the Golgi. We show that the SVP transport carrier biogenesis roles of UNC-16 occur via LRK-1. The sorting roles of UNC-16 and LRK-1 lead to the polarized distribution of SVPs as well as regulation of its composition and size. These two proteins regulate the composition of SVP transport carriers via the exclusion of Golgi resident enzymes and inclusion of relevant SVPs. These roles depend on the AP complexes where the AP-1 complex regulates the size of the carrier and the AP-3 complex regulates composition.
Our study shows that UNC-16 is essential to prevent Golgi resident enzymes such as Mannosidase-II and Sialyl transferase from entering the SVP transport carrier as inferred from the observed mis-trafficking of these enzymes along the neuronal process in unc-16 mutants (Fig 1b and 1d and S2a Fig). Previous studies have shown that retention of Golgi enzymes could occur through several mechanisms such as binding to scaffolding proteins on the Golgi or through the regulation of the lipid membrane composition to favour partitioning of different membrane proteins [46–48]. UNC-16 potentially acts as a scaffolding molecule to recruit effectors such as LRK-1 on the Golgi (Fig 4f), through which it may function to retain certain Golgi resident enzymes. Although lrk-1 mutants themselves do not show exclusion defects (Fig 3b), an excess of LRK-1 in an unc-16 background is sufficient to bypass the requirement of UNC-16 to exclude Golgi resident enzymes from the SVP transport carrier (Fig 4b). This suggests that LRK-1 may have redundant roles in exclusion of Golgi enzymes from SVP transport carriers. Alternatively, LRK-1 could act in an analogous manner to LRRK2, the mammalian homolog of LRK-1, that has been implicated previously in regulating the retromer complex, which sorts proteins from the endosome-lysosome degradation pathway retrogradely to the Golgi complex [49,50]. This likely occurs through the action of certain RABs such as RAB-7 and RAB-9 and through LRK-1’s interaction with VPS35 of the retromer complex [49–51]. In unc-16 mutants Golgi-resident proteins may not be excluded from other compartments as well, such as lysosomes. Our data suggests that UNC-16 mediated LRK-1 localization facilitates the exclusion of Golgi enzymes specifically from the SVP transport carrier as Man-II continues to be ectopically present along the axon potentially in other compartments in unc-16 mutants overexpressing LRK-1.
An important step during protein sorting is the clustering or segregation of proteins to be placed into the same compartment away from the donor compartment proteins. For example, the SVP Synaptobrevin-II and Synaptophysin interact with each other resulting in their co-trafficking [52,53]. Our findings suggest that both UNC-16 and LRK-1 are required for ensuring certain SVPs (such as SNB-1, SNT-1, and RAB-3) are sorted together and included more frequently in the same transport carrier (Figs 1c, 1e and 3c and S1e Fig). Moreover, the inclusion defects seen in unc-16 can be rescued by overexpression of LRK-1 and this rescue depends on presence of a functional AP-3 complex (Figs 4c and 5c). Along with the observation that lrk-1 genetically lies downstream to unc-16 and that exclusion defects are seen only in the unc-16 mutants, we hypothesize that exclusion of Golgi enzymes likely precedes inclusion during the early steps of SVP sorting. In unc-16 animals, the localization of LRK-1 itself at the Golgi is disrupted (Fig 4f, S4d Fig) suggesting that unc-16 may act as a potential hypomorph of LRK-1. This, along with our biochemical evidence that UNC-16 and LRK-1 are present in the same complex (S4a and S4b Fig), suggests that UNC-16 may scaffold LRK-1 at the Golgi in a physical complex that specifically regulates sorting of SVPs.
In addition to protein sorting, a crucial step in the formation of a transport carrier involves the regulation of its size. Previous studies have shown that proteins present on the surface of the TGN are crucial for the recruitment of the machinery involved in size regulation such as the Adaptor protein complexes [8,9,54–57]. We found that UNC-101, the μ-chain of the AP-1 complex in C. elegans, is indeed engaged in regulating the size of the SVP transport carrier formed (Fig 5d and 5e, S3d Fig). Importantly, the localization of UNC-101 on the Golgi is regulated by UNC-16 and LRK-1 (Fig 6). Considering that the Golgi enzyme Man-II in all of these mutants show the presence of 2–3 large puncta juxtaposed to the nucleus in the cell body, like in wild type, the Golgi is likely intact and the appearance of Golgi resident enzymes in the neuronal process can be accounted for due to errors in sorting/retrieval of these proteins. Furthermore, the size regulation by UNC-101 acts downstream to the early sorting steps regulated by UNC-16 and LRK-1 since overexpression of LRK-1 was able to rescue the size in unc-16 animals but not in unc-16; unc-101 mutants (Figs 4e and 5e, S3d Fig). Since exclusion or inclusion defects were absent from unc-101 animals, this also suggests that regulation of size and membrane composition could be independent processes and having an unusually large size does not necessarily incorporate other proteins (Golgi enzymes) typically excluded from these carriers. Earlier studies have also shown that the AP-1 complex is required for trafficking of proteins into dendrites [10,23,38]. Consistent with these observations, the dendritic mis-localization of SVPs in our mutants was largely dependent on UNC-101 (Fig 5a, S2 Table). We report an additional role where it regulates the size of the axonal cargo viz. SVP transport carriers. A recent study showed that the AP-3 complex is necessary at the Golgi for axonal localization of proteins [23]. AP-3 has previously been implicated in the sorting of proteins from early endosomes to lysosomal compartments [58–60]. Our study suggests that, unlike AP-1, the AP-3 complex does not regulate the size of SVP transport carriers (Fig 5e). On the other hand, compared to wild type animals, the apb-3 mutants show inclusion defects wherein the co-transport of SVPs, SNB-1 and RAB-3, is reduced (Fig 5c), which is similar to but more severe than that observed in unc-16 mutants. A recent study showed that AP-3 complex acts downstream of LRK-1 in the endo-lysosomal trafficking pathway [61]. Since overexpression of LRK-1 is able to suppress the defects in unc-16 but not in an apb-3; unc-16 double mutant (Fig 5c), AP-3 likely functions downstream to both UNC-16 and LRK-1 in regulating composition of some SVP transport carriers. Our data suggests that LRK-1 may be a general means to recruit different AP complexes to the appropriate membrane surface. We further hypothesize that UNC-101 at the Golgi is required for formation of an intermediate compartment whose size itself may be regulated by the AP-1 complex. At such an intermediate compartment the AP-3 complex may function to regulate the composition but not the size of vesicles arising from this precursor. Alternatively, UNC-101 may also have an additional role at the intermediate compartment to regulate size of vesicles arising from this compartment. LRK-1, itself may exert its retromer function at such an intermediate compartment for retrieval of Golgi enzymes.
Based on all our observations, the SVP transport carriers formed in unc-16 seem to be visibly aberrant in nature (Fig 8b). These observed defects likely arising due to an altered surface composition might also lead to the recruitment of multiple motors as suggested by Byrd et al., 2001 [29]. Consistent with this, we observed that the aberrant SVP transport carriers formed in unc-16 were no longer dependent exclusively on UNC-104 (Fig 7b). Further, it also appears to depend at least partially on other motors such as UNC-116/Kinesin-1 ([29]; S4f Fig). This suggests that the motor-cargo specificity is lost in unc-16 perhaps as a consequence of the altered surface composition that might now contain adaptors for multiple other motors. The transport carriers formed in lrk-1 and apb-3, have an altered composition, as suggested by defects in the ability to include different SVPs, and are only partially dependent on UNC-104 (Fig 7bvi and 7bx). This could again imply that due to an altered composition the carriers in lrk-1 and apb-3 mutants are not able to stably recruit sufficient numbers of the SV motor. Over expression of LRK-1 in unc-16 completely restores co-transport of RAB-3 and SNB-1, but is still insufficient to make the SVP transport carriers completely dependent on the motor (Fig 7bviii) suggesting the existence of additional factors necessary for forming an UNC-104 motor dependent transport carrier.
Previously, UNC-16 has been suggested to have a “clearance function” wherein it regulates the retrieval of cell soma organelles such as lysosomes and endosomes from the axon, thereby acting as an “organelle gatekeeper” [27,28]. Considering the multiplicity of mis-sorting defects we see in unc-16, we postulate that the unusual accumulation of different organelle proteins in the axon of these mutants could also be contributed by early defects in protein sorting at the Golgi rather than from a retrieval defect alone. Our study supports previous ideas that UNC-16 is involved in multiple trafficking pathways. UNC-16 may achieve these functions via different downstream effector molecules that it can potentially scaffold, regulating different subsets of trafficking pathways. Our study shows that UNC-16 acts via downstream molecules LRK-1, AP-1 and AP-3 to control biogenesis of the SVP transport carriers in the cell body (Fig 8a). Previous biochemical studies in mammalian cells have indicated that LRRK2 was co-purified with both JIP3 and components of clathrin [31,62], suggesting that similar relationships may be conserved in mammals. We uncover a likely hierarchical series of processes regulated by these proteins, which occurs early on at the Golgi and post-Golgi compartments, for the sorting and regulation of SVP trafficking (Fig 8a).
Both, UNC-16 and LRK-1 proteins have been implicated to have roles at the synapse. UNC-16 has been shown to regulate RAB-5-mediated membrane trafficking and contribute to SV maturation at the synapses [34]. Lee et. al., 2010 and Piccoli et. al., 2011 have proposed presynaptic roles for LRRK2 where it regulates synaptic morphology and SV recycling dynamics respectively [62,63]. Thus, both UNC-16 and LRK-1 could have trafficking roles at the synapse in addition to or as a consequence of altered membrane composition arising at the Golgi.
C. elegans strains were grown and maintained at 20°C on NGM plates seeded with the E. coli OP50 strain using standard methods [64]. L4 or 1-day adult animals were used for imaging in all cases. Strains used are listed in S4 Table. Some of the strains were provided by the CGC (https://cgc.umn.edu/acknowledging-the-cgc).
The unc-16 (tb109) allele was isolated from a behavioral suppressor screen carried out in unc-104 (e1265) worms. This was a non-clonal screen of approximately 60,000 haploid genomes. Animals were mutagenized using 50 mM EMS for 4 h. F1 and F2 progenies were screened for improved locomotion. Suppressors were further identified by improved localization of GFP::RAB-3 in PLM neuron (Kumar et al., 2010). The tb109 allele was separated from the background mutation and mapped to chromosome III and tested for non-complementation with unc-16 (e109). The tb109 phenotypes could be rescued by the expression of Punc-16::UNC-16::GFP. Sequencing revealed that allele tb109 contains a point mutation in exon 9 leading to a stop codon at amino acid 423 (Arginine to opal as stop codon).
Young adults of unc-16 (tb109) and N2 were fixed for electron microscopy by high-pressure freezing (HPF) technique [65]. Serial sections were cut and the dorsal and ventral nerve cord regions were imaged using Gatan side mount camera on Tecnai G2 12 BioTwin electron microscope (FEI Company). The cross-sectional width of all the vesicles present in each section was measured using ImageJ [66].
Immunostaining was performed as described previously [16]. For double labeling of the SV proteins, sample was first incubated with mouse anti-RAB-3 (1:2000), followed by incubation with rabbit anti-SNT-1 (1:500) antibody. Appropriate secondary antibodies (1:350) (Alexa 488, Alexa568; Molecular probes) were added and incubated for two days. Images were captured using a Zeiss Axiovert inverted microscope. Images were processed using ImageJ [66].
Samples were prepared by mechanical homogenization using homogenization buffer (15mM HEPES-NaOH pH7.4, 10mM KCl, 1.5mM MgCl2, 0.1mM EDTA, 0.5mM EGTA, 0.05mM sucrose and protease inhibitors (Roche)) followed by mild sonication at 4°C. Sample was then incubated with mouse monoclonal anti–Flag (1:50) (Biovision) or mouse monoclonal anti-GFP sera (1:10) (Genei, Merck) at 4°C for 4 h. The antigen-antibody complex was incubated with Protein-A agarose beads (Genei, Merck) for another 4 h. Antibodies used for Western blots to probe for UNC-16::GFP and LRK-1::Flag were rabbit anti-GFP (1:500–1000, Santacruz, Abcam) and mouse anti-flag (1:1000–2000, Sigma) respectively.
In all graphs, data are presented as mean values ± SEM. Statistical analysis was performed using GraphPad Prism 6 (GraphPad Software). Wherever possible, an independent t-test was used, or for multiple comparisons, one-way ANOVA followed by Dunett’s multiple comparison tests was done. For grouped datasets, two-way ANOVA, followed by Tukey’s multiple comparison tests was used. Differences were considered significant when P < 0.05 (*, P < 0.05; **, P < 0.01; ***, P < 0.001).
|
10.1371/journal.pgen.1002260 | Large-Scale Gene-Centric Analysis Identifies Novel Variants for Coronary Artery Disease | Coronary artery disease (CAD) has a significant genetic contribution that is incompletely characterized. To complement genome-wide association (GWA) studies, we conducted a large and systematic candidate gene study of CAD susceptibility, including analysis of many uncommon and functional variants. We examined 49,094 genetic variants in ∼2,100 genes of cardiovascular relevance, using a customised gene array in 15,596 CAD cases and 34,992 controls (11,202 cases and 30,733 controls of European descent; 4,394 cases and 4,259 controls of South Asian origin). We attempted to replicate putative novel associations in an additional 17,121 CAD cases and 40,473 controls. Potential mechanisms through which the novel variants could affect CAD risk were explored through association tests with vascular risk factors and gene expression. We confirmed associations of several previously known CAD susceptibility loci (eg, 9p21.3:p<10−33; LPA:p<10−19; 1p13.3:p<10−17) as well as three recently discovered loci (COL4A1/COL4A2, ZC3HC1, CYP17A1:p<5×10−7). However, we found essentially null results for most previously suggested CAD candidate genes. In our replication study of 24 promising common variants, we identified novel associations of variants in or near LIPA, IL5, TRIB1, and ABCG5/ABCG8, with per-allele odds ratios for CAD risk with each of the novel variants ranging from 1.06–1.09. Associations with variants at LIPA, TRIB1, and ABCG5/ABCG8 were supported by gene expression data or effects on lipid levels. Apart from the previously reported variants in LPA, none of the other ∼4,500 low frequency and functional variants showed a strong effect. Associations in South Asians did not differ appreciably from those in Europeans, except for 9p21.3 (per-allele odds ratio: 1.14 versus 1.27 respectively; P for heterogeneity = 0.003). This large-scale gene-centric analysis has identified several novel genes for CAD that relate to diverse biochemical and cellular functions and clarified the literature with regard to many previously suggested genes.
| Coronary artery disease (CAD) has a strong genetic basis that remains poorly characterised. Using a custom-designed array, we tested the association with CAD of almost 50,000 common and low frequency variants in ∼2,000 genes of known or suspected cardiovascular relevance. We genotyped the array in 15,596 CAD cases and 34,992 controls (11,202 cases and 30,733 controls of European descent; 4,394 cases and 4,259 controls of South Asian origin) and attempted to replicate putative novel associations in an additional 17,121 CAD cases and 40,473 controls. We report the novel association of variants in or near four genes with CAD and in additional studies identify potential mechanisms by which some of these novel variants affect CAD risk. Interestingly, we found that these variants, as well as the majority of previously reported CAD variants, have similar associations in Europeans and South Asians. Contrary to prior expectations, many previously suggested candidate genes did not show evidence of any effect on CAD risk, and neither did we identify any novel low frequency alleles with strong effects amongst the genes tested. Discovery of novel genes associated with heart disease may help to further understand the aetiology of cardiovascular disease and identify new targets for therapeutic interventions.
| Coronary artery disease (CAD) has a substantial genetic component which is incompletely characterised. Genomewide association (GWA) studies have recently identified several novel susceptibility loci for CAD [1]–[9]. Because GWA studies involve assumption-free surveys of common genetic variation across the genome, they can identify genetic regions responsible for previously unsuspected or unknown disease mechanisms. However, despite the success of the GWA approach, it has potential limitations. Because CAD loci identified through GWA studies have predominantly been found in regions of uncertain biological relevance, further work is required to determine their precise contribution to disease aetiology. Furthermore, in contrast with their high coverage of common genetic variation, GWA studies tend to provide limited coverage of genes with well-characterised biological relevance (“candidate genes”) [2], particularly in relation to lower frequency genetic variants (such as those with minor allele frequencies of 1–5%). Such variants are also often difficult to impute from GWA data. Although candidate gene studies should provide more comprehensive coverage of lower frequency and functional variants than GWA studies, most have been inadequately powered.
To complement GWA studies, we undertook a large-scale gene-centric analysis of CAD using a customised gene array enriched with common and low frequency variants in ∼2,100 candidate cardiovascular genes reflecting a wide variety of biological pathways [10]. The array's potential to identify disease-associated lower frequency variants has been demonstrated by previous identification of strong independent associations with 2 variants in the LPA gene - rs3798220 (minor allele frequency 2%), and rs10455872 (7%) - and CAD risk [11]. We have now investigated this gene array in a further 13 studies comprising a total of 15,596 CAD cases and 34,992 controls. To enable interethnic comparisons, participants included 4,394 cases and 4,259 controls of South Asian descent, an ethnic group with high susceptibility to CAD. For further evaluation of putative novel associations, we attempted to replicate them in an additional 17,121 cases and 40,473 controls.
The experimental strategy used is shown in Figure 1. In the discovery phase we genotyped participants from 12 association studies of CAD/myocardial infarction (MI), including a total of 11,202 cases and 30,733 controls of European descent (10 studies), plus 4,394 South Asian cases and 4,259 South Asian controls (2 studies) (Table 1, Table S1, with further details of the studies given in Text S1).
36,799 SNPs passed QC and frequency checks and were included in the meta-analysis (reasons for exclusion of variants in each study are given in Table S2). The distribution of association P values in the discovery stage analyses are shown in Figure 2. We found significant associations with CAD for several previous GWA-identified loci contained on the array including 9p21.3 (rs1333042, combined European and South Asian P = 1.1×10−37) and 1p13.3 (rs646776, 3.1×10−17; Table S3). We also confirmed associations of other genes with strong prior evidence including the first association of a variant at the apolipoprotein E locus at genomewide significance (APOE/TOMM40, rs2075650, P = 3.2×10−8), as well as associations at apolipoprotein (a) (LPA, rs10455872, P = 1.2×10−20), and low density lipoprotein receptor (LDLR, rs6511720, P = 1.1×10−8; Table S3). However, we found no persuasive evidence of association of several prominently-studied genes and variants for which the previous epidemiological evidence has been inconclusive, even though the majority of these loci were well-tagged (Table S4) and the current study was well-powered to detect associations of modest effect (Figure S1). Notable variants that did not show significant association included the angiotensin converting enzyme (ACE) insertion/deletion polymorphism, the cholesteryl-ester transfer protein (CETP) Taq1B polymorphism and the paraoxonase 1 (PON1) Q192R polymorphism (Table S4). Perhaps contrary to expectation, apart from the LPA variant rs3798220, we did not observe any other strong association (odds ratio >1.5) among the ∼4,500 low frequency (1–5%) variants and/or variants with suspected or known functional impact on protein structure/function or gene expression specifically selected for the inclusion on the array (Table S3).
Based on simulations conducted prior to the analysis (Figure S2), loci were eligible for replication if unadjusted P-values for CAD were <1×10−4 in either the primary (each ethnic group analysed separately) or secondary (combined) analyses and the loci had not been previously established with CAD. This identified 27 loci in total: 15 in the European only analysis, 3 in the South Asian only analysis, and 9 in the combined analysis (Table S5). A recent GWA meta-analysis from the CARDIoGRAM Consortium with some overlapping cohorts to those in our study, reports discovery of three of these loci [12]: COL4A1/COL4A2, ZC3HC1, CYP17A1. The P values observed for the lead variants at these loci in the current study were: COL4A1/COL4A2: rs4773144, P = 3.5×10−8; ZC3HC1: rs11556924, P = 3.1×10−7; CYP17A1: rs3824755, P = 1.2×10−7, providing further strong evidence for the association of these loci with CAD. Hence, only the lead SNPs at the 24 remaining loci were taken forward for replication. This was done in silico in 17,121 CAD cases and 40,473 controls, all of whom were of white European ancestry and derived from non-overlapping cohorts from CARDIoGRAM and EPIC-NL (Text S1, Table S6). The power of our replication sample to confirm significant associations is shown in Figure S1. Of the 24 variants taken forward, four were independently replicated (1-tailed Bonferroni-corrected P<0.05 is P<1.9×10−3; Figure 3, Table S5), comprising variants in or adjacent to: LIPA, IL5, TRIB1 and ABCG5/ABCG8 (Figure 4). For the variant at the LIPA locus, the combined P-value was 4.3×10−9, exceeding conventional thresholds for GWA studies. For each of the IL5, TRIB1 and ABCG5/ABCG8 variants, the P-value was <3×10−6, exceeding array-wide levels of significance (Figure 3). CAD associations in the individual component studies are shown in Figure S3. The CAD associations for these loci did not vary materially by age, sex or when restricted to the MI subphenotype (Figure S4).
To investigate whether the 4 newly identified loci associate with cardiovascular risk traits, we interrogated available data from previous GWA meta-analyses of diabetes mellitus (n = 10,128 individuals) [13], systolic blood pressure (n = 25,870) [14], and low-density (LDL) and high-density (HDL) lipoprotein-cholesterol and triglycerides (n = 99,900) [15]. This showed that the risk allele at the TRIB1 locus was associated with higher triglyceride (P = 3.2×10−53), higher LDL-C (P = 6.7×10−29) and lower HDL-C (P = 9.9×10−17) and that the ABCG5/ABCG8 risk allele was associated with higher LDL-C (P = 1.7×10−47; Figure 5). We also examined the association of the novel risk variants with gene expression in full transcriptomic profiles of circulating monocytes derived from 363 patients with premature myocardial infarction and 395 healthy blood donors from the Cardiogenics study (Text S1). We found a highly significant association (P = 1.0×10−124) of the risk allele at the LIPA locus with LIPA mRNA levels in these cells explaining ∼50% of the variance in the expression of the gene (Figure 6).There were no other highly significant associations between CAD risk alleles and gene expression at the novel loci (Table S7a and S7b).
We explored whether associations of loci with CAD differed between individuals of white European ancestry and South Asian ancestry. For most loci, frequency of risk alleles and pattern of risk associations did not differ qualitatively by ethnicity, although the evidence of association was often weaker in South Asians, perhaps due to lower power (Figure 3, Tables S3 and S5). For the 9p21.3 locus, despite similar risk allele frequencies (Table S3), odds ratios were higher in Europeans than South Asians (rs1333042: 1.27 v 1.14; P = 0.003 for difference), though common haplotype frequencies did not vary by ethnicity (Table S8). The three variants at the TUB, LCT and MICB loci selected for replication on the basis of South Asian-specific results did not show evidence of association in Europeans (Table S5).
Our in-depth study of ∼2,100 candidate genes has yielded several novel and potentially important findings, adding to the emerging knowledge on the genetic determination of CAD. First, we have identified several novel genes for CAD. These genes relate to diverse biochemical and cellular functions: LIPA for the locus on 10q23.3; IL5 (5q31.1); ABCG5/ABCG8 (2p21); TRIB1 (8q24.13); COL4A1/COL4A2 (13q34); Z3HC1 (7q32.3); and CYP17A1 (10q24.3). We have furnished evidence directly implicating the candidacy of these genes, either because the locations of the signals discovered are within a narrow window of linkage disequilibrium or because there is evidence of a mechanistic effect, or both. Second, we have provided large-scale refutation of the relevance of many prominent candidate gene hypotheses in CAD, thereby clarifying the literature. Third, contrary to expectation, we did not observe highly significant novel associations between low frequency variants and CAD risk, despite study of >4,500 such variants. Fourth, we have confirmed the relevance of several previously established CAD genes to both Europeans and South Asians, without finding qualitative differences in results by ethnicity.
LIPA (lipase A) encodes a lysosomal acid lipase involved in the breakdown of cholesteryl esters and triglycerides. Mutations in LIPA cause Wolman's disease [16], a rare disorder characterized by accumulation of these lipids in multiple organs. However, despite evidence that the risk allele was associated with higher LIPA gene expression (suggesting that both under- and over-activity of LIPA increase CAD risk), it was not significantly associated with altered lipid levels. This finding suggests that the impact on CAD risk is either through an alternative pathway, or that the mechanism is more complex than reflected through conventionally measured plasma lipid levels. Two recent studies have also found associations of variants in the LIPA gene with CAD using a GWA approach, strengthening the evidence for this association [17], [18].
Our identification of the association of variants near interleukin 5 (IL5), an interleukin produced by T helper-2 cells, is interesting given the evidence that both acute and chronic inflammation may play important roles in the development and progression of CAD [19]. Most previous human association studies of inflammatory genes and CAD have focused on other cytokines and acute-phase reactants. Nevertheless, some experimental data predict that IL-5 has an atheroprotective effect and this has been supported by association between higher circulating IL-5 levels and lower carotid intimal-medial thickness [20]–[22]. Our findings now highlight the potential importance of IL-5 in CAD, especially as the IL-5 receptor is already a viable therapeutic target in allergic diseases, although we can not rule out the possibility that another gene at this locus may be mediating the association with CAD risk.
The ATP-binding cassette sub-family G proteins ABCG5 and ABCG8 are hemi-transporters that limit intestinal absorption and promote biliary excretion of sterols. Mutations in either gene are associated with sitosterolaemia, accumulation of dietary cholesterol and premature atherosclerosis [23]. Recently, common variants in ABCG8 have also been shown to be associated with circulating LDL-C and altered serum phytosterol levels with concordant changes in risk of CAD [15], [24]. Our findings confirm that this locus affects CAD risk either directly through its effect on plasma phytosterol levels or through primary/secondary changes in LDL-cholesterol.
The association signal on 8q24.13 maps near the TRIB1 gene which encodes the Tribbles homolog 1 protein. Tribbles are a family of phosphoproteins implicated in regulation of cell function, although their precise roles are unclear [25]. However, SNPs in or near TRIB1 - including the lead SNP in our study (rs17321515) - have recently been shown to have highly significant associations with levels of several major lipids [15], providing a possible mechanism for their association with CAD. Our findings confirm the previous suggestion that this variant is also associated with CAD risk [15], [26]. Hepatic over-expression of TRIB1 in mice has been shown to lower circulating triglycerides; conversely, targeted deletion of the TRIB1 gene in mice led to higher circulating triglycerides [27]. The location of the CAD-associated variant downstream of TRIB1 suggests that its effect may be mediated by regulation of TRIB1 expression leading to adverse lipid profiles, although we did not find an eQTL at this locus in monocytes.
Our study brings to 33 the number of confirmed loci with common variants affecting risk of CAD (Figure 7). We estimate that in aggregate these variants explain about 9% of the heritability of CAD which is consistent with the recent analysis by CARDIoGRAM [12]. Interestingly, the odds ratios that we observed for the novel loci were generally lower than those of previously identified loci. This suggests that most of the common variants with moderate effects have been identified and that increasingly larger sample sizes will be required to detect further common variants that affect risk of CAD. However, the modest odds ratios associated with such variants do not necessarily imply that they are not of potential clinical or therapeutic relevance. For example, there are only modest effects of common variants in the LDLR gene on CAD risk (Figure 7); yet this pathway has become a major target for the prevention and treatment of CAD with the development of statins.
Despite the success of the GWA approach in identifying several common variants that affect risk of CAD, such loci explain only a small proportion of the heritability of CAD [5]. It has been hypothesized that some of the unexplained heritability resides in lower frequency (1–5%) variants which are not adequately represented on current genomewide arrays and/or are difficult to impute from GWA data. Because the gene array used in the current study included ∼4500 lower frequency variants as well as known functional variants for the majority of the genes on the array, we were able to examine this issue for CAD, at least in relation to candidate cardiovascular genes. Although we confirmed the previously reported associations of lower frequency variants in LPA and PCSK9 with CAD risk, we did not detect any other strongly associated variants in the 1–5% range or an enrichment of low frequency variants amongst SNPs that showed nominal association with CAD. However, it is important to note that rare variants in the genome (minor allele frequency <1%) were not addressed in this study.
CAD is more common in South Asians and tends to occur at an earlier age than in Europeans, perhaps partly due to genetic factors [28]. Our study provides the first systematic exploration of this issue. We observed a weaker effect size for the 9p21.3 locus in South Asians compared with Europeans, although this did not appear to be related to any obvious differences in haplotype structure at the locus, confirming recent findings in Pakistanis [29]. This difference in effect size between ethnic groups will require further evaluation and replication as other differences between the European and South Asian studies (eg, different sex distributions) could explain this finding. Most of the other disease-associated variants we found had slightly weaker effects in South Asians, although, because power to detect heterogeneity of effect between the ethnicities was low and there were only 2 South Asian studies, this finding will require further evaluation. We observed variants at 3 loci (TUB, LCT and MICB, Table S5) which showed modest (P<10−4) associations in South Asians but were convincingly null in Europeans and will therefore require replication in additional South Asian samples. Overall, we did not find clear evidence of major variation in genetic risk factors for CAD between Europeans and South Asians.
In summary, using a large-scale gene-centric approach we have identified novel associations of several genes for CAD that relate to diverse biochemical and cellular functions, including inflammation and novel lipid pathways, as well as genes of less certain function. Together, these findings indicate that previously unsuspected biological mechanisms operate in CAD, raising prospects for novel approaches to intervention.
Characteristics of the discovery phase studies are summarised in Table 1, Table S1 and the replication studies in Table S6. Further details of all the studies are given in Text S1. All individuals provided informed consent and all studies were approved by local ethics committees.
Using the HumanCVD BeadChip array (Illumina), which is also known as the “ITMAT-Broad-CARe” (IBC) 50K array, we genotyped 49,094 single nucleotide polymorphisms (SNPs) in ∼2,100 candidate genes identified in previous studies of cardiovascular disease, pathway-based approaches (including genes related to metabolism, lipids, thrombosis, circulation and inflammation), early access to GWA datasets for CAD, type 2 diabetes, lipids and hypertension, as well as human and mouse gene expression data [10]. Variants in genes suspected to be associated with sleep, lung and blood disease phenotypes were also included, along with SNPs that were related in GWA datasets to rheumatoid arthritis, Crohn's disease and type 1 diabetes. Human and mouse gene expression data was also used to select variants. Genes were then prioritised by investigators, with ‘high priority genes’ densely tagged (all SNPs with MAF>2% tagged at r2>0.8), ‘intermediate priority genes’ moderately covered (all SNPs with MAF>5% tagged at r2>0.5), and ‘low priority genes’ tagged using only non-synonymous SNPs and known functional variants with MAF>1%.
A “cosmopolitan tagging” approach was used to select SNPs providing high coverage of selected genes in 4 HapMap populations (CEPH Caucasians, Han Chinese, Japanese, Yorubans). For all genes, non-synonymous SNPs and known functional variants were prioritised on the array. Genotypes were called using standard algorithms (eg, GenCall Software and Illuminus) and standard quality control methods were applied to filter out poorly performing or rare (<1% minor allele frequency) SNPs (Text S1). After exclusion of low frequency variants (average 8,354 in each study), non-autosomal variants (average 1,224) and variants that failed quality control (average 842 – predominantly due to high missingness or failure of HWE), the number of SNPs taken forward for analysis in each study ranged from 30,550–39,027 (Table S2).
In each study, unadjusted logistic regression tests using a case-control design assuming an additive genetic model were conducted, with most studies using PLINK [30]. All studies made attempts to reduce over-dispersion. The genomic inflation factor for each study after adjustment was <1.10 with one exception (Table S2). The primary analysis was a fixed-effect inverse-variance-weighted meta-analysis performed separately for each ethnic group using STATA v11. A chi-squared test for between-ethnicity heterogeneity was performed. A secondary analysis combined European and South Asian studies to identify additional variants common to both ethnicities (Text S1).
Based on a simulation study conducted prior to the analysis (Figure S2), variants were selected for the replication stage if they had an unadjusted P<1×10−4 in either the primary analysis or the combined ethnicity analysis. Only the most significant (“lead”) SNP from each locus was taken forward for replication. SNPs at known coronary disease risk loci (eg, 9p21.3, LPA, APOE) were excluded from the replication stage, leaving 27 SNPs to take forward. In silico replication was conducted using non-overlapping participants from the CARDIoGRAM GWA meta-analysis [12] of CAD plus an additional study, EPIC-NL [31] (details in Table S6). In total, the replication stage comprised up to 17,121 coronary disease cases and 40,473 controls. The threshold for independent replication was a 1-tailed Bonferroni-corrected P<0.05 (P<1.9×10−3) from a Cochran-Armitage trend test. P values from the replication and discovery stages were combined using Fisher's method with a chip-wide value of P<3×10−6 considered to be statistically significant based on the simulation study (Figure S2). Adjusted P values accounting for both over-dispersion and heterogeneity in the discovery stage studies were also estimated through correction for study- and meta-analysis-specific inflation factors.
To check for consistency of effect of variants that replicated, subgroup analyses were performed in the discovery stage studies for MI cases only, CAD cases aged less than 50, males only and females only. Replicating SNPs were tested for association with known cardiovascular risk factors such as blood pressure, lipids levels and type 2 diabetes mellitus using existing large-scale GWA meta-analyses data of these traits [13]–[15]. We also assessed the association of these variants with gene expression in circulating monocytes taken from 363 patients with premature myocardial infarction and 395 healthy blood donors (Text S1). To put novel findings from this study in the context of existing knowledge, we summarised associations of common variants established in CAD (P<5×10−8) using available information from the NHGRI's GWA studies catalogue [32].
|
10.1371/journal.ppat.1006136 | Static and Evolving Norovirus Genotypes: Implications for Epidemiology and Immunity | Noroviruses are major pathogens associated with acute gastroenteritis worldwide. Their RNA genomes are diverse, with two major genogroups (GI and GII) comprised of at least 28 genotypes associated with human disease. To elucidate mechanisms underlying norovirus diversity and evolution, we used a large-scale genomics approach to analyze human norovirus sequences. Comparison of over 2000 nearly full-length ORF2 sequences representing most of the known GI and GII genotypes infecting humans showed a limited number (≤5) of distinct intra-genotypic variants within each genotype, with the exception of GII.4. The non-GII.4 genotypes were comprised of one or more intra-genotypic variants, with each variant containing strains that differed by only a few residues over several decades (remaining “static”) and that have co-circulated with no clear epidemiologic pattern. In contrast, the GII.4 genotype presented the largest number of variants (>10) that have evolved over time with a clear pattern of periodic variant replacement. To expand our understanding of these two patterns of diversification (“static” versus “evolving”), we analyzed using NGS the nearly full-length norovirus genome in healthy individuals infected with GII.4, GII.6 or GII.17 viruses in different outbreak settings. The GII.4 viruses accumulated mutations rapidly within and between hosts, while the GII.6 and GII.17 viruses remained relatively stable, consistent with their diversification patterns. Further analysis of genetic relationships and natural history patterns identified groupings of certain genotypes into larger related clusters designated here as “immunotypes”. We propose that “immunotypes” and their evolutionary patterns influence the prevalence of a particular norovirus genotype in the human population.
| Efforts are underway to develop vaccines against norovirus, a leading cause of acute gastroenteritis. The purpose of our study was to understand how norovirus strains within different genotypes evolve and adapt as they are transmitted in the human population. Using large-scale genomics and computational tools developed in our laboratory, we identified two strikingly different evolutionary patterns among norovirus genotypes: “static” and “evolving.” We mined large datasets from infection and outbreak studies in context of these evolutionary patterns and propose a new model for antigenic clustering of norovirus genotypes that could simplify vaccine design.
| RNA viruses evolve quickly, with mutation rates that vary between 10−3–10−4 nucleotide (nt) substitutions/year; up to 1000 times higher when compared with most DNA viruses [1]. This high mutation rate is attributed largely to the inability of their RNA polymerases to correct errors introduced during replication. In addition to nt substitutions, RNA viruses generate diversity by recombination or rearrangements of their genome [2–4]. This diversity is considered required for the virus to (i) escape the immune surveillance of the host [5–7], (ii) reach different organs of the host and/or change tropism, and (iii) induce the pathological effects that could lead to efficient transmission to a new host [8]. Although extreme diversity might be advantageous for a virus to exploit different niches, viruses can also retain their phenotypic characteristics for decades (e.g., dengue virus and respiratory syncytial virus) [5, 9].
Noroviruses are a major cause of acute gastroenteritis worldwide, mainly associated with outbreaks occurring in closed settings, such as hospitals, nursing homes, schools, cruise ships, and military facilities. In countries where rotavirus vaccination has been successfully introduced, norovirus has become the major cause of gastroenteritis in children [10–13]. It is estimated that noroviruses cause between 70,000 to 200,000 deaths per year worldwide, with the majority in children from developing countries [14, 15]. Although symptoms characteristically resolve within 72 hours in healthy immunocompetent individuals, norovirus genomic RNA can be detected in stool for up to 2 months [16–19], and clearance of virus seems to be associated with a specific immune response in the mucosa [18]. In immunocompromised patients, symptoms and virus can persist over years, complicating the management of their underlying disease [20].
Noroviruses are considered fast-evolving viruses [21–24], and present an extensive diversity that is driven by acquisition of point mutations and recombination. The genome consists of a single-stranded positive-sense RNA molecule of ~7.5kb that is organized into three open reading frames (ORFs). ORF1 encodes a polyprotein that is co-translationally cleaved into six proteins required for replication, while ORF2 encodes the major capsid protein (VP1), and ORF3, a minor capsid protein (VP2) [25]. Expression of recombinant VP1 yields virus-like particles (VLPs) that mimic the native virion, which have been important research tools in the absence of tractable cell culture systems and animal models [26–30]. Based on sequence differences of the VP1 protein, noroviruses have been classified into seven genogroups (GI-GVII) and over 30 genotypes [25, 31]. Despite this extensive diversity, a single genotype (GII.4) has been shown to be the most prevalent in humans worldwide [31–33]. Since the mid-1990s, six global epidemics have been documented and each has been associated with the emergence of a new GII.4 variant. The first was characterized by an increased number of norovirus outbreaks worldwide, and associated with the US95_1996 virus. The second epidemic started in 2002 and coincided with the replacement of the US95_1996 virus by the Farmington_Hills_2002 virus. The third epidemic was caused by the Hunter_2004 virus, which was rapidly replaced by two new pandemic strains, namely Den_Haag_2006b and Yerseke_2006a. During 2009, a new strain emerged (New_Orleans_2009) that co-circulated with the 2006 variants for almost three years until the current predominant variant emerged in 2012 (Sydney_2012) [25, 31, 33]. Interestingly, this epidemiological pattern has not been reported for any other norovirus genotype, until recently. A novel GII.17 variant has emerged, potentially displacing an older GII.17, causing large outbreaks in different countries from Asia [22, 34–36]. Although GII.4 is overall the most prevalent genotype in the human population, multiple norovirus genotypes co-circulate in children with low to high incidence. Genotypes GII.3, GII.6 and GII.2 (in addition to GII.4) have consistently been linked to infection in children under 5 years of age [17, 21, 37, 38].
Initial challenge studies in human volunteers suggested a lack of protective responses between strains from the two major genogroups (GI and GII), as cross-challenge between Norwalk virus (the GI.1 prototype strain) and Hawaii virus (the GII.1 prototype strain) did not induce protection. In addition, duration of immunity might be short (less than 6 months), as individuals re-challenged with the same virus became ill during the second exposure [39]. It has been noted that the high titer of challenge virus administered in the early volunteer studies might not reflect that during natural exposure [40, 41] and recent studies have focused on the natural history of noroviruses. Based on epidemiological data, Simmons et al. modeled that norovirus genotype-specific immunity could last up to 9 years [42], which would enhance the duration of vaccine-induced immunity. The diversity of genotypes has also been addressed. Children can be re-infected multiple times during the first 5 years of life [17, 18, 38], with the majority of re-infections occurring with different norovirus genotypes. These data suggest that genotypes may represent distinct serotypes, which would complicate vaccine design.
In this study, we integrated large-scale genomics analysis with natural history data to investigate mechanisms involved in the diversification and evolution of norovirus genotypes and their variants. Most norovirus intra-genotypic variants displayed a striking genetic stability over long periods of time, with GII.4 as the notable exception. We detected patterns of re-infection and susceptibility consistent with genetic and antigenic clustering of certain genotypes and propose that these relationships may be relevant in the design of norovirus vaccines.
To investigate the diversity and evolutionary differences of the distinct norovirus genotypes, more than 2000 sequences of the gene (ORF2) encoding the VP1 were retrieved from GenBank, with 101 and 1909 genes from GI strains and GII strains, respectively (Table 1). Individual sequences were vetted with an online norovirus typing tool that follows a widely-used universal classification and nomenclature system for norovirus genotypes [25], and an effort was made to verify the date of occurrence with the supporting documentation. Genotypes with 10 or more complete (or nearly complete) ORF2 sequences were selected for further phylogenetic analysis at both the nt and amino acid (aa) level, and included 16 out of the 31 current GI and GII genotypes (Table 1). We defined an intra-genotypic variant as a group of strains (≥ 2) that clustered together in the phylogenetic tree and that showed <5% difference in their nt or aa sequences, but ≥5% difference compared to other strains. Most genotypes segregated into 1 to 5 phylogenetic variants when nt sequences were analyzed (Table 1), with the exception of genotype GII.4 that displayed at least 10 different variants (Table 1 and Fig 1A). The number of variants was lower in seven genotypes (GI.1, GI.4, GI.6, GII.2, GII.12, GII.13 and GII.14) when aa sequences were used for tree reconstruction and distance analyses (Table 1). To link the different intra-genotypic variants to phenotypic characteristics in the VP1, we focused on the analysis of aa sequences. Interestingly, non-GII.4 genotypes presented variants with strains that have been detected many years apart (mean: 24.9 years, standard deviation: 12.97 years) while having only a few differences in their aa sequence (Table 1). An example is the GII.6 genotype, with each of the three variants (A-C) containing strains that differed by only a few aa residues (≤1.2%) but that were detected up to 41 years apart (Fig 2A). In contrast, the GII.4 genotype was comprised of variants that were present in the human population from 3 to 8 years (mean: 5.3 years, Fig 1A).
We developed an algorithm to illustrate visually the relationship between amino acid diversity and time among strains from a given genotype. The algorithm generated a heat map in which each square represents the number of strains with such aa difference in their VP1 and plotted against the timespan of detection. Analysis of 16 genotypes with sufficient data in GenBank revealed two distinct patterns of variant evolution; one in which the number of aa differences accumulated continually over time (Fig 1B), and one in which the number of aa differences remained relatively constant over time, regardless of the timespan between strains (Fig 2B). The first pattern (Fig 1B) was related to a constantly changing “evolving genotype” represented by the GII.4, while the second pattern (Fig 2B) was a highly conserved or “static genotype” represented by the 15 other genotypes with sufficient data for analysis (Table 1). The static genotypes resolved readily into distinct intra-genotypic variants, with one exception. The diversity plot of the GII.12 genotype strains displayed a subtle accumulation of differences over a time period of approximately 20 years; however, those differences were not associated with different intra-genotypic variants (S1 Fig). A larger number of sequences over a longer period will be helpful for defining variant diversity in GII.12 and other static genotypes.
During 2014–2015, a sharp increase in the number of gastroenteritis outbreaks was reported in Asia [22, 34, 43] that were associated with the emergence of a new variant of the genotype GII.17, i.e. variant C or Kawasaki_2014 [35, 43]. Our phylogenetic analysis of this genotype revealed four distinct variants, with one of the variants having strains that spanned over 37 years with a low level of sequence diversity consistent with a static pattern (S2 Fig). In contrast, the emerging strains showed multiple substitutions at the nt and aa sequence level [22, 35][44], which led to diversification into two separate phylogenetic variants (C and D). These differences appeared to accumulate over time, with predominant strains that circulated during the 2014–2015 season (variant D) differing by 5.4±1.1% from those in the 2013–2014 season (variant C) in their aa sequence.
To gain insight into the evolutionary differences noted at the aa level, the nt rate of evolution and nonsynonymous substitutions (dN)/synonymous substitutions (dS) ratio were calculated for each of the genotypes included in this analysis. The nt rates of evolution were similar among all the norovirus genotypes (range: 5.40x10-3–2.23x10-4 nt substitutions/site/year). However, differences were seen in the dN/dS ratios, with genotypes GII.4 and GII.17 presenting slightly higher values than any other norovirus genotype (Table 1). Note that in five different genotypes (GI.3, GI.4, GII.4, GII.13, and GII.17), the dN/dS ratio in the P2 encoding region was at least two times higher than the complete ORF2 (encoding VP1) dN/dS ratio. The dN/dS ratio has been used to inform the evolutionary pressures on a gene: a dN/dS > 1 (higher number of non-synonymous mutations) indicates positive evolution as the phenotype is changing due to pressures of the environment (e.g. immune responses), while a dN/dS < 1 indicates a purifying selection (also known as negative selection), where the new phenotype is mostly deleterious and eliminated from the population [45–47]. Despite the small differences in the dN/dS ratios, purifying selection (dN/dS < 1) is strongly acting at the VP1 protein level in all norovirus genotypes, including GII.4.
We next examined whether the different patterns of diversification observed in ORF2 extended to other regions of the genome. We analyzed by NGS full-length norovirus genomes from immunocompetent individuals infected in different settings. The first set of samples was from a child who was consecutively infected with three different genotypes (GII.4, GII.6 and GII.17) over a 3-year period [18, 35]. Although in each episode the child resolved the symptoms within ~ 72 hours, viral RNA was detected in stools for weeks after onset of symptoms. The full-length genome sequences were compared within the first 3 weeks for GII.4 and GII.6 viruses and the first 2 weeks for GII.17 viruses. A total of up to 78,000 reads/site (mean: 14688, standard deviation: 5675) were obtained for each sample by NGS (S1 Table). All consensus sequences were identical to the reference (day 1 [d1]); however, mutations ranging from 5 to 50% of the total reads (S1 Table and Fig 3) were found at later time points for each virus. In GII.4 viruses, 12 nt mutations arose in the subpopulations of the sample collected at d14, with nine of them being non-synonymous mutations. Three aa mutations were located in the P domain, the capsid surface-exposed region of the VP1 protein, with two present in antigenic sites A (E368A) and E (S412N) [48] (S3 Fig). By d21, the GII.4 virus had acquired 20 nt mutations in its subpopulations, with most of them (19/20) new mutations as compared to the d14 sequence. Nine out of the 20 nt mutations changed the aa sequence, with 4 of them mapping to the P domain, near or at antigenic sites (S3 Fig). In contrast, GII.6 and GII.17 viruses presented only four and two substitutions in their subpopulations at d21 and d14, respectively, with no evidence of the accumulation of mutations over time (S2 Table).
Although a large amount of norovirus is shed in stool, the infectious dose in natural transmission is likely low [41]. Thus, during inter-host transmission events noroviruses may undergo an initial reduction in the number of replicating viruses, creating a bottleneck effect. To compare inter-host evolution in individuals involved in outbreaks, we analyzed samples from outbreaks that occurred in the state of Maryland where the causative agents were identified as GII.6 (a hospital outbreak in 1971) [49] or GII.4 noroviruses (nursing home outbreaks in the 1987–1988 winter season) [32]. The samples and their dates of collection are indicated in Fig 4. Comparison of the NGS sequences with the outbreak consensus sequence revealed only a few substitutions, ≤ 5 nt and ≤ 2 aa, among samples from the same outbreak (Fig 4A and Fig 4B). However, when the consensus sequences from the GII.4 outbreaks were compared, a progressive accumulation of mutations (up to 86 nt and 16 aa) were detected in a period of three months, with no aa substitutions detected in the ORF2 (Fig 4B). To confirm these observations, we compared 151 genomes from GII.4 viruses (variant Den Haag) detected during three epidemic seasons in Japan [50, 51], and showed the accumulation of nt and aa substitutions over time (S4 Fig). In both sets of samples, the ORF2 (encoding VP1) acquired fewer amino acid substitutions as compared with ORF1 and ORF3, and thus maintained the VP1 phenotype for the GII.4 variant circulating in that given season. The data from full-length genome analyses were consistent with those from analyses of ORF2 in the GenBank database: different patterns of evolution exist among the norovirus genotypes in an acute outbreak setting.
To reconcile our observations on the different mechanisms of diversification and data on re-infection and epidemiology of noroviruses, we investigated whether additional relationships might exist among the genotypes from the two major genogroups. Our phylogenetic tree constructed with representative strains from each genotype (strains from each lineage described here were included) showed clustering among certain genotypes (e.g. GI.3, GI.7, GI.8 and GI.9), while others appeared as single genotypes (e.g. GI.1, GII.3, GII.6; Fig 5A, S3 Table). The genotypic clustering was reproducible with a second phylogenetic methodology (S5 Fig). We designated each of the separate branches as groups A-L (Fig 5A), and the deduced aa sequences showed an approximate cut-off value of ≥20% aa differences between groups (Fig 5B). In a review of data from research groups that have documented norovirus re-infections and determined the genotype for each infection [17, 18, 38, 52–54], we observed that the pattern of re-infection might be consistent with the new grouping system as a predictor of antigenically-distinct strains. To test the hypothesis that these groups, provisionally designated here as “immunotypes,” might play a role in norovirus immunity, we developed a matrix that recorded the data from each of the consecutive re-infection cases with documented norovirus genotyping [17, 18, 22, 38, 52–54]. For example, a child consecutively infected with a GII.4 (“immunotype” G), GII.6 (“immunotype” H), and GII.17 (“immunotype” J) norovirus would count as one individual for the cell of the matrix that compares immunotype G and H, and as one individual for the cell that compares immunotype H and J. A matrix was constructed using re-infection data available from 116 children and 2 adults (Fig 5C, S6 Fig). Overall, the majority of re-infections occurred with viruses from different immunotypes, with re-infection rare from strains within an immunotype. A notable exception was immunotype G, which is comprised of genotypes GII.4 and GII.20. Re-infection of eight children (as shown by the 8 individuals in the black cell) was documented to have occurred with different variants of GII.4 viruses [17, 38].
Viruses are genetically and structurally diverse. Depending on their genome and/or replication strategies, viruses can present different rates of evolution (range: 10−2–10−9 nt substitutions/site/year) [3, 47, 55]. As with many other RNA viruses, noroviruses have been regarded as rapidly evolving viruses [22, 48, 56]. The overall rate of evolution for the norovirus genotypes included in this study ranged from 5.40x10-3–2.23x10-4 nt substitutions/site/year for the VP1 encoding region, which were similar to those described previously for norovirus GII.4, GII.3, GII.6, and GI.1-GI.6 [21, 57–61], and within the range for positive-strand RNA viruses [3]. Despite this high nt mutation rate, the number of non-synonymous substitutions were on average ~18 times lower than the synonymous substitution (dN/dS average: 0.06), suggesting that purifying selection (dN/dS <1) acts strongly in the VP1 protein. Similar observations have been made for other RNA viruses, where the rate of evolution reached up to 10−2 nt substitutions/site/year (depending on the region of the genome used for analyses) but was mostly dominated by high synonymous substitution rates [46, 55, 62]. In noroviruses, positive selection has been reported for certain codons of the VP1 for GII.4, GII.3, GII.6 and GII.17 viruses [21, 22, 57, 58, 63], and codon changes in the antigenic sites of GII.4 viruses (which are located in loops of the P2 domain) have correlated with the emergence of new variants [24, 27, 48]. Taken together, our findings suggest that the capsid protein of all noroviruses evolve with strong structural constraints, with only a limited number of codons that can evolve and, perhaps confer adaptive advantages to infect human hosts.
Epidemiological studies coupled with sequence data from field isolates have indicated that the most predominant norovirus genotype, GII.4, is evolving similarly to influenza H3N2 viruses; i.e. with a temporal replacement of predominant variants that is driven by the immune response of the host [27, 48, 64]. By exploring the intra-genotypic diversity from representative human norovirus genotypes we verified that GII.4 noroviruses produce the largest number of intra-genotypic variants, and that these variants last (on average) ~5 years in the human population. In contrast, non-GII.4 noroviruses sustain a low number of intra-genotypic variants with a limited number of aa differences among strains within that given variant; even if decades apart in occurrence. Interestingly, different variants from a given genotype can often be co-circulating within the same year and geographical location causing gastroenteritis [37, 44, 65, 66]. The GII.4 viruses, and to a lesser extent one variant of the GII.17 viruses, acquired aa substitutions over time that created phenotypically different variants. In contrast, all other genotypes retained similar sequences within variants that might have arisen early in the origin of that genotype and that persisted over time. This led us to discriminate two different patterns of evolution in norovirus: evolving and static. Evolving viruses continually accumulate mutations in their genome over time, and static viruses do not.
The concept of evolving versus static norovirus genotypes may be helpful in understanding the spread of pandemic strains. The recent emergence of GII.17 viruses resulted in the rapid replacement of one variant (variant C) with another (variant D) [22, 44]. This pattern of very rapid replacement, occurring within two consecutive seasons, in the emerging GII.17 viruses is notably different from that of GII.4 viruses, in which each emerging GII.4 variant is replaced every 3 to 8 years. Thus, since the GII.17 genotype presents other variants shown to be “static,” the recent global spread of the GII.17 genotype might be the moment when a new genotypic variant (variant C) emerged and is quickly adapting to reach maximum fitness in the human host (variant D) to become static. Since the emergence of this GII.17 strain has only recently occurred and most of the available GII.17 sequences (136/143) correspond to these two variants, more information on pre-2013 strains and the future epidemiological behavior of the GII.17 strains will be helpful in establishing the evolutionary pattern of this genotype. Because recombination has been suggested to play an important role on the emergence of many GII.4 variants [67], and the emerging GII.17 strains presented a novel polymerase (encoded by ORF1) [22, 34, 35, 44], further studies should be conducted on the role of recombination in norovirus VP1 diversification into variants.
To determine the role of intra-host evolution at the genomic level, we developed a method to generate and analyze full-length norovirus genomes with NGS technologies and bioinformatics. The strategy of amplification was similar to that published by Eden et al. [67] for GII.4 viruses, and our method was robust for a number of GII noroviruses (GII.1, GII.2, GII.3, GII.4, GII.6, GII.12, and GII.17), and from samples stored for over 40 years [35]. Several groups have explored the intra-host diversity of noroviruses by NGS using partial regions of the genome [23, 68]; however, our approach extended these findings by allowing high-resolution analysis at every nt position in the coding sequence of the genome. We first examined the intra-host evolution of GII.4, GII.6 and GII.17 noroviruses within a single patient, and observed that only the GII.4 viruses presented a gradual increase in the number of mutations, which in some cases resulted in aa substitutions in areas regarded as important antigenic sites. The limited intra-host diversity found during the shedding phase of an infection in immunocompetent individuals contrasts with the vast diversity of viruses found in immunocompromised patients [68]. Due to the diversity found in immunocompromised patients and prolonged shedding, it was suggested that they might be a source of new GII.4 variants to the human population [69]. Noroviruses are highly transmissible; however, there is little evidence that norovirus can be efficiently transmitted during the chronic phase of the infection [19]. A more likely source for new GII.4 variants might be immunocompetent individuals, where we show that mutations can arise during inter-host transmission events, and accumulate during the intra-variant period. Although noroviruses belonging to the “static” genotypes can also accumulate mutations during inter-host transmission events, those mutations would likely be eliminated from the viral population by purifying selection. Viruses that better tolerate the introduction of mutations are regarded as genetically robust, and this robustness has been shown to be beneficial for virus survival and prevalence [70]. Overall our data suggest that GII.4 noroviruses are genetically robust. In contrast, noroviruses with “static” genotypes may be genetically fragile, which limits their antigenic diversity and prevalence.
How do “static” genotypes prevail in the human population, in the face of limited antigenic diversity within the genotype? To address this question, genotypes were grouped together based on phylogenetic clustering and aa differences in their capsid proteins. These groups, or “immunotypes,” were applied to the interpretation of epidemiological observations. When examining data from a birth cohort study, or reports where children and adults were followed for years to study norovirus re-infections, genotypes belonging to the same immunotype generally did not re-infect these individuals. Thus, most of these individuals were re-infected with a varying series of genotypes (predominantly containing combinations of GII.4, GII.6, GII.3, GII.17 or GII.2), but all of them belonging to different immunotypes as defined in Fig 5C. The exception to this was the GII.4 strains in immunotype G, in which a few re-infections were observed, albeit with different GII.4 variants. Based on these data, we propose a model for norovirus re-infection in which naïve children are constantly exposed and infected with strains from each of the different immunotypes until a broad immunity develops. In contrast, older individuals (i.e. older children and adults) are more likely to become ill from evolving genotypes, as they have already acquired immunity against a number of static genotypes (Fig 6). This model not only explains the differences in the genotype distribution often seen when comparing children and adult populations [17, 37, 38], but also suggests that immunity against norovirus may be longer than initially suggested [39, 42].
For decades understanding of norovirus immunity was based on human volunteer challenge studies and animal models or in vitro surrogates of neutralization tests [27, 39, 71, 72]. Initial cross-challenge studies, conducted in the 1970s using the prototype GI.1 Norwalk virus and GII.1 Hawaii virus, showed a lack of protection between these two genogroups [39]. Further epidemiological data and in vitro assays, such as antibody blockage of carbohydrate binding to VLPs, suggested a role for immunity against the different intra-genotypic variants of GII.4 [27, 33, 58]. Norovirus vaccines are currently based on the premise to include at least two major antigens for noroviruses representing GI and GII [29, 30, 71, 73, 74]. However, recent data indicating that certain genotype-specific immune responses were unable to confer natural protection against disease raised concerns that a prohibitive number of components (almost 30) might be needed in a norovirus vaccine [17, 18, 35, 38]. Although additional studies will be needed to confirm the existence of shared antigenic groups among the norovirus genotypes, preferably by neutralization assays or animal models, our analysis provides a new perspective on the genetic and antigenic diversity of noroviruses that could lead to the identification of cross-protective strains and inform vaccine design.
Stool specimens from the child were obtained with the written informed consent of the parent, and enrollment in National Institutes of Health (NIH) clinical study NCT01306084. Archival stool samples stored in the Laboratory of Infectious Diseases Calicivirus Repository were waived as exempt from IRB review by the NIH Office of Human Subjects Research and Protection (OHSRP 11833). Epidemiological information relating to the sample collection has been published elsewhere [18, 32, 35, 49].
The full-length (or nearly full-length) ORF2 sequences (encoding for VP1) from each of the 31 genotypes described for GI and GII were retrieved from GenBank (accessed on March 2015) for analyses. Alignments were performed with Clustal W as implemented in MEGA v6 [75]. Sequences from each genotype were aligned separately to minimize the presence of insertions or deletions (indels), which can arise when different genotypes are compared. Phylogenetic trees were constructed using Kimura 2-parameter as method of nt substitution and Neighbor-Joining as algorithm of reconstruction as implemented in MEGA v6 with default settings. Phylogenetic trees that used aa sequences were reconstructed using a Poisson method of aa substitution. Bootstrap analyses were used to support the clustering of the variants. Information on the strains used for the phylogenetic analyses is provided in S4 Table. Evolutionary rates (nt substitutions/site/year) for each genotype were estimated using the ORF2 sequences and the Bayesian Markov Chain Monte Carlo (MCMC) approach as implemented in the BEAST package [76]. For each set of data the General Time Reversible (GTR) model with gamma rate distribution and invariable sites parameter was used and the MCMC was run for a sufficient number of generations to reach convergence of all parameters. All evolutionary rates were calculated using strict clock model and coalescent constant size tree prior, except for genotypes GI.4, GI.6, GII.14 and GII.16, which reached convergence using Bayesian Skyline and random local clocks. Selection pressures acting in the VP1 sequences were investigated by estimating the mean rate of nonsynonymous substitutions (dN) and synonymous substitutions (dN) and the dN/dS ratio as implemented in MEGA v6. The nearly full-length genome sequences from 151 GII.4 viruses detected in Japan during 2006–2009 [50, 51] were downloaded from GenBank and analyzed using MEGA v6 and Prism software (GraphPad Prism version 7).
To visualize the aa substitutions within each genotype, a Python script (available upon request) was developed to calculate the number of aa differences and the isolation year differences between two individual strains. Isolation years were extracted from strain descriptions. The difference values were added into a matrix where the y-axis represents the isolation year differences and the x-axis the amino acid differences. Note that some cells will present more than one comparison, since strain pairs presenting the same number of aa differences and the same year difference, despite the years detected, will be included in the same cell. Heat map plots were calculated for each genotype using GraphPad Prism version 7 (GraphPad Software, La Jolla California USA), with the values representing the number of strains compared.
A platform was developed to analyze the plasticity of norovirus genotypes at the full genome level. Briefly, viral RNA was extracted from 10% (w/v) stool suspensions using the MagMax Viral RNA Isolation Kit (Ambion, California, USA) following manufacturer’s recommendations. Complementary DNA was synthesized from the viral RNA using the Tx30SXN primer (GACTAGTTCTAGATCGCGAGCGGCCGCCCTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT [77]) at 5μM final concentration, and the Maxima H Minus First Strand cDNA Synthesis Kit (Thermo Fisher Scientific, California, USA) following manufacturer’s recommendations except that only 0.1 μL of Enzyme Mix was used per reaction. Amplification of the full-length genome was performed using 5 μl of the RT reaction, a set of primers that target the conserved regions of the 5’- and 3’-end of GII noroviruses (GII1-35: GTGAATGAAGATGGCGTCTAACGACGCTTCCGCTG, and Tx30SXN), and the SequalPrep Long PCR Kit (Invitrogen, California, USA) following manufacturer’s recommendations. Amplicons were excised from an agarose gel and purified with the QIAquick Gel Extraction Kit (Qiagen, California, USA).
Ion Torrent libraries were prepared by using 300–500 ng of full-length genome PCR amplicons following standard Ion Torrent library prep protocol. DNA was fragmented followed by the introduction of ligation barcode adapters. Adapted-ligated libraries were amplified using 13 PCR cycles, and size selected from agarose gels. Final libraries were quantified by Qubit (Invitrogen, California, USA), Bioanalyzer (Agilent), and qPCR. Libraries were normalized to 1nM, pooled at an equal molar ratio, and loaded onto a 318 v2 Chip in an Ion OneTouch2 machine. The sample from Ion OneTouch2 was transferred to an Ion OneTouch ES and then to an Ion PGM for sequencing with a 400bp kit (Life Technologies, California, USA). Ion Torrent sequence reads were de-multiplexed, and each individual set of reads was aligned to reference sequences using Bowtie2 and SAMtools [78, 79]. Aligned reads were visualized in the Integrative Genomics Viewer (IGV) [80] for single nt polymorphisms (SNPs) identification. Consensus sequence for each full-length genome was calculated using IGV. Read coverage (reads/nt position) was calculated using the genomecov command from BEDTools [81]. Sequence analyses were performed using MEGA v6 and Sequencher 5.4 (Gene Codes Corporation, Michigan, USA). The consensus sequence was calculated using default settings in Sequencher v5.4, and genomic sequences determined in this study were deposited into GenBank under Accession numbers KY424328 through KY424350. All other relevant data are within the paper and its Supporting Information files.
|
10.1371/journal.pcbi.1005823 | Automated classification of dolphin echolocation click types from the Gulf of Mexico | Delphinids produce large numbers of short duration, broadband echolocation clicks which may be useful for species classification in passive acoustic monitoring efforts. A challenge in echolocation click classification is to overcome the many sources of variability to recognize underlying patterns across many detections. An automated unsupervised network-based classification method was developed to simulate the approach a human analyst uses when categorizing click types: Clusters of similar clicks were identified by incorporating multiple click characteristics (spectral shape and inter-click interval distributions) to distinguish within-type from between-type variation, and identify distinct, persistent click types. Once click types were established, an algorithm for classifying novel detections using existing clusters was tested. The automated classification method was applied to a dataset of 52 million clicks detected across five monitoring sites over two years in the Gulf of Mexico (GOM). Seven distinct click types were identified, one of which is known to be associated with an acoustically identifiable delphinid (Risso’s dolphin) and six of which are not yet identified. All types occurred at multiple monitoring locations, but the relative occurrence of types varied, particularly between continental shelf and slope locations. Automatically-identified click types from autonomous seafloor recorders without verifiable species identification were compared with clicks detected on sea-surface towed hydrophone arrays in the presence of visually identified delphinid species. These comparisons suggest potential species identities for the animals producing some echolocation click types. The network-based classification method presented here is effective for rapid, unsupervised delphinid click classification across large datasets in which the click types may not be known a priori.
| Health of marine mammal populations is often considered an indicator of overall marine ecosystem health and resilience, particularly in highly-impacted regions such as the Gulf of Mexico. Marine mammal populations are difficult to monitor given the many challenges of observing animals at sea (e.g. weather, limited daylight, ocean conditions, and expense). An increasingly common approach is the use of underwater acoustic sensors capable of recording marine mammal calls at remote locations for months at a time. Acoustic sensors generate large datasets in which dolphin echolocation clicks are commonly present. Dolphins are the most diverse family of marine mammals, and distinguishing click characteristics have only been described for a small subset of species. We developed a workflow to automatically identify distinct dolphin click types within large datasets without prior knowledge of their distinguishing features. Our algorithm then recognizes these click types in novel recording data across a range of monitoring locations. Known species-specific click types emerge from the data using this approach, as well as new click types potentially associated with additional species. This technique is a key step toward determining species identification for passive acoustic monitoring of offshore populations of dolphins and other toothed whales under a big data paradigm.
| Dolphins produce echolocation clicks while socializing, foraging and traveling [1]. The prevalence of echolocation clicks makes these signals useful for monitoring delphinids using passive acoustic methods; however, only a few delphinid click types can currently be classified to species. Echolocation clicks have a suite of characteristics that make them challenging to classify in acoustic recordings. For example, echolocation clicks are highly directional signals which can be received “on-axis” (animal oriented in the direction of the recording sensor while clicking) or “off-axis” (animal oriented away from the sensor), leading to differences in amplitude and interference patterns [2]. Dolphin echolocation click signals also significantly attenuate over relatively short distances due to their high frequency acoustic content; therefore, the orientation and proximity of a clicking animal relative to an acoustic sensor has a large effect on the frequency structure of the recorded click [3, 4]. Behaviorally, individual dolphins may vary click source levels and beam widths [5–8]. Furthermore, dolphins are typically found in large, sometimes multi-species groups in which animals vocalize simultaneously. All of these factors contribute to click variability and therefore complexity in classification.
Despite these sources of variability, echolocation clicks of a few delphinid species as well as many beaked whale species have known species-specific spectral features [9–12]. Consistent features have typically been recognized by expert analysts manually reviewing large amounts of data. Previously identified characteristic spectral features include mean frequency, bandwidth, and peaks or troughs in frequency spectra indicating dominant or diminished frequencies. Typical inter-click interval (ICI) ranges also differ between beaked whale species [13], and ICI is used to identify porpoise click trains to species [14, 15], although ICI may vary as a function of depth or behavior in some cases [1, 16, 17].
A challenge in echolocation click classification is to overcome the many sources of variability to recognize underlying consistent patterns. One approach is to train analysts to recognize patterns. Humans are particularly adept at pattern recognition tasks: With enough training time, contextual information and training data, an analyst can distinguish within-type and between-type click variations, and develop a sense of the major click categories in a dataset. However this is an iterative, time-consuming and potentially subjective process.
An alternative is to develop automated methods to perform echolocation click classification. Within a computational framework, one approach to the click variability problem is to consider a set of clicks as a group of objects that are similar but not identical to one another. In a simple example with five clicks labeled A through E, consider a case where clicks A, B and C are very similar, click D is slightly different, and click E is very different than A-C, with some similarity to D. In this case, clicks A, B and C are regarded as the most informative for classification, as they contain consistent features among them, while clicks D and E are likely outliers. We might consider A, B and C to be members of a group characterized by their common feature set. In practice, an actively echolocating dolphin produces multiple clicks per second. Therefore, a similar but more complex case exists in which a subset of those clicks will be highly interrelated, while others are only weakly associated.
This approach to the variability problem can be represented as a weighted network [18], in which clicks are represented by nodes and the lines or edges between nodes represent the strength of the similarity between them. In the example above of echolocation clicks A through E, the click characteristic inter-relationships are represented by a network with larger edge weights among similar clicks A-C and lower value edge weights among clicks D and E and their neighbors which show their greater dissimilarity from clicks A-C and each other (Fig 1). A network of N nodes can also be represented as an adjacency matrix G in which G(i,j) represents the weight of the edge between nodes i and j, for i and j ∈ the set of nodes N [19].
Once the relationships between a set of clicks are represented as a network, an unsupervised learning algorithm can be used to identify clusters of highly similar clicks. Here we use an agglomerative clustering routine [20] that seeks to identify structure within the network without a priori information about what that structure might be. Using this method, nodes within the network are iteratively grouped together based on the strengths of the edges between them. This method can converge to a single large cluster if all nodes are highly interrelated, but multiple clusters can be identified if interrelationships are not evenly spread across the network.
In this work, unsupervised network-based classification methods are applied to the problem of delphinid echolocation click classification in the Gulf of Mexico (GOM). Long-term passive acoustic monitoring efforts using autonomous near-seafloor hydrophones at five sites in the GOM have resulted in a dataset of over 52 million unlabeled dolphin echolocation clicks. Thirteen delphinid species are known to inhabit the GOM, including five members of the genus Stenella, and five species belonging to the subfamily Globicephalinae (Table 1). Three of these five species, Risso’s dolphin (Grampus griseus), false killer whale (Pseudorca crassidens) and short-finned pilot whale (Globicephala macrorynchus) can be distinguished based on echolocation click characteristics [11, 21]; however, few other species have been conclusively identified.
Our objectives are to develop a technique for recognizing candidate click types in this dataset which may be associated with species that are not yet acoustically identifiable, and to demonstrate a method for recognizing these click types automatically in novel data. Further, we support the utility of this method by comparing automatically identified types with clicks recorded using towed hydrophone arrays in the presence of vocalizing animals from the western Atlantic whose species identity has been verified by trained visual observers. The described click types are informative for passive acoustic delphinid population monitoring efforts, while the methods offer an approach for automated classification of variable signals in large unlabeled acoustic datasets.
Long term passive acoustic recordings were collected at three continental slope sites (sites MC, GC, and DT), and two shelf sites (sites DC and MP). Delphinid clicks were automatically detected in large numbers during all deployments at each site, with click counts ranging from 5.2x105 to over 8.1x106 analyst-confirmed detections per deployment (between 6,000 and 67,000 clicks per day; Table 2). Detections were grouped into 5-minute bins marked as click-positive or negative. The number of click-positive 5-minute bins per deployment varied from almost 5,000 to close to 12,000 bins (unnormalized for recording effort). The average number of delphinid echolocation encounters (periods of continuous click detections bounded before and after by at least 15 minutes without click detections) per recording day ranged from 1.4 to 7.9 across deployments. Average encounter durations were generally shorter at the shelf sites MP and DC; however, encounter durations were highly variable at all sites and ranged from 1 to 640 minutes. Across all deployments, between 0.1% and 10.1% of click-positive bins contained more than 5000 clicks and were sub-sampled for classification purposes. The most sub-sampled site was site DT.
Phases 1 and 2 were run on the full training set following the exploratory analysis. In Phase 1, the average number of automatically identified clusters per time bin ranged from 1.02 to 1.14 (CV = 0.14 and 0.35 respectively) across sites and deployments (Table 2). In Phase 2, seven dominant and recurrent click types (A-G) characterized by consistent spectral shapes and modal ICIs were identified (Table 3, Fig 3). We define the modal ICI as the most frequently observed ICI during a period of clicking.
Click type A was identified in the training data from the three deep sites, and one shallow site. Most instances came from site GC. This type was characterized by a minor narrow low frequency peak near 12 kHz, dominant energy between 20 and 35 kHz, and 0.15 sec modal ICI.
Click type B was identified in the training data from all sites except site GC. This click type, presumed to be Risso’s dolphin based on Soldevilla et al. [11] and has distinct narrow energy peaks at ~ 22, 26, and 33 kHz. The ICI distribution for this type was bimodal with shorter ICIs near 0.12 sec at the northern sites, and longer ICIs over 0.23 sec at the southern site DT.
Click type C was identified in the training data from the deep sites only. This click type had the lowest frequency content of dominant energy between ~15 and 30 kHz, and a modal ICI of 0.16 sec.
Click type D was identified in the training data from site DC, and in one bin from site MP. This click type was characterized by two low frequency peaks at 12 and 18 kHz, dominant energy between 30 and 50 kHz, and a bimodal modal ICI with peaks at 0.03 and 0.09 sec.
Click type E was identified in the training data from all five sites and represented 22% of the training set. It was particularly common at the southern site DT. Click type E was characterized by minimal energy below 20 kHz, a dominant spectral peak near 30 kHz, and a modal ICI of 0.06 sec. Spectral variability below 20 kHz may indicate the presence of multiple subtypes, or overlap with click type F.
Click type F was identified in the training data from all five sites and represented 47% of the training set. This type was similar to type E, had a minor energy peak at approximately 20 kHz. Some examples had a third peak at 16 kHz. High variability of this type in the 10–25 kHz band suggests that it may include multiple subtypes. This type had a strong modal ICI at 0.06 sec.
Click type G was only identified in the training data from the two shallow sites only: Sites DC and MP. It was characterized by a broad high energy band between 15 and 52 kHz, and a peak frequency of 36 kHz and a modal ICI of 0.03 sec.
Delphinid clicks are very short duration, highly variable signals which contain limited information when considered individually. The automated clustering strategy was designed to mimic a human analyst by comparing large numbers of clicks to identify persistent features. Leveraging multiple sources of information such as spectral shape and ICI distributions across bins of similar clicks further facilitated pattern recognition and click type distinction. The two-step training process tackled the large dataset by reducing the number of comparisons necessary through use of filtered means and modes.
A variety of different pruning and clustering techniques were implemented during the algorithm development process. In the final implementation, edge pruning was executed using a dynamic metric in which the weakest N% of edges were pruned from each network. Using this approach, networks of highly similar nodes and networks of weakly similar nodes were pruned by the same amount. An alternate approach would be to prune all edges weaker than a static threshold value. Using the static approach, a network of weakly interrelated nodes would be pruned more heavily than a network of strongly interrelated nodes. Both approaches were tested during development of the clustering protocol, but the dynamic metric was ultimately chosen as the more conservative pruning method for preserving click types with smaller sample sizes. More aggressive pruning at site MP might reduce inclusion of false positives associated with snapping shrimp and improve classification accuracy if snap spectra are more variable than click spectra.
A more complex, greedy clustering algorithm [modularity; 24, 25], preliminarily used during the development process, was not able to reliably identify clusters of different sizes. The simpler CW algorithm used in the final implementation identified both small and large clusters within a network, which is essential in identifying less common click types. Further click type separation may be possible however. In this dataset, some click types had very different spectral shapes and ICIs from one another such as type A and B clicks, while others were similar, such as type E and F clicks. This is a challenging situation for clustering purposes, because some types separate well, while others remain intermingled, as in the case of types E and F where spectral variability may represent multiple sub-types. In Phase 2, a multi-pass clustering approach in which thresholds were incrementally increased might enable better distinction between similar types such as those within type E without over-pruning highly distinct types. Reduced within-cluster variability would probably also reduce classifier confusion and improve accuracy.
ICI and spectral similarities (both values between 0 and 1), were combined in Phase 2 of the automated classification process by simple multiplication. The multiplicative rule was used because analysts typically needed both robust ICI and spectral information to make a confident classification. The two metrics did not necessarily contribute equally to the overall similarity scores because although they are both values between [0,1], they did not have identical distributions. Transforming the distribution of either parameter prior to multiplication would modify the influence of the parameter on the Phase 2 network. For example, if spectra were deemed more reliable than ICI, SSPEC could be transformed prior to Eq (2) to give it more influence on the network. For classification of the test set, the multiplication method requires that both score high to achieve a high overall similarity score.
An earlier implementation of this algorithm used correlation distance between ICI distributions instead of distance between modal ICIs. This strategy produced similar results but performance suffered when classifying bins with high click counts. As the number of detections per bin increased, click trains tended to become interleaved, resulting in higher numbers of low ICIs. While true ICIs from a single animal’s click train may be a species-specific feature [26], the interval between clicks received from multiple individuals’ trains is not informative. Similarly, high false positive rates associated with snapping shrimp at site MP affected ICI distributions. Modal ICI, which likely represents individuals’ ICIs, was found to be less sensitive to differences in click counts per bin and more robust to false positives. Modal ICI may be more difficult to detect for species that are often found in very large groups.
The unsupervised click classification routine identified seven distinct delphinid click types in the training data across five sites in the Gulf of Mexico based on frequency content and modal ICI. All types were identified at a minimum of two sites, and over half were identified at four or more sites.
One hypothesis of what is driving the persistent features leading to the click type clusters is site-specific propagation and noise conditions; however, a number of features demonstrated here are inconsistent with this explanation. First, site-specific noise and propagation do not explain why multiple click types were found at each site, often within the same day or in overlapping encounters, nor do they explain why the same click types were found at multiple sites, despite differences in noise, site depth, and site location. Second, site-specific propagation and noise would be expected to affect all clicks in the same way; therefore, they do not explain why some click types have complex spectra with peaks and troughs, or why frequency distributions differ between types under similar noise conditions. Third, site-specific conditions do not offer an explanation for the consistent relationships between click type spectral shape and ICI distributions across deployments spanning multiple years, or why ICI distributions have consistent modal values.
Alternative hypotheses are that the distinct click types identified in this dataset represent different dolphin species or echolocation clicks used in different contexts [e.g. 27]. Species differences may explain these observations. Echolocation click frequency content and click rates have been shown to differ between odontocetes such as sperm whales, beaked whales, dolphins, and porpoises [e.g. 11, 12, 13, 28]; therefore, it is reasonable to expect that these features may also differ between delphinid genera and/or species. Consistent ICIs have been reported for beaked whale species [e.g. 13] and similar consistency may be typical of some delphinids [29]. Spectral content may vary depending on target prey [9], and ICI may be related to click source level, frequency content, and/or prey search distance [e.g. 30, 31]. Low frequency, high amplitude clicks have the potential to propagate farther than high frequency or low amplitude clicks. This may result in a longer two-way travel time for each click. Delphinids may compensate with a longer ICI to allow for the longer travel times.
The majority of clicks detected at the three deepest sites were associated with types E and F which had similar spectral shapes and modal ICIs. According to the most recent NOAA stock assessments [22, 23] based on summer visual surveys, approximately 80% of offshore delphinids in the GOM are members of the Stenella genus, of which spinner and pantropical spotted dolphins are the most common species. Two additional Stenellid species, striped and Clymene dolphins, are also found offshore, although population estimates vary widely between surveys. A fifth species, Atlantic spotted dolphin, is found primarily on the continental shelf. Based on the high abundance of Stenellids as a proportion of GOM delphinids, Stenellid dolphins are the most likely match for type E and F clicks. Considerable variability below 20 kHz within sites in the type E and F clusters suggests that they may include multiple subtypes, possibly representing different species. Towed hydrophone array recordings made in the presence of pantropical and Atlantic spotted dolphins revealed ICIs that were consistent with type E and F clicks. Distributions of the various Stenellid species differ in the GOM [32], and this may account for the different ratios of these types across sites.
Based on visual survey data, species composition and abundance is expected to differ between the three deeper slope sites (GC, MC, and DT) and two shallower shelf sites (MP and DC). Primary species at the shallow sites include Atlantic spotted dolphin (also a member of the genus Stenella) and bottlenose dolphin [32]. Rough-toothed dolphins have also been observed near site DC, although in much lower numbers. Click type G which was common at the two shallow sites but was not identified at deeper locations, and click type D which was predominantly identified at site DC, are likely associated with some of these species. Snapping shrimp snaps were a common source of false positives at site MP, where click type G was primarily detected. Distributions associated with this click type may have been contaminated by snap signals. In future work, click train tracking could be used to improve ICI estimates in noisy, shallow water environments, and encounters with very high click counts.
Click Type B likely represents Risso’s dolphin clicks as it contains the consistent peaks and notches described for Risso’s dolphins in the Southern California Bight, and further matches the peak structure documented here from a towed array recording of visually-verified Risso’s dolphins from the western Atlantic. Modal ICI differed between the three northern sites (MC, DC, and MP) and the southern site (DT), suggesting possible behavioral or population differences.
Click type A may represent short-finned pilot whale clicks as it is similar to Atlantic pilot whale (presumed short-finned) recordings collected using towed hydrophone arrays. However, it differs from a recent description of Pacific short-finned pilot whale clicks which found spectral peaks at 12 and 18 kHz collected in the Hawaiian Islands [21]. Click type A was most common at site GC in this dataset, which is consistent with short-finned pilot whales’ predominantly eastern GOM distribution based on visual surveys [32].
The narrower bandwidth of click type C centered at lower frequencies is consistent with published descriptions of false killer whale (Pseudorca crassidens) echolocation clicks [9, 21] from the Eastern Pacific. However, there are no published estimates of modal ICI for false killer whales. Across all sites, 1.3% of bins were classified as Type C. The most recent stock assessment estimates place false killer whales as approximately 1% of offshore GOM delphinids.
Melon-headed whales are expected in low densities the GOM, but information regarding distinguishing features of these clicks is limited [12], and no clear match was identified. Killer whale, pygmy killer whale and Fraser’s dolphin, although present in the GOM, may be too rare at these sites to be identified using these methods [23]. Use of a larger training set with a multi-pass strategy in which dominant types, such as E and F, were identified and removed could facilitate recognition of rare types.
A subset of the identified click types had characteristics in common with clicks recorded in the presence of visually-identified species recorded using the towed hydrophone array. Unfortunately, with the exception of the pantropical spotted dolphin data, these recordings were collected in the Atlantic and can only be tentatively compared with GOM click types. Towed array hydrophones are typically much shallower than seafloor instruments, therefore the effect of acoustic propagation on recorded signals differs. Further work will seek to solidify and extend comparisons between seafloor sensor types and towed array recordings of known species, with an emphasis on collecting recordings of visually identified species in the GOM.
The towed array environment is different from that of the seafloor sensor. Towed array recordings are much more affected by vessel, ship-based electronic and wind-generated sea-surface noise, and shallow sound-speed profiles than autonomous seafloor recordings. The orientations of animals relative to the sensors differ between the two types of recordings, for example, during a ship survey, dolphins are often oriented toward the bow, while the sensor is towed behind the vessel; whereas seafloor instruments are located below dolphins traveling near the sea surface, and do not typically influence dolphin orientations. Animal behaviors likely differ as well because marine mammal surveys require daylight for visual marine mammal identification, but seafloor sensor recordings typically show that most delphinid clicks are detected at night [29]. In addition, comparisons of simultaneous towed array and HARP recordings in the GOM have shown that towed array detection ranges may be limited by signal refraction associated with warm surface mixed layer [33]. Towed array delphinid click recordings were often characterized by short encounters and highly variable click amplitudes. When animals were close enough to the towed array to be detectable, both on-axis (transmission beam oriented directly toward the sensor) and off-axis clicks were likely received, and on-axis clicks could be clipped due to high amplitudes at close range. In contrast, delphinid encounters recorded by near-seafloor HARPs were often longer in duration due to larger detection ranges. Click amplitudes tended to be lower, because delphinids were farther from the sensor, and off-axis clicks were less detectable according to click propagation simulations [34].
Several improvements could be made to the automated classification approach in future work. First, different distance metrics could be evaluated. In this study, a correlation distance metric was used to assess similarity between spectra as it was found to capture shape similarities more effectively than a simpler Euclidean distance. However, the correlation distance used assigns equal weight to all frequencies in the spectra, while high frequency amplitudes are expected to vary more than low frequencies because of acoustic attenuation. To account for this expectation, a weighted distance metric could be used that emphasizes low frequency shape. Alternatively, click shapes could be summarized as cepstra (inverse FFT of spectra, e.g. [28]) to emphasize particular aspects of overall shape. Preliminary experiments using cepstra and perceptual weighting were conducted as part of this study, however clustering results were poor. Nonetheless, more complex weighting strategies might be useful in future work.
Another improvement that could be considered is to evaluate the impact of pre-filtering spectra prior to classification. In this implementation, frequencies below 10 kHz were removed by a bandpass filter. Future classification efforts might benefit from including lower frequency spectral content. Recent work by Finneran et al. [4] suggests that delphinid clicks may have consistent spectral features below 10 kHz which may be useful for click classification [e.g. 21].
Improvements could also focus on using different metrics to capture persistent features of ICIs. In this study, clear modal ICI peaks were associated with the click types, and ICI previously has been found to be useful for classifying beaked whale clicks [13]. While delphinids have been shown to vary their ICI in wild and captive studies [1, 16], this typically occurs during terminal buzzes [35] which are much lower amplitude and occur less frequently than regular clicks [35, 36] and therefore, are much less likely to be detected in wild recordings [34]. Deep seafloor instruments (at depths of roughly 80 m or more) often receive only a single animal’s click train at a given time due to the animals’ narrow transmission beam patterns and distance from seafloor sensors; therefore ICI often is accurately calculated and modal ICI is representative of persistent features. On occasions when a group of animals is large and/or close to the sensor, multiple click trains will overlap and modal ICI values may become subject to saturation. Click train tracking [37] could be used to improve modal ICI estimates in saturated cases and in noisy or shallow environments.
Additional improvements could be made to fully automate the classification process. For example, false positives were manually removed from this dataset prior to classification. However, many sources of false positives, including beaked whales, sperm whales, and ships, have their own spectral and ICI characteristics. A similar approach to that described here could be used to build template clusters for false positive sources, allowing them to be identified and excluded automatically during classification. In addition to accelerating the analysis process, this approach could improve the removal of false positives within bouts of true detections (such as at shallow sites), facilitate false positive rate calculations, and provide certainty scores for removed detections. Future work will likely seek to combine clustering with deep learning methods as a possible refinement for improved classification.
Finally, future improvements should focus on evaluating sources of variability within click types and on linking distinct click types with delphinid species identity or behavior states. This work focused on identifying distinct click types, however, more work needs to be done to describe within-type variability. Delphinids have been shown to vary their clicks depending on context [e.g. 6, 16, 27]. The types described here are broad groupings, and are not meant to indicate a lack of variability within each type. Concurrent visual identifications with HARP recordings are needed to conclusively validate potential species associations. Future steps should include applying this method to a labeled dataset associated with visually-identified species to ground truth the approach. Continued development of unsupervised learning strategies for identifying consistent dolphin click types will advance marine mammal monitoring programs by facilitating delphinid and toothed whale species identification in data collected using autonomous passive acoustic sensors.
Long-term autonomous datasets were collected using High-frequency Acoustic Recording Packages (HARPs) at three continental slope and two shelf locations in the GOM between 2010 and 2012 (Fig 6). Details of each HARP deployment are presented in Table 2.
HARPs are autonomous bottom-mounted acoustic recorders containing a hydrophone, data logger, battery power supply, ballast weights, acoustic release system, and flotation [39]. All of the seafloor recording instruments used in this study were of the same type with equivalent hardware and software. Each instrument used a calibrated hydrophone (Channel Group Technologies, Santa Barbara, CA, ITC-1042) buoyed approximately 10 m above the seafloor. Hydrophones had an approximately flat (±2 dB) sensitivity from 10 to 100 kHz of -200 dB re V/μPa. Preamplifier calibrations were performed at Scripps Institution of Oceanography and at the U.S. Navy’s Transducer Evaluation Center facility in San Diego, California [38]. All HARPs sampled continuously at 200 kHz throughout each deployment.
Towed hydrophone array recordings were collected in 2011 and 2012 (Table 6) during National Oceanographic and Atmospheric Administration’s (NOAA) National Marine Fisheries Service (NMFS) Southeast Fisheries Science Center (SEFSC) marine mammal abundance surveys aboard the R/V Gordon Gunter, conducted in the eastern GOM and within the southeastern U.S. Atlantic coastal exclusive economic zone (EEZ). A team of visual observers identified dolphins to species whenever possible, thereby providing ground-truthed species identifications which acousticians could associate with concurrent array recordings. A five-element hydrophone array was towed 274 m behind the ship, at an estimated depth of 15 to 18 m at typical survey speed (10 kn). Hydrophone elements consisted of custom-built preamplifiers, with band-pass filters set for 3 dB roll-off at 1 kHz and 200 kHz, connected to an omni-directional spherical hydrophone (HS-150 Sonar Research and Development, Ltd., Beverley, UK). Two hydrophones separated by 2.12 m were sampled continuously at 500 kHz using a data acquisition board (NI USB 6251, National Instruments Corporation, Austin, TX) and recorded directly to hard disk drives using Logger 2000 (International Fund for Animal Welfare, IFAW, Yarmouth Port, MA). The towed array recording setup differs considerably from the seafloor sensors, therefore any comparisons are considered tentative.
The set of summary nodes identified using in the training set were used to automatically classify clicks in the test dataset (Table 2). As in the classifier training, Phase 1 of the automated clustering routine was executed on click-positive bins from test data to produce a set C of n test summary nodes spanning each test deployment. To classify each test summary node Ci in C (for i = 1 to n) from the test data to one of the click type clusters T from the training data, the spectrum and modal ICI of the test node was compared to all of the training nodes in each click type Tj of P, (for j = 1,…, m), to obtain a similarity metric following similar methods as for Phase II described above. The set of similarity scores was pruned, and Ci was automatically assigned to the cluster Tj with the highest mean similarity score between the test and training summary nodes.
The mean similarity between Ci and its matching cluster Tj was retained as a metric of classification certainty. In this classification exercise, the goal was to find the best click type match for Ci, even if Ci was a poor quality example (e.g. noisy or sparse) so a lower pe threshold (pe = 0.90) was used to allow matching across a range of qualities by retaining more edges. Note also that this method of fusing spectral and ICI similarity scores using a product requires both scores to be strong in order to produce a strong match. Implications of this approach are further explored in the discussion.
To evaluate classifier performance, a systematic random sample of 200 test summary nodes from each site were manually assigned to a template cluster by a trained analyst reviewing mean spectra and ICI distributions of the test nodes. Test nodes that were not clearly similar to any of the click type clusters were labeled “unknown” by the analyst and counted as disagreements. The manual classifications were then compared with the automated classifications to evaluate classification confusion and to examine the relationship between automated classifier certainty and agreement between automated and manual classifications. Based on the evaluation, a minimum certainty threshold of 0.3 was established for automated classification. When evaluating classification confusion from the test subset, test summary nodes identified as unknown by either the manual or automated method were considered mismatches. Total detection rates of each click type at each site were evaluated for the full test set.
|
10.1371/journal.pbio.1000032 | The Interscutularis Muscle Connectome | The complete connectional map (connectome) of a neural circuit is essential for understanding its structure and function. Such maps have only been obtained in Caenorhabditis elegans. As an attempt at solving mammalian circuits, we reconstructed the connectomes of six interscutularis muscles from adult transgenic mice expressing fluorescent proteins in all motor axons. The reconstruction revealed several organizational principles of the neuromuscular circuit. First, the connectomes demonstrate the anatomical basis of the graded tensions in the size principle. Second, they reveal a robust quantitative relationship between axonal caliber, length, and synapse number. Third, they permit a direct comparison of the same neuron on the left and right sides of the same vertebrate animal, and reveal significant structural variations among such neurons, which contrast with the stereotypy of identified neurons in invertebrates. Finally, the wiring length of axons is often longer than necessary, contrary to the widely held view that neural wiring length should be minimized. These results show that mammalian muscle function is implemented with a variety of wiring diagrams that share certain global features but differ substantially in anatomical form. This variability may arise from the dominant role of synaptic competition in establishing the final circuit.
| Conventionally, the organization of a neural circuit is studied by sparsely labeling its constituent neurons and pooling data from multiple samples. If significant variation exists among circuits, this approach may not answer how each neuron integrates into the circuit's functional organization. An alternative is to solve the complete wiring diagram (connectome) of each instantiation of the circuit, which would enable the identification and characterization of each neuron and its relationship with all others. We obtained six connectomes from the same muscle in adult transgenic mice expressing fluorescent protein in motor axons. Certain quantitative features were found to be common to each connectome, but the branching structure of each axon was unique, including the left and right copies of the same neuron in the same animal. We also found that axonal arbor length is often not minimized, contrary to expectation. Thus mammalian muscle function is implemented with a variety of wiring diagrams that share certain global features but differ substantially in anatomical form, even within a common genetic background.
| The nervous system's connectivity is believed to be a fundamental determinant of its function [1,2], but in general it is not readily accessible. One way to characterize neural circuits is to extract statistical properties of connectivity, often by pooling data from multiple animals [3–6]. This method assumes that connectional specificity at the level of classes of cells suffices to account for the properties of circuits [7–9]. It also assumes that within a class, each neuron's connectivity is established independently, without correlations with that of other cells. While such models may provide interesting ideas about how the nervous system works, their underlying assumptions are probably oversimplified. Neurons, for example, often innervate a nonrandom subset of cells within their target, rather than stochastically innervating such a group of cells [10,11]. In many circuits, neurons innervating the same group of cells do not establish connections independently, as evidenced by interneuronal competition observed during development [12,13]. Thus some neuroscientists have concluded that “any attempt to interpret neuronal connectivity purely in terms of probabilities … must be doomed to failure [14].” The obvious alternative is to obtain complete wiring diagrams (connectomes) of either the entire nervous system of an individual animal, or a well-defined subnetwork of the nervous system, based on direct observation rather than statistical inference. It is possible that such maps might ultimately reveal that neural circuits are stochastic in certain aspects and thus amenable to probabilistic descriptions. On the other hand, such maps may reveal organizational specificity that may not be detectable by statistical analysis, especially when the structure and connectivity of individual neurons need to be characterized in the context of the entire circuit (see below).
The first attempt to directly describe a connectome was undertaken in the parasitic nematode Ascaris lumbricoides with optical microscopy [15–17], but it produced only “enigmatic wiring diagrams” [18] because of inadequate resolution. The only successful connectomic reconstruction was accomplished in another nematode, C. elegans, using serial electron microscopy [2,18–20]. This map has proven to be a valuable resource for further analysis of circuits underlying C. elegans behaviors [21–23]. Therefore, it is likely that mammalian connectomes will also provide important information.
The advent of transgenic technologies to label neurons [24], combined with automated optical microscopy and computer-assisted image analysis tools [25], provides an avenue for the reconstruction of mammalian connectomes. Nevertheless, given the enormous complexity of mammalian nervous systems, it is necessary to begin this endeavor with tractable circuits. In this work we attempt to generate the complete wiring diagram of a peripheral neuromuscular circuit. This circuit consists of the full set of α-motor axons and the full complement of muscle fibers in the single muscle innervated by these axons. It can be captured in its entirety because each muscle's innervation is nonoverlapping. In contrast, any finite volume of circuitry in the central nervous system (CNS) contains neuronal processes entering and leaving the volume, so completeness of reconstruction cannot be achieved locally. Another advantage of the neuromuscular circuit is that its functional organization has been studied intensively, which culminated in the discovery of the size principle [26], namely, the recruitment of motor neurons proceeds in the order of increasing twitch tensions. The anatomical underpinnings of the graded tensions elicited by the group of motor neurons, however, have not been demonstrated. An additional rationale for studying the neuromuscular connectome is that the mature wiring diagram emerges from an extensive postnatal reorganization of axonal arbors known as synapse elimination. Previous imaging studies [27,28] suggested that the fate of different axons that co-innervate the same NMJ is influenced by the interactions of these axons at other NMJs with other axons. Therefore, predicting which branches are retained and which are pruned requires analyzing the competitive relationships among the entire group of neurons. In this work we took the first step of unraveling the rules of this competition by generating the adult neuromuscular connectome, which is the end product of this aforementioned process. Lastly, comparing corresponding connectomes between different animals or in the same animal (e.g., left versus right side) may help clarify the extent to which genetics, epigenetic factors, and random fluctuations impact circuit structure.
We chose to study the connectome of the mouse interscutularis, a muscle that attaches to the base of the ear and to the middle of the skull, because it is small, very thin, and is innervated by relatively few neurons. We used YFP-16, one of the few transgenic mouse lines that express fluorescent protein in 100% of motor neurons [24] to catalogue every motor unit in individual muscles (recently developed Brainbow lines do not label all motor neurons; see [29]). With confocal microscopy and semi-automated 3D reconstruction tools (Figure 1), we obtained complete connectomes of six interscutularis muscles in four mice (e.g., Figures 2 and 3, and Figure S1). Typical datasets for one muscle consisted of ∼150 image stacks, each stack containing on average ∼150 images (16 bit, 1K × 1K), totaling ∼40 GB of data. The accuracy of this reconstruction method was confirmed in three different ways (Figure 1E–1G and see Materials and Methods).
Because motor axons may branch en route to the target muscles [30], in some samples we followed each axon's trajectory far back (up to ∼0.5 cm) along the posterior auricular branch of the facial nerve (cranial nerve [CN] VII). We found that 6% (10/162) of these traced axons branched extramuscularly (Figure 2). Among these ten axons, three branched <1 mm away from where they entered the muscle, three branched 1–2 mm away, two branched 2–3 mm away, and two branched >3 mm away. None of these axons branched more than once, and all branches eventually entered the interscutularis muscle. Furthermore, we found no tendency for the axons to branch at any particular sites, such as sites where the nerve branched to supply other muscles. These data support the idea that extramuscular branching in each axon occurred independently.
The reconstructed interscutularis connectomes (Figure 3 and Figure S1) provided an atlas of neuromuscular connectional diagrams of all the axons within the muscle. This atlas included information about the number and position of all the postsynaptic targets, as well as branching topology, neighbor relations, and segmental geometry of each axon. Some of the information, such as motor unit sizes (number of neuromuscular junctions [NMJs] innervated by one axon) and statistical properties of axonal tree structures, may be obtainable by pooling single axon data from many sparsely labeled muscle samples, if homogeneity among animals and unbiased sampling are assumed. In this case, the connectomic approach provides a compendium of such information efficiently. More importantly, however, other aspects of neuromuscular circuit organization, such as neighbor relations in the fasciculation and innervation pattern, can only be understood by placing each individual axon into the context of the whole circuit's structure, and thus require the connectomic approach. Moreover, comparison between identified neurons across mice cannot be achieved by random, sparse labeling (see below).
To summarize the data: each connectome contained 14.5 ± 1.5 axons and 198 ± 11 muscle fibers. The axons exhibited a wide range of motor unit sizes, with a predominance of smaller motor units over larger ones (Figure 4A). The smallest motor units had only one NMJ (2/87 axons); the largest motor units had 37 NMJs (2/87 axons).
We found that among the 979 instances of axonal branching, most (88.5%) were binary, with progressively smaller fractions of higher degree branching (tri-furcations 10.7%, 4-furcations 0.6%, 5-furcations 0.2%, see Figure S2). We then analyzed branching symmetry of axons that innervated more than three NMJs (n = 83). This symmetry was evaluated with the imbalance index I [31], which is 0 for a completely symmetric tree (e.g., each branching point gives rise to exactly two daughter branches), and one for a completely asymmetric tree (e.g., each branching point gives rise to one terminal branch and one branch that further bifurcates). Most axons were relatively symmetric (I = 0.31 ± 0.21), and axons with larger motor unit sizes tended to be more symmetric (Spearman test, p < 0.0001).
The total intramuscular length of axonal arbors ranged from 1,583 μm to 13,320 μm (7,256 ± 2,352 μm, mean ± standard deviation [SD]), with a positive correlation with motor unit sizes. Axonal segments between branching nodes became progressively shorter with increasing branch orders (Figures 4B and S3).
We noted that the range of motor unit sizes (7.7 ± 2.8-fold) was similar to that of twitch tensions recorded in previous physiological studies of mammalian muscle contraction (e.g., 8.3-fold [32], 12.3-fold [33]). Furthermore, both the twitch tension distribution and the motor unit size distribution shared the same shape: unimodal and skewed towards the smaller end (Figure 4A). These results strongly argue that motor unit sizes are the anatomical underpinning of the observed distribution of twitch tensions. As the entire collection of motor units in each muscle was known in our dataset, we could ask whether all connectomes follow the same motor unit size distribution. Indeed we found that motor units in all six connectomes were distributed in the same way (p > 0.2, Kruskal-Wallis test). Given that motor neurons are recruited in a fixed order (weak to strong, see [26]), the correspondence between motor unit size and twitch tension mentioned above allowed us to establish the functional correspondence between individual axons in different muscle samples.
Based on conduction velocity studies, axons generating larger twitch tensions appear to possess larger calibers [33,34]. We thus anticipated that axonal cross-sectional area should correlate with motor unit size. We measured the mean cross-sectional area of each axon right before its first intramuscular branch and normalized the area to the total cross-sectional areas of all axons innervating the same muscle. We found that the normalized cross-sectional area A was correlated with the motor unit size M obeying a power law: A scaled approximately as the square root of M (Figure 4C). Furthermore, the cross-sectional area of first order axonal branches was correlated with the number (N) of downstream NMJs by a similar scaling relationship: A ∼ N0.536, (n = 47, 95% confidence interval [CI] of the exponent: 0.4663–0.6051). This similarity suggests that the scaling relationship is a fundamental property of motor axon branching.
In order to better understand the origin of this relationship, we measured the total axonal arbor length L distal to the point where the axon enters the muscle. We found that L scaled linearly with A (Figure 4D). Furthermore, L also scaled as the square root of M (Figure 4E). Taken together, these results support the idea that the principal determinant of axonal cross-sectional area is the energy cost associated with axonal membrane (see Discussion for details).
As the arrangement of different motor units in a muscle affects the mechanical properties of force delivery, we proceeded to address how motor units are deployed relative to each other in the interscutularis. The positions of NMJs in most motor units were distributed uniformly in the endplate band (Figure S4), both across the width of the muscle (80/83, 96.4%) and along the muscle's length (79/83, 95.2%). Statistical test also suggested that some motor units (18/83 across the muscle, 4/83 along the muscle) were “super-uniform,” i.e., the distribution was too regular to be from a random uniform sample. Therefore, the interscutularis muscle does not seem to possess compartments as described in certain larger muscles [35,36].
In distinction to entire motor units, primary subtrees of individual axons were not uniformly distributed. In most cases they appeared to invade nonoverlapping territories (for example, see Figure S5). We compared the distribution of subtree terminals of 27 axons in which each subtree had at least four terminals. We found that in 20 axons (74%) the distribution of terminals of the two subtrees was different (p < 0.05, generalized Wald-Wolfowitz test [37]). In particular, in 12 axons (44.4%) the territories of the two subtrees were completely segregated. On the other hand, when primary subtrees belonging to different axons were compared, their territories tended to be overlapping (78/112 pairs, 69.4%, p < 0.05, generalized Wald-Wolfowitz test). This arrangement of subtrees suggests that developmental mechanisms prevent multiple branches of the same axon from projecting to the same region, while permitting branches of different axons to intermingle in the same region. Such mechanisms may explain the observation that multiple axons innervate the same muscle fiber at early developmental stages [13], whereas rarely do two branches of the same axon innervate a single muscle fiber.
The intramuscular nerve fasciculation patterns reflect the collective behavior of all the axons. We found that the relationship between branching structures of individual axons and nerve fascicles was surprisingly complicated. Individual axons' branching behavior was not strictly coupled to the fasciculation pattern of the nerve. At some nerve branching points no axons branched; different axons simply followed one of the paths (Figure 5A). On the other hand, some axons branched inside a nerve segment and the resultant branches traveled in parallel along the same segment over some distance (Figure 5B). The most conspicuous example of such behavior was the extramuscular branching of axons discussed previously (Figure 2). Moreover, although most fasciculated nerve segments travel in a proximal-distal direction, some axonal branches contained in them did not follow the same direction. For example, in Figure 5C three axons entered the nerve fascicle from the left and two axons traveled in the opposite direction. Taken together, among 85 branching axons, 69 (81.2%) branched at least once within a nerve segment, and 29 (34.1%) contained at least one branch that traveled against the direction of some nerve segments. Overall, 89.4% of the axons deviated in some way from being a proper subgraph of the nerve fasciculation pattern.
It is possible that intramuscular nerve fasciculation reflected predetermined patterning similar to the highly stereotyped nerve structures seen more proximally (e.g., the brachial plexus). We thus tested whether there might be a conserved core fasciculation pattern in the interscutularis muscle. We assigned to each segment of the nerve a weight proportional to the total number of downstream NMJs (Figure 5D). We found that the extracted “skeletons” were topologically distinct in each muscle including left-right pairs in the same animal (Figure 5D insets). Therefore it seems unlikely that nerve fasciculation patterns in a muscle are genetically predetermined.
As mentioned above, the knowledge of motor unit sizes of all axons allowed us to identify exact neuronal counterparts in different muscle samples. This knowledge enables exploration of a question that has been investigated in invertebrates but, to our knowledge, never in terrestrial vertebrates: the degree to which an individually identified neuron shares the same branching structure with its counterparts in other samples. Although nerve fasciculation patterns differed from sample to sample as shown above, the possibility that the branching structures of axonal counterparts be identical could not be ruled out, as axonal branching structures are not necessarily subgraphs of the nerve fasciculation pattern (see section above).
In order to compare neuronal counterparts, we first determined whether there were systematic differences between left and right copies of the interscutularis muscle. The number of muscle fibers on left and right sides was not significantly different (left 201.7 ± 20.7 versus right 199.0 ± 18.0, n = 6 pairs, two-tailed p = 0.53, paired Student's t-test; Figure S6A), nor was there a difference in the distribution of muscle fiber types (type I: left 40.3%, right 42.4%; type IIA: left 20.1%, right 19.1%; type IIB+IIX: left 39.6%, right 38.5%; two pairs). The number of innervating motor neurons was not significantly different either (left 14.7 ± 1.5 versus right 14.0 ± 1.4, six pairs, two-tailed p = 0.47, paired Student's t-test; Figure S6B). We thus proceeded to identify each neuron and its counterparts based on motor unit size and/or its rank within the connectome.
We analyzed four connectomes in two animals (left-right pair for each animal). We first compared the largest motor unit with its contralateral counterpart in the same animal. In one case their sizes were similar (animal M3, left 25 NMJs [12.8% of all NMJs in the muscle] versus right 29 [14.8%], Figure S1B), but in the other case less so (animal M4, left 37 [18.8%] versus right 28 [15.2%], Figure 3B). Moreover, their appearances did not exhibit appreciable similarity upon visual inspection (e.g., Figure 3B, L1/R1 pair). We then compared smaller motor units with their contralateral counterparts and again found no appreciable similarity in the branching structures. Whether the counterpart was defined by rank order (Figures 3B and S1B) or by absolute motor unit size made no difference; in each case there was no evidence for a common branching pattern.
In order to investigate whether left-right pairs of axons with the same rank or same motor unit size are similar in less obvious ways, we focused on their topologies, ignoring geometric features (e.g., length and angle of branches). We found a wide range of different topologies between axons with the same rank (Figure 6A) and even among axons with the same motor unit size (Figure 6B). We used tree-editing distance (TED [38]) to quantify the topological difference between axons. We found that left-right pairs of axons in the same animal (intra-animal pairs) were no more similar to each other than interanimal pairs of same-sized axons (Figure 6C). Furthermore, intra-animal pairs of axons were no more similar than pairs of synthetic “axons” randomly selected from an ensemble of tree structures generated by a Monte Carlo simulation (Figure 6D) (intra-animal pair TED, 9.00 ± 2.62; Monte Carlo TED, 8.87 ± 2.50; two-tailed p = 0.89; unpaired Student's t-test). These results indicate that the topologies of intra-animal left-right pairs of axons were not correlated.
In addition to the variability of branching topology we found that axonal trajectories did not adhere to the principle of minimization of total wiring length, initially proposed by Cajal [1] and supported by the full reconstruction of the C. elegans nervous system [39–41] but see [42]. Even superficial visual inspection of the interscutularis connectome showed that almost every axon's total length could be shortened by following different nerve fascicles or altering the location of branching points. Remarkably, some axons took highly tortuous routes to their target muscle fibers even when more direct paths seemed possible (Figure 7). Moreover, ∼6% of axons branched extramuscularly as previously mentioned (Figure 2), which is also suboptimal, as the two resultant branches of the same axon invariably continued together into the same muscle. In these cases, removal of the extramuscular branching could have saved 2,072 ± 1,116 μm (n = 4 axons) of wiring length, which is equivalent to 25 ± 9.5% of the intramuscular wiring length of these axons. As the intramuscular wiring length is comparable to the distance from the cell body to the muscle, its contribution to the total metabolic cost of the cell is substantial. Therefore, the 25% extra wiring length imposes significant additional metabolic load to the cell. However, from the perspective of the neuromuscular system as a whole, this additional cost may be insignificant, since the metabolic load of muscle contraction far exceeds that of axonal conduction.
There are several reasons for mapping the connectome, i.e., the entire wiring diagram of a neural circuit. Most importantly, this map represents the complete inventory of connectional information in one particular sample. The conventional approach to circuit analysis, in contrast, infers connectivity by pooling partial data from many samples, and thus relies on assumptions such as homogeneity of neurons in a population, or stereotypy of circuits among different individuals. The connectomic approach is advantageous because it does not require such assumptions. Moreover, as every cell is identified, the way in which each cell integrates into the organization of the circuit is revealed. Lastly, comparison between different instantiations of the same circuit can reveal those aspects of connectivity that are physiologically relevant, and those that are not.
In this work we present an initial attempt to reconstruct mammalian subnetwork connectomes. We chose the mouse interscutularis muscle as the starting point because of its simplicity and accessibility. Its simplicity lies in the fact that being an end organ, it does not have strong recurrent components. This feature allows us to achieve “completeness” within a finite volume, as opposed to the situation in most CNS circuits, where recurrent connections originating from distant sites are commonplace. Furthermore, it is simple because the input is purely divergent: each axon innervates multiple muscle fibers but each muscle fiber has only one input. This pattern is only present at a few places of the CNS.
Although the interscutularis muscle represents one of the smallest possible connectomes in mammals, it still presented significant technical challenges for reconstruction. Often axonal branches were tightly fasciculated with each other, the distance between which approached the resolution limit of confocal microscopy. This problem was aggravated by scattering especially when imaging deeper structures. Therefore axonal profiles sometimes bled into each other, so computer segmentation had to be monitored and complemented by manual intervention, which significantly reduced the speed of reconstruction. Optimally, the Reconstruct program we modified for automatic segmentation traced out 4 mm of axonal length per hour. In practice, however, the requirement of human monitoring and editing reduced it to ∼0.5 mm per hour.
We anticipate that the automated imaging and semi-automated reconstruction undertaken in this work will also be generalized to the study of CNS connectomes. However, the aforementioned technical difficulties in imaging and image analysis would be greater in the CNS, where the length scale of neural structures is much smaller and the packing of neuropil is much denser. Thus future technical innovations are required to facilitate fully automated reconstruction. For example, different colors may be introduced to spectrally separate different neurons [29]; imaging resolution may be improved through super-resolution techniques [43–46]; aberrations induced by scattering in deep tissues may be overcome by serial sectioning followed by either electron microscopy [47,48] or optical microscopy [49], by adaptive optics [50], or by tissue clearing [51].
The reconstructed connectomes demonstrated four organizational principles of neuromuscular circuits. First, the motor unit size distribution in each connectome paralleled previous results from physiological recordings of twitch tensions, providing an anatomical correlate for Henneman's size principle, which until now was a physiological concept. The skewed distribution of twitch tensions, obtained by pooling data from many different samples, demonstrated that statistically most motor units generate small twitch tensions, and a few generate large twitch tensions. However, the degree to which the set of motor units within each muscle obeys the same distribution has not been directly demonstrated. Our measurement of all motor units in each muscle shows that the skewed motor unit size distribution holds for each sample.
Second, we found robust, quantitative relationships between axonal caliber, arbor length, and motor unit size. The cross-sectional area of an axon proximal to its intramuscular arborization scaled linearly with its intramuscular length (Figure 4D). Because axonal caliber is proportional to axoplasmic transport [52], it may scale with downstream metabolic expenditure. The energy expenditure is primarily devoted to resting and action potentials instead of synaptic transmission, therefore proportional to the surface area of the axonal membrane [53]. As long as the axonal caliber remains relatively constant, the surface area is proportional to arbor length. This may explain the linear relationship between axonal caliber and arbor length.
In addition, we found that the total intramuscular length of an axon scaled with the square root of its motor unit size (Figure 4E), akin to the prediction based on optimization considerations [54]. This power law scaling may be the result of the fact that average branch lengths progressively decreased as branch orders increased (Figure 4B). Therefore as motor unit size increases, the required increment in axonal arbor length is reduced. This relationship, combined with the proportionality between axonal caliber and arbor length, explains why axonal caliber scales sublinearly with motor unit size (Figure 4C).
Third, the axonal branching structure of each motor neuron was unique. We compared each axon with its functional counterparts, as defined by the size principle, in other muscles, and found substantial topological differences. Left-right pairs of corresponding neurons in the same animal showed no less variation than ipsi- or contralateral pairs from different animals. Such intra-animal variance is surprising, as each pair of neurons had identical genetic background and presumably experienced an identical environment. This result suggests that the branching pattern of these neurons was not predetermined, which contrasts strongly with the situation in invertebrates. For instance, the C. elegans connectome revealed remarkable stereotypy in the structure of the neural circuit. Worm neurons that are ontogenetic counterparts share almost identical branching patterns and connectivity both within an individual and across different animals, even though they may not be exact replicas of each other [18,19]. In annelids [55,56], insects [57–62], and crustaceans [63,64] individual neurons can also be identified, and their axonal branching patterns are stereotyped. In particular, this mammalian result contrasts with the stereotypy of neuromuscular innervation in invertebrates. For example, although there are fine structural differences in the terminal branching of axons at NMJs of any particular muscle fiber in insects, even these branches seem to have morphological regularities that are recognizable between different animals [65,66]. In mammals not only is the preterminal branching highly variable (as shown in this paper), but our experience suggests that no two NMJs look the same. Thus axonal branching in this mammalian system seems fundamentally different from that found in invertebrates.
Fourth, many axons exhibited tortuous trajectories en route to target muscle fibers, contrary to the notion that neural circuits should minimize total wiring length [67]. The layout of axonal arbors did help to minimize wiring length by preventing significant overlaps between territories of subtrees. However, other aspects of wiring, in particular extramuscular branches, wasted substantial wiring length. In contrast, in C. elegans neural wiring approximates the optimal solution fairly well [40]. The suboptimality in wiring length found in this work does not imply that the optimization principle per se is inapplicable; it rather suggests that factors other than wiring length also play a significant role. For instance, invertebrate nervous systems are under tight genetic control, and particular mutations in a single gene can lead to stereotyped alterations in neural wiring [62]. The mammalian neuromuscular system, on the other hand, may rely more strongly on activity-dependent reorganizations for each individual neural circuit to settle down on a particular wiring scheme. This strategy does not guarantee the establishment of optimal wiring, but only arrives at a solution that is functionally acceptable.
In conclusion, the interscutularis connectome reveals that in mammals, muscle function is implemented with a variety of wiring diagrams that share certain global features but differ substantially in anatomical form. Even the left and right copies of this neuromuscular circuit in the same animal exhibited significant variation. Nevertheless, the multitude of wiring diagrams exhibited no appreciable functional difference. Does this fact imply that, a posteriori, the observed variability in this system is inevitable, as there is no functional reason to impose a particular wiring diagram? We believe that this may not be the case. In general, the nervous system contains features that may have no adaptive value but tend to remain conserved [68,69]. This conservation of structure may be due to tight developmental constraints, as random changes during development may lead to dysfunction. Therefore, the rationale for the observed wiring variability may lie beyond the lack of functional significance.
This variability may result from the peculiarities of the nervous system of terrestrial vertebrates [13]. The neuromuscular circuit, for example, has a reduplicated arrangement of elements: each neuron belongs to a group of similar cells (the motor neuron pool) and projects to a population of similar postsynaptic targets (muscle fibers). At early developmental stages there is extensive fan-out (each neuron innervates a large number of muscle fibers) and fan-in (each muscle fiber is innervated by many neurons). The final circuit, however, retains only a small fraction of the initial connections—those that survived the pruning phase of synapse elimination [12,70]. This reorganization process is unidirectional (connections are lost but never regained), and the fate of an axonal branch is related to the identity of its competitors [27,28]. If a different input were eliminated from even one muscle fiber early on, there might be substantial divergence in the structure of the connectome when synapse elimination is complete. This sensitivity to developmental history may be the engine that generates diversity in neural wiring.
From this perspective, the variability is not a sign of lack of regulation, but rather indicates a different developmental strategy. Instead of genetically specifying the optimal wiring diagram for all individuals, this strategy allows a different instantiation to emerge in each case. Given the important role of interneuronal competition in the developing CNS [71–74], this strategy could well be a common theme in the entire mammalian nervous system. The value of this vertebrate innovation may be that it unfetters the structure of the nervous system from strict genetic determinism.
All animal experiments were performed according to protocols approved by Harvard University Institutional Animal Care and Use Committee (IACUC). Young adult (∼30 d old) transgenic mice of thy-1-YFP-16 line received IP injections of 0.1 ml/20g ketamine-xylazine (Ketaset, Fort Dodge Animal Health) or 0.2 ml/30 g sodium pentobarbital (64.8 mg/ml in sterile water). Once anesthetized, the animals were transcardially perfused with 4% paraformaldehyde (PFA) in 0.1 M phosphate-buffered saline (PBS [pH 7.4]). The interscutularis muscle along with a segment of its innervating nerve was removed and postfixed in 4% PFA for 30 min. Muscles were rinsed in PBS (25 °C, 30 min × 2) and mounted on slides with Vectashield mounting medium (Vector Laboratories). Mounted slides were slightly squeezed between a pair of small magnets for 12 h in order to flatten the tissue so that the distance from tissue surface to the coverslip was minimized and roughly constant.
For identification of muscle fiber type, muscles were removed as above and postfixed with 1% PFA for 7 min, frozen, and sectioned at 20 μm using a Leica Cryostat. Then sections were incubated with blocking solution (2% BSA + 1% goat serum + 0.3% triton) at 25 °C for 3 h, and incubated with monoclonal antibodies against myosin type I and 2A (mouse anti-myosin I IgG1, 1:20, Novocastra; mouse anti-myosin 2A IgG1, 1:10, Iowa Hybridoma Bank) at 4 °C for 6–8 h. After several washes in 0.1% PBS-triton, samples were incubated with secondary antibody (Alexa-488 anti-mouse IgG1, 1:1,000; Molecular Probes) for 3 h. Finally muscle sections were rinsed in PBS (25 °C, 30 min × 2) and mounted on slides with Vectashield mounting medium.
Samples were imaged using a confocal laser scanning microscope (Zeiss Pascal, Carl Zeiss) equipped with a motorized stage. We used a 63× 1.4 NA oil-immersion objective and digitally zoomed-in so that each pixel was 0.1 μm (Nyquist limit). YFP florescence was excited with a 488-nm Argon laser and detected through a band-pass emission filter of 530–600 nm. The images were oversampled by a factor of 1.5 in the Z direction (Z-step sizes = 0.2 μm), with 12 bit dynamic range. Stack montages were obtained using the motorized stage controlled by the MultiTimeZ macro (Carl Zeiss), which set up the coordinates and imaging conditions for each stack. Adjacent stacks had 10% overlap to guarantee the precision of later alignment and tracing.
Using custom-written Matlab (The MathWorks, Inc.) programs, image stacks were median-filtered and resized to 512 × 512 in XY to have cubic voxels (0.2 × 0.2 × 0.2 μm). Each stack was then digitally resampled along either x- or y-axis to generate a series of cross-sections that were approximately orthogonal to the direction of most axons. All axons were reconstructed from the series of cross-sections using custom-modified Reconstruct program (freely available from http://synapses.clm.utexas.edu/tools/reconstruct/reconstruct.stm; J. Lu, J. C. Fiala, J.W. Lichtman, unpublished data). Briefly, the modification incorporated a region-growing algorithm based on intensity threshold. The user presets the threshold and selects a point (seed) in an axon on one cross-section image. The program applies the region-growing algorithm to detect the contour of the axon on the image. Then it calculates the centroid of the contour, and propagates the centroid to the next image as the seed for the next cycle of edge-detection. The user can interrupt the progress of tracing at any time if aberrant region-growing occurs, and resets the threshold. Once all axons were traced out, they were rendered in 3D in Reconstruct and projected into 2D images (one image per axon), which were manually assembled into complete montages for each axon in Adobe Photoshop (Adobe Systems Inc.).
The 2D montage of each axon was retraced with the NeuronJ plug-in (http://www.imagescience.org/meijering/software/neuronj/ [75]) to ImageJ (http://rsb.info.nih.gov/ij/, NIH) to label and measure each axonal segment. The length and connectivity of axonal segments were transformed into a tree representation with custom-written Matlab programs.
Reconstruction accuracy was confirmed in three different ways. First, different persons independently traced a series of overlapping stacks, and the tracing results contained no gross-level discrepancy that would have led to different interpretations of the connectivity relationship between axonal branches. Second, we traced individual axons from tri-color mice using only one channel, and compared the results to the tri-color images. In the line of tri-color mouse (thy-1-KOFP × thy-1-YFP-H × thy-1-CFP-S), Kusabira-Orange fluorescent protein (KOFP) is expressed in 100% of motor axons (J. Livet and J.W. Lichtman, unpublished data); CFP and YFP are expressed each in a random subset of motor axons (24). Images were taken from all three fluorescent channels, and gray-level data from the KOFP channel only (Figure 1E) was used for reconstruction of axonal profiles. The monochromatic tracing results for doubly labeled axons (Figure 1F, yellow, KOFP + YFP; lavender, KOFP + CFP) were identical to that shown in the RGB images (Figure 1G, CFP, YFP, and KOFP were mapped to blue, green, and red, respectively). Third, we checked for abnormalities in tracing results such as axons looping back onto themselves or branches unconnected to any axon, and did not find any.
We quantified the symmetry level of axonal arbors using the imbalance index I, which is defined as I = 2 × (∑all interior nodes |TR − TL|)/(n − 1)(n − 2). Here TR and TL are the number of terminals belonging to the right and left subtrees of the branching node, respectively, and n is the total number of terminals (NMJs) in the entire axon.
Intramuscular axonal arbor length was defined as the total length downstream of a reference point common to all axons in the muscle. This reference point was chosen to be the first branching point of the axon that branched most proximally (close to cell body) in the muscle. Arbor length of axons that branched extramuscularly was defined to be the sum of arbor length of primary branches distal to the reference point. Axonal caliber was measured by dividing the volume V of an axonal segment (length ∼ 50 μm) slightly proximal to the reference point by the length of the segment. Volume was calculated as V = ∑i Ai × d, where Ai was the area of the axonal profile on the i-th image section and d is section thickness. The calculated axonal caliber was normalized to the sum of calibers of all axons entering the muscle. Relationships between axonal caliber, arbor length, and motor unit size were fitted with GraphPad Prism 5 for Windows (GraphPad Software, Inc.).
The spatial distribution of NMJs in each motor unit was parameterized along the lateral-medial axis (along muscle) and the rostral-caudal axis (across muscle). The relative position of a NMJ was defined as its rank order in the connectome along the axis. These relative positions were used to test whether a motor unit is distributed uniformly in the endplate band (Kolmogorov test [76]).
In order to test whether the two primary subtrees of an axon tend to “exclude” each other, we implemented the generalized Wald-Wolfowitz test in Matlab. For each axon, a minimal spanning tree (MST) was constructed from the distance between NMJs using Prim's algorithm [77]. Edges connecting NMJs of different subtrees were removed, and the number of resultant disjoint subgraphs was counted. Significance level p was obtained through a Monte Carlo simulation in which NMJs were reshuffled between subtrees. Two subtrees were considered completely segregated if removing one edge partitioned the MST into 2 disjoint subgraphs, each corresponding to a subtree.
TED is defined as the minimal number of operations (insertion, deletion, and relabeling of nodes) required to transform one tree into another. TED was calculated with a custom implementation in Matlab of a dynamic programming algorithm. Tree structures of axons belonging to the right-side muscles were flipped horizontally so as to compare them with the ones on the left side, with which they would overlap if there were no branching differences.
We used Monte Carlo simulation to generate a large ensemble of axons all with the same number of terminals but random topologies consistent with the data. The simulation was based on a branching process model with level-dependent branching probabilities. In particular, probabilities of axons to terminate, bifurcate, trifurcate, etc, at each branching level were calculated from the full ensemble of reconstructed axons. 50 random “axons” were generated and their pair-wise TEDs (1,225 pairs in total) were compared to that of real axons with the same number of terminals.
|
10.1371/journal.pgen.1006770 | An R2R3-type MYB transcription factor, GmMYB29, regulates isoflavone biosynthesis in soybean | Isoflavones comprise a group of secondary metabolites produced almost exclusively by plants in the legume family, including soybean [Glycine max (L.) Merr.]. They play vital roles in plant defense and have many beneficial effects on human health. Isoflavone content is a complex quantitative trait controlled by multiple genes, and the genetic mechanisms underlying isoflavone biosynthesis remain largely unknown. Via a genome-wide association study (GWAS), we identified 28 single nucleotide polymorphisms (SNPs) that are significantly associated with isoflavone concentrations in soybean. One of these 28 SNPs was located in the 5’-untranslated region (5’-UTR) of an R2R3-type MYB transcription factor, GmMYB29, and this gene was thus selected as a candidate gene for further analyses. A subcellular localization study confirmed that GmMYB29 was located in the nucleus. Transient reporter gene assays demonstrated that GmMYB29 activated the IFS2 (isoflavone synthase 2) and CHS8 (chalcone synthase 8) gene promoters. Overexpression and RNAi-mediated silencing of GmMYB29 in soybean hairy roots resulted in increased and decreased isoflavone content, respectively. Moreover, a candidate-gene association analysis revealed that 11 natural GmMYB29 polymorphisms were significantly associated with isoflavone contents, and regulation of GmMYB29 expression could partially contribute to the observed phenotypic variation. Taken together, these results provide important genetic insights into the molecular mechanisms underlying isoflavone biosynthesis in soybean.
| Isoflavones are bioactive substances with various benefits, and increasing isoflavone content is one of the major aims of soybean quality improvement. Isoflavone biosynthesis is regulated by multiple genes and complex metabolic networks. The modification of certain structural genes in the isoflavone pathway by genetic engineering has been unable to significantly improve isoflavone content. Thus, the identification and application of transcription factors specific to the isoflavone pathway may effectively resolve this problem. Here, a genome-wide association study (GWAS) was used to identify an R2R3-type MYB transcription factor, GmMYB29, associated with soybean isoflavone contents. Transient expression and plant transformation results confirmed that GmMYB29 positively regulates isoflavone biosynthesis in soybean. A candidate-gene association analysis identified 11 probable causative GmMYB29 polymorphisms. The identification and functional characterization of GmMYB29 not only improves our understanding of the genetic molecular mechanisms underlying isoflavone synthesis but also provides a direct target for both genetic engineering and selection for the improvement of isoflavone content in soybean.
| Isoflavones are a group of secondary metabolites predominantly distributed in leguminous plants, including soybean [Glycine max (L.) Merr.] [1]. In plants, isoflavones play important roles in microbial interactions, functioning as phytoalexins to protect plants from pathogen infection [2, 3]. They also act as signal molecules in the formation of nitrogen-fixing root nodules in leguminous plants [4]. For humans, isoflavones have health benefits in the prevention of several diseases, such as cancer [5], cardiovascular disease [6], and climacteric syndrome [7], which are associated with their phytoestrogenic and antioxidant properties [8]. However, isoflavones are undesirable in soy-based infant formulas [9]. In soybean breeding, an improved understanding of the mechanism of isoflavone biosynthesis would be of great value, as it may allow the manipulation of isoflavone biosynthesis and the production of cultivars that can meet various needs.
In soybean, there are three core isoflavone aglycones: daidzein, genistein, and glycitein [10]. They are synthesized via the general phenylpropanoid pathway that exists in all higher plant species and a branch of the isoflavonoid biosynthesis pathway specific to leguminous plants [11]. Isoflavone biosynthesis begins with the deamination of phenylalanine by phenylalanine ammonia lyase (PAL). After steps catalyzed by a series of enzymes, the critical branch point enzymes chalcone synthase (CHS) and isoflavone synthase (IFS) lead substrates to the isoflavone synthesis branch and finally generate isoflavones and their derivatives [12]. In addition to isoflavone biosynthesis, the phenylpropanoid pathway is also involved in the synthesis of lignins, stilbene, phlobaphenes, proanthocyanidins and anthocyanins via specific branches. Hence, the biosynthesis of isoflavones involves an intricate network reconciling many competing branch pathways. Thus, the modulation of a single gene does not necessarily alter the metabolic flux to target branch pathways [12–14].
The isoflavone biosynthesis pathway is complex, and functional differentiation is found in the isoflavone synthesis-related gene families due to two recent whole-genome duplication events: a soybean-lineage-specific duplication 13 million years ago and an early-legume duplication 59 million years ago [15, 16]. Therefore, researchers have focused on the discovery and application of isoflavone regulation-related transcription factors (TFs) instead of the manipulation of a single gene. Various TFs have been identified to regulate the biosynthesis of phenylpropane substances in higher plants, such as MYB, bZIP, WRKY, MADS box and WD40. Some MYB TFs involved in the regulation of the isoflavonoid biosynthesis pathway have been identified in soybean. For example, the R1-type MYB TF GmMYB176 has been shown to affect isoflavonoid synthesis by regulating CHS8 gene expression [1]. The R2R3-type MYB TFs GmMYB39 and GmMYB100 have been reported to negatively regulate isoflavonoid biosynthesis by suppressing the expression of structural biosynthesis genes [17, 18]. The soybean genome contains 4343 putative transcription factors, which account for 6.5% of the total predicted genes [19]. It is therefore challenging to discover and identify the key isoflavone regulation-related transcription factors at the genomic level.
Isoflavone content is a complex quantitative trait controlled by multiple genes and affected by both genetic and environmental factors. The primary mapping method for isoflavone-related quantitative trait loci (QTLs) is linkage analysis based on family lines, which is a classical method used to investigate complicated quantitative traits [19]. Previous studies have identified many QTLs controlling the biosynthesis of isoflavones in soybean seeds [20–25]. However, no isoflavone synthesis-related QTLs have been cloned due to limited allelic variation between recombinant inbred population parents. The application of genome-wide association study (GWAS), a more accurate method than linkage analysis, could enhance the power of functional gene identification [26, 27].
The development of the large genome-wide NJAU 355K SoySNP array in our previous study provides a useful tool facilitating GWAS in soybean [28]. In this study, we used this array to perform a GWAS for isoflavone contents and revealed a number of potential loci controlling isoflavone biosynthesis in soybean. We then demonstrated that one candidate gene, an R2R3-type MYB TF designated GmMYB29, played an important role in the regulation of isoflavone biosynthesis in soybean. Transcription analyses revealed a close correlation between the expression of GmMYB29 and IFS2 under normal and stressed conditions as well as between the expression of GmMYB29 and the accumulation of isoflavones. Transient reporter gene assays and overexpression of GmMYB29 in soybean hairy roots also strongly supported its key roles in the regulation of IFS2 and CHS8 expression and the isoflavone accumulation. Additionally, by combining a GmMYB29-based association analysis with an analysis of GmMYB29 expression in seed samples of 30 natural soybean varieties, we confirmed the positive regulatory role of GmMYB29 in isoflavone biosynthesis.
To determine the range of variation of isoflavone contents in soybean, the total isoflavone contents (TIC), daidzein contents (DAC), genistein contents (GEC) and glycitein contents (GLC) in soybean seeds were determined using 196 soybean accessions. To address the potential environmental influence, the soybean accessions were grown in two locations: Nanjing and Nantong (designed as two environments) (Table 1). A broad variation in isoflavone contents was observed in the population. For example, the GLC varied from 10.36 μg g-1 to 1794.00 μg g-1 in Nanjing. The average TIC, DAC, GEC and GLC was 5445.74 μg g-1, 3596.17 μg g-1, 946.73 μg g-1, and 423.85 μg g-1, respectively. The isoflavone contents showed continuous variation and normal distribution (S1 Fig), with skew and kurtosis less than one in the different environments. An analysis of variance (ANOVA) revealed that genotype and the genotype-by-environment interaction significantly influenced the major isoflavone contents (P < 0.001). This result supported the idea that isoflavone content is a complex trait controlled by multiple factors. However, the broad-sense heritability (h2) values of TIC, DAC, GEC and GLC were 74.1%, 76.3%, 67.8% and 83.8%, respectively, indicating that isoflavone content was primarily affected by genetic factors.
To identify the loci associated with isoflavone contents, GWAS was conducted using TIC, DAC, GEC, GLC and 207,608 SNPs with a minor allele frequency (MAF) > 0.05. These SNPs were obtained from the genotyping results of the 196 soybean accessions acquired using the NJAU 355K SoySNP array [28]. Twenty-eight SNPs significantly associated with the major isoflavone components were not only detected in the NJ or NT environment but also repetitively detected in the best linear unbiased prediction (BLUP) data set under a threshold of P < 4.82×10−6. These SNPs were considered as potentially reliable SNPs for further analysis (Table 2 and Fig 1). Additionally, 22 significant SNPs detected only once in the NJ environment, NT environment or the BLUP data set are presented in S1 Table. The 28 significant SNPs were located on chromosomes 5, 6, 11 and 20 and assembled into clusters on chromosomes 11 and 20. Among the significant SNPs, 17, 10, and 11 SNPs were associated with TIC, DAC, and GLC, respectively. Notably, the 10 SNPs significantly associated with DAC were overlapped with those associated with TIC. Unfortunately, no detected SNPs were significantly associated with GEC. The phenotypic variation explained by each of these significant 28 SNPs ranged from 10.20% to 14.98%, suggesting that major QTLs for isoflavone contents may exist.
Based on the linkage disequilibrium (LD) decay calculated previously, the genes within 130 kb flanking the significant SNPs were selected [28]. Among these genes, no known genes in the isoflavone biosynthesis pathway were identified. Therefore, it was speculated that there could be novel genes related to isoflavone biosynthesis or regulation in these loci. Consistently, numerous TF-encoding genes, including MYB, NAC, bZIP, and WRKY were identified (Table 2). These TF-encoding genes could function in the regulation of isoflavone biosynthesis.
Among the 28 significant SNPs, there was only one SNP detected in NJ, NT and the BLUP data set. Notably, this SNP (AX-93910416) was detected within the 5’-untranslated region (5’-UTR) of Glyma20g35180 (GmMYB29). Interestingly, the homologous gene LjMYB14 has been characterized as a TF regulating isoflavonoid biosynthesis in Lotus (Lotus japonicas) [29]. These results suggested that Glyma20g35180 could be a candidate gene controlling isoflavone biosynthesis in soybean.
The full-length open reading frame (ORF) of GmMYB29 was 819 bp and encoded a protein of 272 amino acid residues with a calculated mass of 31.15 kDa and a pI of 5.77. The GmMYB29 protein was predicted to belong to the R2R3-type MYB subfamily. A multiple alignment of GmMYB29 with R2R3-type MYB TFs known to regulate isoflavonoids or flavonoids from various plant species showed a high homology in the N-terminal MYB domain (S2 Fig). GmMYB29 was clustered with AtMYB13, AtMYB14, AtMYB15, NtMYB2, DcMYB1, LjMYB13, LjMYB14, LjMYB15, VvMYB14, and VvMYB15, and they shared a C-terminal conserved motif found in subgroup 2 of the R2R3-type MYB gene family in Arabidopsis (the SG2 motif, DxSFW-MxFWFD), which has previously been described as a stress response motif (S2 Fig) [30–32]. Consistent with this finding, in the phylogenetic tree, the proteins from various plants that were grouped in the same cluster with GmMYB29 have been reported to respond to biotic or abiotic stresses (S3 Fig) [30]. Notably, LjMYB13, LjMYB14 and LjMYB15, which regulate isoflavonoid biosynthesis in Lotus [29], were found to form a cluster with GmMYB29.
To determine the subcellular localization of the GmMYB29 protein, the GmMYB29 cDNA was fused with green fluorescent protein (GFP) under the control of the CaMV 35S promoter. This construct was then transformed into Arabidopsis mesophyll protoplasts using polyethylene glycol (PEG) and into onion epidermal cells using a gene gun. Consistent with the putative function of TFs, the GmMYB29::GFP fusion protein was localized in the nucleus, while in cells transformed with a GFP control plasmid, fluorescence was detected in both the cytoplasm and the nucleus (S4 Fig).
Previous studies have revealed that glutathione (GSH) treatment could induce isoflavonoid production [33] and that biotic stress could influence isoflavone content [34]. To determine whether the expression of GmMYB29 was induced by GSH and biotic stress and whether the expression pattern of GmMYB29 was consistent with that of isoflavone synthase 2 (IFS2), a key gene in isoflavone biosynthesis, we examined the expression of GmMYB29 and IFS2 in soybean leaves treated with GSH and common cutworms (Fig 2). After 3 h of GSH treatment, GmMYB29 and IFS2 showed 7- and 4-fold higher expression in treated leaves than in the control samples, respectively. After 7 h of treatment, these two genes showed 26- and 10-fold increases in expression, respectively. After insect-mediated induction, both GmMYB29 and IFS2 displayed marked up-regulation at 1 h and 6 h. Therefore, both GmMYB29 and IFS2 showed induced expression by GSH elicitation and insect feeding, and they were co-expressed under these two stresses. These results indicated that GmMYB29 and IFS2 could be involved in similar or the same biological processes.
To further examine the correlation between GmMYB29 and IFS2 expression in different tissues and developmental stages of soybean, we investigated the expression patterns of GmMYB29 and IFS2 and the isoflavone content in different soybean tissues (Fig 3). The expression of GmMYB29 was closely associated with that of IFS2. These two genes showed relatively higher expression in roots and seeds than in other tissues, and the expression noticeably increased with seed development. Notably, the expression of GmMYB29 and IFS2 was consistent with the isoflavone content in different tissues, suggesting that the expression of GmMYB29 and IFS2 is closely related with isoflavone accumulation in soybean.
To examine whether GmMYB29 could regulate the expression of isoflavone biosynthesis-related genes, transient expression using Arabidopsis mesophyll protoplasts and a dual luciferase reporter gene assay was performed. The promoters of two critical genes (IFS2 and CHS8) in the isoflavone biosynthesis pathway were amplified from 1790 bp and 1663 bp upstream of the start codons of IFS2 and CHS8, respectively, to study the interaction between these promoters and GmMYB29. Several MYB binding elements and stress-related cis-regulatory elements were predicted in the IFS2 and CHS8 promoters using the PLACE database (http://www.dna.affrc.go.jp/htdocs/PLACE/) [35]. As shown in Fig 4, transient expression demonstrated that overexpression of GmMYB29 increased the activity of both the IFS2 and CHS8 promoters by 100- and 200-fold, respectively, compared with the controls. These results suggested that GmMYB29 plays a critical role in the transcriptional regulation of key genes in the soybean isoflavone biosynthesis pathway.
To further identify the GmMYB29 recognition regions in the IFS2 promoter, eight fragments (IFS2ΔP1-IFS2ΔP8) were generated by gradual 5’ deletions of the promoter, which were then used to drive luciferase (LUC) expression (S5 Fig). The reporters IFS2ΔP1-IFS2ΔP8pro:LUC and IFS2fullpro:LUC were co-transfected into Arabidopsis protoplasts with 35Spro:GmMYB29, and the LUC activity was measured. The vectors containing IFS2ΔP1 and IFS2ΔP2 showed similar LUC activity to that containing IFS2full. However, the LUC activity dramatically decreased for IFS2ΔP3 and the further deletions. These results indicated that the 208 bp region between -885 and -1093 in the IFS2 promoter contained the motif required for promoter activity. In this region, a cis-regulatory element, MYBCORE (containing the CNGTTR sequence), was predicted by PLACE as a MYB binding element, suggesting that GmMYB29 could bind the IFS2 promoter and activate IFS2 expression via the recognition of this element.
To determine the role of GmMYB29 in isoflavone accumulation, overexpression and RNAi-mediated silencing of GmMYB29 were performed using a soybean hairy root system (S6 Fig). The transgenic hairy roots were verified by PCR amplification, and the positive lines were selected to conduct further studies (S7 Fig). We performed quantitative RT-PCR to study the effect of overexpression and RNAi silencing on the transcription levels of GmMYB29 and isoflavone biosynthesis-related genes, including PAL, cinnamate 4-hydroxylase (C4H), 4-coumarate coenzyme A ligase (4CL), CHS8, chalcone isomerase (CHI), chalcone reductase (CHR) and IFS2, in hairy roots obtained from several independent transgenic lines. The transcription level of GmMYB29 was significantly increased by 14- to 47-fold in GmMYB29-overexpressing transgenic hairy roots (S8A Fig) and significantly reduced by 3- to 7-fold in transgenic hairy roots with RNAi-mediated GmMYB29 silencing (S8B Fig). The GmMYB29-overexpressing transgenic roots also showed increased transcription levels of PAL, 4CL, CHS8, CHR, and IFS2, but no significant change in C4H and CHI expression was observed between overexpressing and control roots (S9A Fig). Interestingly, the transcription levels of all the monitored isoflavone biosynthesis genes were markedly decreased in GmMYB29-silenced transgenic roots (S9B Fig). Furthermore, we measured the total isoflavone contents in GmMYB29-overexpressing and GmMYB29-silenced lines by high-performance liquid chromatography (HPLC). The isoflavone content increased by 1.6- to 3.3-fold in GmMYB29-overexpressing roots (Fig 5A) and decreased by 2-fold in the gene-silenced roots (Fig 5B) (P < 0.01).
To further investigate the association between the allelic variation of GmMYB29 and isoflavone contents, we sequenced the GmMYB29 gene in a subset of 30 soybean accessions, representing varieties with high, medium and low isoflavone contents. An approximately 2.4-kb genomic region, spanning the 5'- to 3'-UTR of GmMYB29, was analyzed. A total of 12 SNPs and 11 indels (insertions and deletions) were identified and filtered out for the subsequent association analyses (Fig 6A). The association study showed that 11 probable causative sites, including Site-102 (located 102 bp upstream from the translation start codon, S-102), S-46 and S-12 in the 5'-UTR, S99 in exon1, S489 in exon2, Indel645, S679 and S1167 in intron2, S1619 in exon3 and Indel2134 and S2135 in the 3'-UTR, were significantly associated with variations in the TIC (Table 3). S-12 (corresponding to SNP AX-93910416 in our GWAS results) and S-46 were significantly correlated with the TIC, both contributing to 49.99% of the phenotypic variation in the representative subset. A single-base transversion at S1619 resulted in an amino acid substitution of lysine to asparagine at amino acid position 133, which contributed to 14.91% of the variation in TIC.
Furthermore, based on the 11 significant variants with strong LDs (Fig 6A), the 30 soybean genotypes were classified into seven haplotype classes (Hap1-Hap7) (Fig 6B). Haplotype 3 (Hap3, n = 19) is the largest group, and Hap5 (n = 6) is the second largest group, whereas the other five haplotype classes are minor groups, each comprising one soybean accession. Statistically, Hap3 accessions had significantly higher TIC than Hap5 accessions (Fig 6C). Among the different sites between Hap3 and Hap5, the most significant variants are S-46 and S-12, which are located in the 5'-UTR. Considering that the expression of GmMYB29 may cause phenotypic variation, we subsequently measured the expression of GmMYB29 in seeds from these 30 soybean accessions. The expression of this gene was positively correlated with the isoflavone content (r = 0.63, P < 0.01) (S2 Table). Additionally, we observed that the Hap3 accessions exhibited higher GmMYB29 expression than the Hap5 accessions (Fig 6D). Therefore, these data suggested that the expression of GmMYB29 could at least partially explain the phenotypic variation in isoflavone content.
The isoflavone biosynthesis network is extremely complicated, and isoflavone accumulation is dependent on pathway enzymes and interactions among enzymes [36, 37]. However, modification of a single enzyme does not significantly alter isoflavone content [12]. Thus, the identification and application of specific transcription factors in the isoflavone pathway could be an effective method to resolve this problem. In this study, we successfully cloned a GWAS-identified transcription factor (GmMYB29) that was responsible for the isoflavone contents in soybean. We then combined expression analysis, transient expression analysis, soybean hairy root transformation, and candidate-gene association analyses to confirm that GmMYB29 plays a positive regulatory role in soybean isoflavone biosynthesis. Our results reveal that an effective strategy for identifying the key QTL genes and provide a reference for cloning the rest of the isoflavone regulation-related loci in soybean as well as in other plants.
Previously, a number of QTLs associated with soybean isoflavone-related traits have been identified by linkage mapping, but few of these have been cloned or functionally characterized, perhaps because of the QTL resolution, which is limited by lower density molecular markers [38]. GWAS with high-density markers can overcome this limitation and have recently been successfully applied in studies of Arabidopsis thaliana, rice and maize [39–41]. In the present study, using a diverse natural population genotyped with high-density markers (292,053 SNPs, one SNP/3.3 kb) and phenotyped in various environments, we identified an important TF related to soybean isoflavone biosynthesis, clarified its molecular mechanism and determined the favorable alleles/haplotypes.
Studies have shown that the selection of appropriate mapping populations genotyped with saturated markers is important for performing GWAS to identify complex QTLs [42]. The 196 accessions used in this study have been reported to identify QTLs associated with seed shape, phosphorus efficiency and yield, among other features [43–45], suggesting that this panel might contain diverse genetic variations in complex quantitative soybean traits. As expected, many genetic variations in TIC, DAC, GEC and GLC were observed in the association mapping population. In addition, DAC was always the highest, followed by GEC and then GLC across various environments, which was consistent with a previous report [25]. Although isoflavone content was affected by both the genotype and the interaction between genotype and environment, isoflavone content also maintained a high heritability (0.68–0.84), which agreed with recently reported results [22–25]. These studies reveal that the heritability of isoflavone content is high enough to be considered in breeding practices to genetically improve cultivars effectively.
GWAS based on high-density SNP markers can be used to finely map quantitative trait genes, even to the genes themselves. Recently, an 82-bp (MITE) insertion in the promoter region of a NAC gene (ZmNAC111) detected by a GWAS has been determined to be associated with maize drought tolerance [46]. In our study, a highly significant SNP, AX-93910416, was identified to be associated with soybean isoflavone contents across two environments (NJ and NT) and their BLUP. A strong LD was detected in the region around this SNP (S10 Fig), indicating the existence of artificial selection and a potential target gene responsible for phenotypes in this region. More importantly, this SNP was located in the 5’-UTR of the transcription factor GmMYB29. In addition, GmMYB29 is homologous to LjMYB14, a transcription factor reported to regulate isoflavone biosynthesis in Lotus corniculatus [29], indicating that GmMYB29 is possibly involved in isoflavone regulation.
It is known that similar and conserved protein functions are derived from conserved motifs from a common origin. Comparative genomic analyses have shown that GmMYB29 not only maintains the highly conserved R2R3 domain but also has a small amino acid motif in the C-terminal region. This small motif is SG2 (DxSFW-MxFWFD), which has been reported to be related to stress resistance in plants [30–32, 47, 48]. For example, AtMYB15, which contains an SG2 motif, reportedly can improve stress resistance by increasing the expression of genes related to ABA synthesis and the ABA signaling pathway in Arabidopsis plants exposed to drought and salinity [49, 50]. Similarly, our results showed that the expression of GmMYB29 was significantly increased under abiotic and biotic stress. Interestingly, the expression pattern of GmMYB29 was similar to that of IFS2, a key structural gene in the isoflavone biosynthesis pathway, which suggests that GmMYB29 could be involved in the same regulation pathway as IFS2 [29, 30]. Furthermore, the expression profile of GmMYB29 determined by quantitative RT-PCR analysis in different tissues showed that GmMYB29 was expressed in every isoflavone-accumulating tissue [51]. However, the expression of GmMYB29 preceded IFS2 in different developmental stages of soybean seeds. For instance, the highest expression levels of GmMYB29 and IFS2 occur on the 40th and 50th day after flowering, respectively. This is consistent with the hypothesis that the expression of regulators occurs in advance of their target genes [30]. These results indicate that GmMYB29 may regulate isoflavone biosynthesis in soybean.
Transcription factors often act on the promoter region of their target genes and regulate their expression [52, 53]. As expected, we found that GmMYB29 can interact with the promoters of IFS2 and CHS8 and activate the expression of these two genes. Furthermore, co-transfection of promoter deletion fragments showed that a 208-bp fragment (from -885 bp to -1093 bp), which contains the MYB TF binding cis-acting element MYBCORE (CNGTTR), was necessary for the activation of IFS2pro:LUC, indicating the important role of this element in MYB recognition and gene transcriptional regulation. In addition, two other MYBCORE elements, located in the -1158 bp to -1790 bp and -1093 bp to -1158 bp regions of the IFS2 promoter, respectively, were identified by cis-acting element prediction software. However, gradual deletions of these two elements (generating IFS2ΔP1 and IFS2ΔP2, respectively) showed no significant effects on LUC activity (S5 Fig). The IFS2ΔP3 construct, in which all of the MYBCORE elements were deleted, exhibited almost no LUC activity, confirming that the MYB elements are the key sites recognized by GmMYB29, thereby affecting IFS2 transcription. The promoter sequences of other structural genes (PAL1, C4H, 4CL, CHS8, CHI, CHR and IFS1) in the isoflavone pathway were also investigated, and various MYBCORE elements were identified. Thus, further experiments are required to confirm whether GmMYB29 directly interacts with other isoflavone pathway-related genes.
The isoflavone biosynthesis-related R2R3-type MYB TFs reported by previous studies have generally been negative regulators in soybean [17, 18]. For example, GmMYB100 was found to inhibit isoflavonoid production by down-regulating the expression of CHS, CHI and IFS [17]. In addition to the negative regulators, the R1-type MYB TF GmMYB176, which could activate the promoter activity of CHS8, was also observed. RNAi-mediated silencing of GmMYB176 in transgenic soybean hairy roots resulted in reduced levels of isoflavonoids. Unfortunately, overexpression of GmMYB176 was not sufficient to increase CHS8 transcription and isoflavonoid levels in hairy roots [1]. In our study, overexpression of GmMYB29 increased the activity of IFS2 and CHS8 promoters; moreover, the isoflavone content was increased in GmMYB29-overexpressing hairy roots and decreased in GmMYB29-silenced hairy roots. These results imply that we identified a novel R2R3-type MYB TF, GmMYB29, which acts as a positive regulator to activate isoflavone production. Surprisingly, the level of isoflavone production in GmMYB29-overexpressing hairy roots was found to be only 3.3-fold higher at most than the control lines. One possible explanation for this observation may be the possible phytotoxic effect of isoflavonoid accumulation on the growth of hairy roots, as reflected in the relatively slow growth of GmMYB29-overexpressing transgenic hairy roots and the loss of several lines with high GmMYB29 expression.
Soybean hairy roots overexpressing GmMYB29 showed a marked increase in the expression of PAL, 4CL and CHS8 as well as CHR and IFS2 (S9A Fig). This suggests that in addition to its role in the regulation of isoflavone biosynthesis, GmMYB29 may also be involved in the regulation of upstream phenylpropanoid pathway genes to ensure the availability of substrates for isoflavone biosynthesis. It has been reported that a single TF could regulate multiple genes in the phenylpropanoid pathway and that expression of a single target gene in the pathway might be regulated by multiple TFs [54–58]. The transcriptional regulation of the anthocyanin and proanthocyanidin pathway genes is conducted by a complex in which R2R3-type MYB TFs, WD40 proteins, and bHLH proteins all interact to activate gene transcription [59–61]. Thus, we cannot exclude the possibility that there are other TFs that can activate the biosynthesis of isoflavones. Further characterization of other TFs identified in this research would provide deeper insight into the regulatory mechanisms underlying isoflavone biosynthesis.
In addition to transformation, the selection and accumulation of elite alleles of key genes functioning in isoflavone biosynthesis may be an effective strategy for soybean breeding. Similar investigations have been reported for maize, rice, soybean, and Arabidopsis, among others [46, 62–65]. For example, a sequence analysis of the drought tolerance gene ZmVPP1 in 140 inbred maize lines identified a 366-bp insertion in the promoter, which was associated with maize drought tolerance and conferred drought-inducible expression of ZmVPP1 in drought-tolerant accessions. Although some isoflavone biosynthesis-related TFs have been characterized [1, 17, 18], the polymorphism and haplotype analyses of these genes and the potential regulation mechanisms have not been reported. In this study, haplotype analysis showed that the GmMYB29 gene can be found as seven haplotypes (Hap1-Hap7), and Hap3 had higher levels of isoflavone content and GmMYB29 expression than the others, indicating that Hap3 might be significant to breeding soybeans with higher isoflavone content. Here, the 30 soybean accessions used to identify the favorable haplotype of GmMYB29 were selected to represent soybeans with different levels of isoflavone content. However, as more than 20,000 soybean accessions have been preserved [66], other elite alleles of GmMYB29 might be discovered using the stored soybean germplasms. The optimal haplotypes and alleles of this gene could therefore be detected by investigating the genetic differences in GmMYB29 expression and transcriptional activity in additional soybean accessions. Taken together, these results could lead to the development of molecular markers for the breeding of soybeans with optimized isoflavone content.
A natural population of 196 representative cultivated soybean (Glycine max) accessions with broad variations in isoflavone contents was selected from Wang et al. to develop the association mapping panel (S3 Table) [28]. These accessions (including 164 landraces, 24 improved accessions and 8 accessions with unknown evolution type) originated from all three ecological habitats of soybean in China. The seeds of each accession were provided by the Germplasm Storage of the Chinese National Center for Soybean Improvement (Nanjing Agricultural University, Nanjing, China). The trials were conducted in 2014 at two locations: Jiangpu Station of Nanjing Agricultural University in Nanjing (32°12'N, 118°37'E) (designated as environment NJ) and the Experimental Farm of Jiangsu Yanjiang Institute of Agricultural Sciences in Nantong (31°58'N, 120°53'E) (designated as environment NT). In each environment, 196 soybean accessions (corresponding to 196 plots) were planted in a randomized complete block design with three replicate blocks. Each accession was planted in four rows per plot, and each row was 200 cm long, with a row-spacing of 50 cm. The inter-plot spacing was also 50 cm. All field management requirements during the growing period, including watering, weed management, and fertilization, were performed similarly at both test locations. After maturity, four individuals from each plot were randomly screened for isoflavone content analysis.
The extraction and determination of isoflavones was performed according to the protocol described by Sun et al. [67]. First, approximately 20 g of dried seeds from each accession was ground to a fine powder using a cyclone mill. Fifteen milligrams of this powder was added to a 2 mL tube containing 1.5 mL of 80% methanol. The mixture was spun for 30 s, subjected to ultrasound treatment (frequency 40 kHz, power 300 W) for one hour at 50°C, and rotated every ten minutes. After centrifugation at 12,000 rpm for 10 min, the supernatant was filtered using a YMC Duo-filter (YMC Co., Kyoto, Japan) with 0.22 μm pores. The sample was injected into a 2 mL Agilent auto sampler and stored at -20°C before use.
Samples were analyzed with a high-performance liquid chromatography (HPLC) system (Column: Zorbax SB-C18, 5 μm, 80 Å, 4.6 mm×150 mm) under the following conditions: solvent A was 0.1% aqueous acetic acid and solvent B was 100% methanol; the solvent system was 0–2 min 27% B (v/v), 2–3 min 27%-38% B, 3–10 min 38% B, 10–12 min 38%-39% B, 12–14 min 39% B, and 14–15 min 39%-27% B. The solvent flow rate was 2 mL/min, and the UV absorption was measured at 254 nm. The column temperature was set at 36°C, and the injection volume was 10 μL.
The identification and quantification of each isoflavone component was based on the standards of 12 isoflavone components provided by Sigma-Aldrich. The 12 isoflavone standards were daidzin (D), glycitin (GL), genistin (G), daidzein (DE), glycitein (GLE), genistein (GE), malonyldaidzin (MD), malonylglycitin (MGL), malonylgenistin (MG), acetyldaidzin (AD), acetylglycitin (AGL), and acetylgenistin (AG). Different concentration gradients (0, 5, 10, 20, 50, 100, 500, 1000 ng/sampler) of the 12 isoflavone standards in 80% methanol were produced. Twelve standard curves were generated to calculate the 12 kinds of isoflavone monomer content. The precise contents of these 12 isoflavone components per gram of dry seeds (μg g-1) were calculated with the formula described in detail by Sun et al. [67]. The total isoflavone contents (TIC) were calculated as the sum of the 12 isoflavone components. The contents of daidzein (DAC), genistein (GEC) and glycitein (GLC) were the sum of four corresponding components: malonyl glycosides, acetyl glucosides, β-glycosides and aglycones.
An analysis of variance (ANOVA) of all phenotypic data was performed using PROC GLM of SAS/STAT 9.2 (SAS Institute, 2002) with environment, replication within environment, genotype and genotype × environment as random effects. The ANOVA was based on the model yijr = μ+Gi+Ej+(GE)ij+Br(j)+εijr, where yijr is the phenotype value of the ith genotype in the jth environment and the rth block, μ is the population mean; Gi is the effect of the ith genotype, Ej is the effect of the jth environment, (GE)ij is the genotype-by-environment interaction effect, and Br(j) is the effect of the rth replicate block in jth environment; and εijr is the random error. The broad-sense heritability values (h2) were estimated as h2 = σ2g/(σ2g+σ2ge/n+σ2/nr), where σ2g is the genetic variance, σ2ge is the interaction of genotype with environment, σ2 is the residual error, n was the number of environments, and r is the number of replications [68]. The best linear unbiased predictor (BLUP) for each genotype across two environments was predicted with the lme4 package in R (http://www.R-project.org/) and used as the phenotypic input in the genome-wide association study (GWAS).
The 196 soybean accessions used in this study were genotyped using the NJAU 355K SoySNP Array (S1 File). After excluding SNPs with a MAF < 0.05, 207,608 SNPs were left. The mean values of phenotypic data for all genotypes in the NJ environment and NT environment were separately used to perform the GWAS. Meanwhile, the BLUP values across these two environments were also used to perform the marker-trait association analysis. The GWAS was performed using a compressed mixed linear model (CMLM), which accounted for the complex population structure and familial relatedness [69]. For the CMLM analysis, we used the equation y = Wv+Xβ+Zu+e, where y is a vector of a phenotype; v and β are unknown fixed effects representing marker effects and population structure effects, respectively; and u is a vector for unknown random polygenic effects. W, X and Z are the incidence matrices for v, 0 and u, respectively, and e is a vector of random residual effects. All analyses were conducted with an R package called Genome Association and Prediction Integrated Tool (GAPIT) [70]. The population structure was accounted for by a principle component analysis (PCA), and the first five principal components were included in the GWAS model. The kinship matrix was calculated by the VanRaden method [71] and used as the covariance structure of random polygenic effects. The threshold for significant association was set to 1/n (n is the marker numbers, P < 4.82×10−6) [72].
The expression pattern analysis of GmMYB29 and IFS2 in response to reduced glutathione (GSH) and insect as well as in different soybean tissues was conducted in NJAU-C041 (Jianliniumaohuang), which was randomly selected from the 196 soybean accessions. The soybean seedlings used for both GSH and insect induction expression analyses were grown in growth chambers under the conditions of 16/8 h (day/night), 28/23°C (day/night), and 70% relative humidity. Soybean Jianliniumaohuang used for the tissue expression analysis was grown under natural conditions in the field at Nanjing Agricultural University.
To analyze the expression of GmMYB29 and IFS2 in response to GSH, six pieces of healthy and fully expanded upper-third leaves from different individual 30-day-old soybean plants were excised and immediately submerged in 100 mL of a GSH preparation (10 mM, pH 5.8) containing 0.005% Silwet to reduce leaf surface tension in each beaker flask (250 mL). Leaves from the same location in the control plants were also detached and submerged in a solution (pH 5.8) without GSH. Both control and treated leaves were incubated in an oven-controlled crystal oscillator at 25°C in the dark with gentle shaking (100 rpm). Samples were collected by filtration at four sampling times (0, 3, 6, and 7 h after incubation) [29, 33].
The expression of GmMYB29 and IFS2 in response to insects was analyzed by placing two third-instar common cutworm larvae of a uniform size on each fully expanded upper-third leaf of intact 30-day-old soybean seedlings in growth chambers. Control plants were not exposed to common cutworms. The damaged upper-third leaves of treated plants and undamaged leaves at the same location on control plants were excised at eight sampling times (0, 1, 2, 4, 6, 8, 12 and 24 h after induction) for the identification of induced expression [73].
To analyze the expression of GmMYB29 and IFS2 in different soybean tissues, RNA samples were isolated from roots, stems, leaves, and flowers during the full-blossom period, pod walls on 10th day after flowering (DAF), seeds at 10, 20, 25, 30, 40 and 50 DAF, and mature seeds.
The expression level of GmMYB29 was also detected at 40 DAF in the seeds of a subset of 30 soybean accessions, representing varieties with high, medium and low isoflavonoid contents from the 196 accessions.
All collected samples were placed in 2 mL cryopreservation tubes, immediately frozen in liquid nitrogen and stored at -70°C. A total of 100 mg of each sample was used for RNA isolation with the plant RNA Extract Kit (TIANGEN, Beijing, China). cDNA was synthesized using a TaKaRa Primer Script™ RT reagent kit with gDNA Eraser according to the manufacturer’s instructions. Gene expression was determined by RT-PCR assays using an ABI 7500 system (Applied Biosystems, Foster City, CA, USA) with SYBR Green Realtime Master Mix (Toyobo). The constitutively expressed gene Gmtublin (GenBank accession number: AY907703) was used as a reference gene for RT-PCR. Three replicates were performed for each reaction, and the data were analyzed using the ABI 7500 Sequence Detection System (SDS) software version 1.4.0. The normalized expression, reported as the fold change, was calculated for each sample as ΔΔCT = (CT, Target-CT, Tubulin)genotype-(CT, Target-CT, Tubulin)calibrator [74]. The primers used are listed in S4 Table.
Glycine max var. Williams 82 was the first whole-genome sequenced soybean with the most complete genome information. To obtain the accurate GmMYB29 sequence, we cloned it from this cultivar. To determine the subcellular localization of GmMYB29 in soybean, GmMYB29 was amplified from the cDNA, including the 5’- and 3’-UTRs, of Williams 82 and cloned into the pAN580 vector containing a GFP expression cassette (pAN580-GFP) to generate the recombinant plasmid pAN580-MYB29-GFP. The recombinant plasmid and the empty control plasmid pAN580-GFP were introduced into onion epidermal cells by gene gun and Arabidopsis mesophyll protoplasts by polyethylene glycol (PEG). Transgenic cells were analyzed by a laser scanning confocal microscope using a Zeiss LSM780 camera.
GmMYB29 amplified from Glycine max var. Williams 82 was inserted into the BamHI-NotI-digested GFP-removed pAN580 vector to generate the effector vector CaMV 35S::MYB29. The open reading frame (ORF) of luciferase (LUC) was cloned from the pGL3 vector (XbaI-XmaI) (Promega, Madison, WI, USA) and introduced into the GFP-loss pAN580 vector to produce the CaMV 35S::LUC plasmid. The CaMV 35S::LUC plasmid was digested by SacI and NheI to remove CaMV 35S, and then the promoter sequence of IFS2 or CHS8, amplified from Glycine max var. Williams 82, was inserted to form the reporter vectors IFS2pro::LUC and CHS8pro::LUC. A Renilla luciferase reporter was used as an internal control for normalization. The transient promoter activity in protoplasts derived from Arabidopsis suspension cells was analyzed by following the Dual Luciferase Assay protocol (Promega).
GmMYB29 was inserted into pBA002 with the CaMV 35S promoter to produce the pBA002-MYB29 overexpression vector. The RNAi vector was constructed using the Gateway technology with a Clonase II Kit (Invitrogen, Carlsbad, CA). A specific 442-bp fragment of the GmMYB29 cDNA sequence was amplified from Williams 82, and attB1 and attB2 adapters were added. Next, through the BP and LR reactions, we cloned the specific fragment into the pB7GWIWG2(II) vector to obtain the pBI-MYB29Ri vector. As the soybean cultivar Jack is known for its high transformation efficiency, the soybean hairy root transformation was performed using this accession with the pBA002-MYB29 overexpression vector, the pBI-MYB29Ri vector, and the respective empty vectors as controls. The positive hairy roots detected by PCR from several independent transgenic lines were harvested separately and used for gene expression or isoflavone content analysis.
Using the 5'- and 3'-UTR sequences of GmMYB29, which shared relatively low sequence similarity with a paralogous gene, a pair of gene-specific primers (GmMYB29-SF and GmMYB29-SR) were designed (Prime 5.0) to amplify GmMYB29 from 30 soybean genotypes (S3 and S4 Tables). The primers used to sequence the GmMYB29 PCR products are listed in S4 Table. The sequences were assembled and aligned using ContigExpress in Vector NTI Advance 10 (Invitrogen, Carlsbad, CA) and MEGA version 6 [75], respectively. Polymorphisms, including SNPs and indels, with a MAF ≥ 0.05 were identified among these genotypes, and their association with isoflavone content and pairwise LDs were calculated by Tassel 5.0 [27, 76]. Markers were defined as being significantly associated with the phenotype on the basis of a significant association threshold (-LogP > 1.30, P < 0.05).
|
10.1371/journal.pcbi.1000076 | An Allosteric Mechanism for Switching between Parallel Tracks in Mammalian Sulfur Metabolism | Methionine (Met) is an essential amino acid that is needed for the synthesis of S-adenosylmethionine (AdoMet), the major biological methylating agent. Methionine used for AdoMet synthesis can be replenished via remethylation of homocysteine. Alternatively, homocysteine can be converted to cysteine via the transsulfuration pathway. Aberrations in methionine metabolism are associated with a number of complex diseases, including cancer, anemia, and neurodegenerative diseases. The concentration of methionine in blood and in organs is tightly regulated. Liver plays a key role in buffering blood methionine levels, and an interesting feature of its metabolism is that parallel tracks exist for the synthesis and utilization of AdoMet. To elucidate the molecular mechanism that controls metabolic fluxes in liver methionine metabolism, we have studied the dependencies of AdoMet concentration and methionine consumption rate on methionine concentration in native murine hepatocytes at physiologically relevant concentrations (40–400 µM). We find that both [AdoMet] and methionine consumption rates do not change gradually with an increase in [Met] but rise sharply (∼10-fold) in the narrow Met interval from 50 to 100 µM. Analysis of our experimental data using a mathematical model reveals that the sharp increase in [AdoMet] and the methionine consumption rate observed within the trigger zone are associated with metabolic switching from methionine conservation to disposal, regulated allosterically by switching between parallel pathways. This regulatory switch is triggered by [Met] and provides a mechanism for stabilization of methionine levels in blood over wide variations in dietary methionine intake.
| Methionine is an essential amino acid that is highly toxic at elevated levels, and the liver is primarily responsible for buffering its concentration in circulation. Intracellularly, methionine is needed for the synthesis of S-adenosylmethionine (AdoMet), the major biological methylating agent. Methionine used for AdoMet synthesis can be replenished via remethylation of homocysteine. Alternatively, homocysteine can be converted to cysteine via the transsulfuration pathway. A specific feature of liver methionine metabolism is the existence of twin pathways for AdoMet synthesis and degradation. In this study, we analyzed the dependence of methionine metabolism on methionine concentration in liver cells using a combined experimental and theoretical approach. We find a sharp transition in rat hepatocyte metabolism from methionine conservation to a disposal mode with an increase in methionine concentration above its physiological range. Mathematical modeling reveals that this transition is afforded by an allosteric mechanism for switching between parallel metabolic pathways. This study demonstrates a novel mechanism of trigger behavior in biological systems by which the substrate for the metabolic network switches metabolic flux between parallel tracks for sustaining two different metabolic modes.
| Methionine, an essential amino acid, plays a significant role in intracellular one-carbon and sulfur metabolism consequently linking antioxidant and methylation homeostasis (Figure 1). Since methionine is involved in many intracellular processes its levels in blood and in different organs need to be tightly regulated. In fact, methionine levels in blood and liver do not change over a several-fold variation in dietary methionine intake [1]–[3]. Aberrations in methionine, and therefore in methylation and antioxidant metabolism, are associated with a number of complex diseases, including cancer, anemia, neurodegenerative diseases, and developmental abnormalities [4],[5]. Hence, elucidating the multiple switches that regulate methionine metabolism is key to understanding their dysregulation. Such information may also permit intervention that would reverse disease states.
Liver is considered to be the main organ that removes excess methionine from the organism and regulates the methionine level in blood. The metabolism of methionine in liver is more complex than in other organs; it is characterized by a unique, liver-specific enzyme profile (Figure 1), which includes the methionine adenosyltransferase I (MATI) and methionine adenosyltransferase III (MATIII) isoforms of methionine adenosyltransferase. Methionine adenosyltransferase II (MATII) is present only in extrahepatic tissues. The enzymatic reaction of the MATIII isoform exhibits a rate dependence on AdoMet concentration that is the converse of MATI and MATII. Thus, while MATI, like MATII, is inhibited by AdoMet, MATIII is activated at high concentrations of AdoMet [6],[7]. Liver also possesses glycine-N-methyltransferase (GNMT), betaine homocysteine S-methyltransferase (BHMT), and cystathionine β-synthase (CBS) at levels that are much higher than those found in other organs and tissues [8]. Thus, an interesting feature of liver methionine metabolism is that there are parallel tracks for the synthesis and utilization of AdoMet, and at the homocysteine junction, methionine can either be recycled or committed to the synthesis of cysteine, which can then be either utilized biosynthetically or catabolized (Figure 1).
Despite the large number of studies on methionine metabolism in liver [8] its regulation as well as the kinetics of the methionine disposal, are not well understood. Furthermore, the dependence of liver methionine metabolism on fluctuating methionine concentration ([Met]) within a physiologically relevant range has not been investigated. To elucidate the molecular mechanism that controls metabolic fluxes via parallel tracks in liver methionine metabolism, we had developed a simple mathematical model [9], which revealed the possibility of a threshold dependence of liver methionine metabolism on [Met]. This simple qualitative model predicted the existence of two modes in liver methionine metabolism characterized by low metabolic rates and metabolite levels at methionine concentrations equal to or below its normal physiological value and by high metabolic rates and metabolite concentrations at methionine concentrations above its physiological value. The model predicted that under a specific set of conditions, methionine metabolism switches sharply from one mode to another when [Met] slightly exceeds its physiological value. The switch from the “low” to “high” mode is associated with a sharp increase in steady-state AdoMet concentrations. It was presumed that this metabolic switch enables conservation at low methionine levels and disposal at high methionine levels, and is associated with a redistribution of metabolic flux between remethylation and transsulfuration. However, a detailed analysis of the metabolic switch could not be undertaken with the simple model because it did not include descriptions of remethylation and transsulfuration fluxes in an explicit form. Moreover, the simple model had significant limitations such as the loss of stability at methionine concentrations exceeding the normal physiological value by ∼10% and unrealistically slow kinetics for transitional processes [9], raising questions about its utility for experimental verification at the high end of physiologically relevant methionine concentrations and on a real time scale.
In this study, we have analyzed the experimental dependencies of [AdoMet] and methionine consumption rate on [Met] in native murine hepatocytes within a physiologically relevant concentration range (40 to 400 µM). We have found that both [AdoMet] and methionine consumption rate do not respond gradually with increasing [Met] but rise sharply, increasing ∼10-fold within the narrow [Met] concentration interval from 50 to 100 µM in accordance with the predictions of our preliminary mathematical modeling of this pathway [9]. We used our experimental results for construction and quantitative verification of the extended mathematical model of liver methionine metabolism, which included a detailed description of the kinetics of all the relevant enzymes, a simplified description of folate metabolism and a new rate equation for the MATIII reaction. The new rate equation for the MATIII reaction was developed to describe the experimental data at methionine concentrations above its normal physiological value. Analysis of the model reveals that the sharp increase in [AdoMet] and methionine consumption rate observed in hepatocytes at increasing [Met] are associated with a sharp transition in methionine metabolism from a conservation to a disposal mode (i.e. from remethylation to transsulfuration). The transition is controlled by allosteric regulation of enzymes that results in switching metabolic flux between parallel pathways for AdoMet synthesis and utilization. This metabolic switch is triggered by methionine and provides a mechanism for stabilization of methionine levels in blood over wide variations in dietary methionine intake.
Our study reveals an excellent correspondence between the experimental and predicted behavior of liver cells in response to variations in [Met] and demonstrates the existence of a novel allosteric mechanism for switching metabolic fluxes between parallel pathways.
In a suspension of murine hepatocytes incubated with 40 µM methionine, the intracellular concentration of AdoMet was quite stable over 3 h. The average value of [AdoMet] obtained under these conditions in 24 independent experiments was 79±37 µmol/l cells (mean±SD). At higher methionine concentrations, the intracellular [AdoMet] increased within 20–30 min and then reached a plateau within ∼1 h, representing a new steady-state level, which did not change further during the following 1–2 h (Figure 2A). As the initial [Met] in the suspension was increased, the steady-state AdoMet level increased monotonically. However, the major increase in [AdoMet] was observed at an initial [Met] above 200 µM. When the initial [Met] was 200–240 µM or 400 µM, the intracellular AdoMet level reached values of 590±180 (n = 9) and 930±350 µmol/l cells (n = 12) respectively, exceeding the concentration obtained at 40 µM methionine by 11- to 15-fold, on average.
Under control conditions, i.e., at 40 µM methionine, the average intracellular concentration of S-adenosylhomocysteine (AdoHcy), obtained in 7 independent experiments was 7.1±7.0 µmol/l cells. The concentration of AdoHcy also increased with the increase in initial [Met] in the cell suspension, reaching a value of 76±21 µmol/l cells at an initial [Met] of 400 µM. This exceeded the concentration obtained at 40 µM methionine, on average, by 10-fold (not shown).
The concentration of methionine in the incubation medium decreased with time, due to methionine consumption by hepatocytes (Figure 2B). The dependence of [Met] on time was approximated by linear kinetics, and the slope of the corresponding line normalized to cell volume was taken as the rate of methionine consumption by hepatocytes. The rate of methionine consumption increased with an increase in the initial [Met]. At 40 µM methionine, the average rate of methionine consumption, obtained in 15 independent experiments, was 0.86±0.40 mmol/h·l cells. The rate of methionine consumption by hepatocytes increased to 6.0±2.4 mmol/h·l cells at an initial [Met] of 400 µM (n = 12), exceeding its value at 40 µM methionine by 8.1±3.6-fold. Our mathematical model, which is described in the Methods section and in Supporting Text S1 and S2, provides a satisfactory description of the [AdoMet] and [Met] kinetics observed experimentally in hepatocytes in response to changes in [Met] (Figure 2C, D).
To assess the experimental dependence of steady-state AdoMet concentration on methionine, we plotted the average intracellular [AdoMet] measured between 60 and 160 min of incubation versus the average [Met] measured during the same time interval. The AdoMet concentration obtained in each experiment was normalized to the maximum AdoMet levels obtained at an initial [Met] of 400 µM with the same batch of cells. Data from 12 independent experiments are presented in Figure 3A. As can be seen, a distinctive feature of the dependence of [AdoMet] on [Met] in mouse hepatocytes is the sharp increase in [AdoMet] within a narrow range of [Met] between 50 and 100 µM. Below and above this trigger zone, the steady-state [AdoMet] increases only slightly with an increase in [Met]. The relatively low slope of the experimentally observed dependence of [AdoMet] on [Met] at low methionine concentrations (Figure 3A, inset) is incompatible with a simple hyperbolic dependence.
The quantitative mathematical model developed in the present work exhibits a sharp increase in [AdoMet] within a narrow concentration range of methionine above its physiological level (Figure 3A), in good agreement with the experimental results. We have similarly examined the dependence of steady-state AdoHcy concentration on methionine (Figure 4). This dependence is similar to the dependence of [AdoMet] on [Met] (Figure 3A) and is well described by the model (Figure 4).
The experimentally derived rate for methionine consumption in hepatocytes was plotted versus the average [Met] measured between 60 and 160 min, as described above for AdoMet (Figure 3B). The values obtained for the methionine consumption rate in each experiment were normalized to the rate obtained with the same batch of cells at an initial [Met] of 400 µM. As seen for the behavior of the intracellular AdoMet concentration, the methionine consumption rate increased several-fold in a relatively narrow concentration interval of methionine and then did not change so significantly. The experimental data are in good agreement with the prediction of the mathematical model (Figure 3B). The relatively low slope at low [Met] is inconsistent with the methionine consumption rate being a simple hyperbolic function of [Met] (Figure 3B, inset).
The extended model supports analysis of specific reaction rates and metabolites in the sharp transition between low and high [AdoMet] and methionine consumption rate within a narrow [Met] range. The model reveals that at physiological [Met] (∼50 µM), almost all the AdoMet is produced by the MATI isoform and consumed by cellular methylases, thereby meeting the demands for methylation reactions (Figure 5A). Under these conditions, the futile metabolic flux through GNMT is very low because [AdoMet] is low and high levels of 5-methyltetrahydrofolate (MTHF) inhibit GNMT (Figure 1, Figure 5C). Note that the steady-state rate of AdoMet production (VMATI+VMATIII) (for complete definition of terms, see Methods) is not equal to the net rate of methionine consumption (VMATI+VMATIII−VMS−VBHMT) because a significant amount of AdoMet is produced due to recirculation of metabolites via methionine synthase (MS) and betaine homocysteine S-methyltransferase (BHMT) (Figure 5). A small increase in [Met] above its normal value leads to a moderate increase in [AdoMet]. Due to the opposing influence of AdoMet on MATI and MATIII isoforms (Figure 1), the rate of MATI decreases, while the rate of MATIII increases and the total rate of AdoMet production increases slightly. At a threshold value for methionine (∼70 µM in the model) the rate of the MATIII reaction exceeds the rate of the MATI reaction and the rate of AdoMet production exceeds the total activity of cellular methylases. This leads to a rapid and auto-accelerated accumulation of AdoMet and an increase in the rate of the MATIII reaction due to positive feedback regulation of MATIII by AdoMet, which cannot be compensated by a further decrease in the MATI reaction rate. Also, increased AdoMet levels inhibit methylenetetrahydrofolate reductase (MTHFR) that leads to a significant decrease in [MTHF]. Consequently, GNMT is strongly activated due to its sigmoidal dependence on [AdoMet] and decreased inhibition by MTHF under these conditions (Figure 1, Figure 5). The sharp activation of GNMT results in consumption of excess AdoMet produced by activated MATIII providing a new steady state of methionine metabolism characterized by high [AdoMet] and a high metabolic rate. This regulation provides the metabolic switch in methionine metabolism from the “low” to the “high” metabolic mode. Thus, above the threshold concentration for methionine, the rate of AdoMet synthesis increases several-fold, due mainly to the activity of MATIII. The contribution of MATI to total AdoMet synthesis decreases to <10% and excess AdoMet is metabolized via GNMT. Below the threshold [Met], a large fraction of the methionine used in AdoMet synthesis is regenerated via transmethylation reactions catalyzed by MS and BHMT, and the net rate of methionine consumption is low (Figure 5, Figure 3B, theoretical curve). Above the threshold, the net rate of methionine consumption increases dramatically (Figure 3B, theoretical curve), and methionine is primarily catabolized via the transsulfuration pathway (Figure 5B). The increased AdoMet concentration inhibits MTHFR that leads to a sharp decrease in the MS substrate, MTHF. This in turn, leads to a decrease in the MS reaction rate despite the increased level of the other substrate, homocysteine (Hcy). Meanwhile, the committing enzyme in the transsulfuration pathway, CBS, is activated due to an increase in the concentration of its substrate, homocysteine and its allosteric activator, AdoMet. These results are in good agreement with the published experimental data demonstrating the decrease in the fraction of methionine produced via remethylation and activation of methionine catabolism via the transsulfuration pathway at high [Met] in liver or hepatocytes [2],[10],[11].
The metabolic switch between methionine conservation and disposal buffers methionine levels in liver and provides a mechanism for stabilizing liver and blood methionine concentrations over a wide range of dietary methionine intake. Indeed, rapid normalization of blood and liver methionine is observed after food consumption [12] or methionine injection [1],[13]. Blood and liver methionine levels do not vary significantly over a several-fold range in dietary methionine intake [1]–[3]. To demonstrate methionine stabilization in the extended model, we analyzed the dependence of steady-state methionine metabolism on the rate of methionine influx. Figure 6A shows the computed dependence of the steady-state [Met] on the normalized rate of methionine influx into hepatocytes. As one can see, [Met] is stabilized over a ∼6-fold increase in the rate of methionine influx. The rate of the CBS reaction increases proportionally with the increase in the rate of methionine influx while the rate of transmethylation, contributed collectively by MS and BHMT, does not change significantly decreasing only at very high values of methionine influx (Figure 6B). This indicates that stabilization of [Met] is provided via catabolism by adjusting the flux through the transsulfuration pathway.
The methionine metabolic pathway represents a useful paradigm for studying gene-nutrient interactions since it is richly dependent on the input of dietary factors (viz. amino acids and B-vitamins). Moreover, several genetic polymorphisms have been described in the pathway's enzymes that are correlated with risk for various complex diseases [14]–[17]. An integrated understanding of the regulation of methionine metabolism is important because of its critical role in modulating the two major homeostatic systems governing methylation and antioxidant metabolism, which are often dysregulated in complex diseases. Attaining this understanding is, however, a challenge because methionine metabolism is complex; the pathway has branches and cycles in addition to parallel fluxes at several steps of intermediate transformation. Further, regulation of the pathway enzymes by intermediates increases the complexity of the system. Mathematical modeling is a powerful tool for studying the mechanisms of regulation in complex metabolic systems and for analyzing the behavior of a system under different natural and experimentally induced conditions. Previously, simple mathematical models of liver methionine metabolism have been developed and used to analyze and predict the response of the system to changes in [Met] and other variables [9],[18],[19].
The most striking prediction of our previously published simple mathematical model describing methionine metabolism is that a substrate (methionine) can induce the sharp transition between two modes in liver methionine metabolism characterized by low metabolic rates and metabolite levels at methionine concentrations equal to or below its normal physiological value and by high metabolic rates and metabolite concentrations at methionine concentrations above its physiological value [9]. The model predicted that this transition is triggered within a narrow methionine concentration range. In addition to the limitations of the simple model discussed in the introduction, we could not verify or adjust the model parameters to obtain a quantitative description of methionine metabolism because of the scant information available on the dependence of metabolic fluxes on [Met] in liver or in hepatocytes. Even the rate of methionine consumption under normal physiological conditions was unknown.
To address these gaps, we experimentally determined the responsiveness of methionine metabolism in murine hepatocytes to varying concentrations of methionine. Our study revealed a sharp dependence of [AdoMet] and the rate of methionine consumption in hepatocytes on [Met] within a narrow range of extracellular [Met] that lies just above the physiological concentration of circulating methionine (∼50 µM). These data supported the development of an extended quantitative model of liver methionine metabolism, which also includes rate equations for MS, BHMT and CBS, and a simplified treatment of folate metabolism. Importantly, without the link between methionine and folate metabolism the extended model failed to provide a realistic description of the GNMT reaction rate, confirming the experimentally established regulation of GNMT by the folate derivative, MTHF. Enzymatic reaction rates were described using mechanism-based equations and the kinetics parameters were adjusted from ranges reported in the literature and our own experimental data. In addition, a new equation had to be developed to describe the MATIII reaction, including both activation and inhibition of the enzyme at high [AdoMet], in order to describe the experimental data at [Met] above its normal physiological value. A significant attribute of the extended model is that it provides a good quantitative description of our experimental results, including steady-state values and transitional processes over a wide range of [Met].
The extended model was used to analyze two types of experiments. To assess the response of hepatocytes ex vivo, to the parameter, i.e., varying [Met], the behavior of pathway metabolites and reaction rates were analyzed. To describe in vivo conditions, [Met] was considered as a model variable and the rate of methionine intake was considered as an independent parameter. A good agreement was observed between the simulated and observed ex vivo experimental data. Our experimental results demonstrate that a sharp increase in [AdoMet] in hepatocytes is accompanied by a several-fold increase in the methionine consumption rate (Figure 3). This behavior is reversible and liver metabolism switches back to a conservation mode when the [Met] decreases below the threshold level. Our analysis of the in vivo response to varying intake of methionine reveals that the metabolic switch between methionine conservation and disposal provides a mechanism for stabilizing liver and blood methionine concentrations over a wide range of dietary methionine intake. This mechanism is associated with the regulation of the CBS reaction rate (Figure 6) in addition to regulation by MATI/III and GNMT. Homeostasis requires that the steady-state transsulfuration flux must be equal to or less than the net rate of methionine consumption. The normal rate of methionine consumption observed in murine hepatocytes, was ∼1 mmol/h·l cells, which represents ∼1.3% of murine liver CBS activity under maximal velocity conditions [20]. In the model, under physiological conditions, the rate of the CBS reaction is ∼0.6% of CBS activity (Figure 5B and Table S1). Thus, our experimental and theoretical data reveal that under normal physiological conditions, liver CBS works at a very small fraction of its maximal capacity, and that it can be activated several fold at high [Met] to increase the volume of sulfur flowing through the transsulfuration pathway.
Thus, liver methionine metabolism exhibits trigger behavior that results from allosterically regulated switching between two sets of enzymes that catalyze parallel metabolic fluxes. In the methionine conservation mode, the metabolic flux is determined by one set of enzymes, which includes MATI and functional methylases, while in the methionine disposal mode, the metabolic flux is determined by a second set of enzymes that includes MATIII and GNMT (Figure 7). To our knowledge, such a mechanism of trigger behavior that causes metabolic flux to switch between parallel pathways has not been reported previously in other metabolic systems.
A sharp transition from one steady state to another in the background of a monotonic change in one parameter (or stimulus) is observed in many biological processes including ontogenesis, cell cycle and signal transduction. An ideal mechanism for trigger behavior is provided in a system that exhibits bistability, i.e., when an area of instability separates two stable steady states. In this case, the system can jump from one steady-state to the other, and stable intermediate states do not exist in the pathway. There are a number of theoretical and experimental reports of bistability in different biological processes [21]–[25]. In fact, within a fairly wide range of parameters, both the simple and extended models predict the existence of bistability in the system, leading to the trigger behavior in metabolite concentrations and metabolic fluxes, including [AdoMet] that jumps from its physiological concentration to significantly higher levels when the [Met] crosses a threshold value. This bistability results in a hypersensitivity of methionine metabolism to changes in methionine levels within a narrow concentration interval. The theoretical mechanism of bistability, which can be realized in liver methionine metabolism, is described in Text S3 and Figures S3, S4, and S5.
However, bistability is not absolutely essential for effective regulation of liver methionine metabolism. Our model shows that even sharp monotonic dependence of methionine metabolism on methionine in a narrow range of [Met] produces metabolic switching between methionine conservation and disposal modes, providing a mechanism for effective stabilization of methionine levels. The results demonstrate that both [AdoMet] and the rate of methionine consumption in hepatocytes increase slowly with an increase in methionine level at low [Met] (0–50 µM) and then increase sharply in a narrow concentration range of 50–100 µM (Figure 3) in accordance with model predictions. While the experimental data cannot distinguish between a jump versus a monotonic increase in [AdoMet] and the rate of methionine consumption with an increase in [Met] they are not well described by a simple hyperbolic behavior (Figure 8). Formal approximations of the initial part of our experimental points at [Met] from 0 to 100 µM with the simple power function: Y = A+B*Xn, provide the best fit at n = 2.30 for AdoMet and at n = 1.43 for methionine consumption rate. This is consistent with a sharp, non hyperbolic increase in [AdoMet] and methionine consumption rate in a narrow interval of [Met], confirming our model predictions. Similar power function approximation of regular Michaelis-Menten hyperbola Y = X/(100+X) gives n = 0.64.
The sharp change in methionine metabolism observed in hepatocytes at increasing methionine concentration is in striking contrast to the relatively weak dependence we have previously observed in hepatoma cells [18]. A notable difference between primary versus transformed hepatocytes is that the latter express the MATII isoenzyme instead of MATI/MATIII; they also lack GNMT [26]–[28]. Thus, our theoretical and experimental results demonstrate that the presence of the MATI/MATIII and GNMT is crucial for normal regulation of liver methionine metabolism and for achieving a normal balance between methionine conservation and disposal, thereby buffering methionine levels.
We note that our mathematical model describing methionine metabolism is still relatively simple. For instance, it does not include factors that can potentially affect liver and blood methionine levels, such as the modulation of enzyme levels by dietary factors [8] and methionine turnover in organs other than liver. Nevertheless, the predictive power of the model in revealing an unexpected mode of substrate-triggered regulation is validated by the results reported in this study, affirming the utility of a combined theoretical and experimental approach for the study of metabolic regulation.
Our mathematical model is a system of ordinary differential equations that describe the kinetics of extracellular and intracellular concentrations of methionine as well as intracellular concentrations of other intermediates of methionine metabolism (Figure 1).(1)
Here [Met]ext and [Met] are the extracellular and intracellular methionine concentrations, Vinflux is the rate of methionine influx into the system, Vtrans is the rate of methionine transport into hepatocytes, whep and wmed are total volumes of hepatocytes and external medium, respectively, Vprot is the net rate of methionine consumption in protein turnover, VMATI, VMATIII, VGNMT, VMeth, VAHC, VCBS, VMS, VBHMT are the rates of reactions, catalyzed by MATI, MATIII, GNMT, functional methylases, adenosylhomocysteinase (AHC), CBS, MS, and BHMT respectively.
The reversible AHC-catalyzed reaction is assumed to be at equilibrium because the activity of AHC is much higher than the activity of other enzymes in the metabolic system (see Table S1). The rate of methionine consumption in protein turnover and the total rate of functional methylases are described by simplified equations. Equations for rates of other enzymatic reactions are based on known kinetics mechanisms for the respective enzymes. We developed a new equation for the MATIII reaction in order to describe the experimental data at [Met] above its normal physiological value. All equations for the enzymatic reaction rates are presented in Text S1. The development of the equation for MATIII is described in Text S2 and Figures S1 and S2.
We do not take into account AdoMet consumption for polyamine synthesis because its rate is about 1% of flux in methionine metabolism [29] in normal liver.
The distribution of methionine between hepatocytes and incubation medium can be evaluated using the following published data [30]–[33]. It was shown using radioactive methionine distribution that methionine transport in hepatocytes is fast, passive, reversible and that the equilibrium between intracellular and extracellular methionine is established within a few minutes (Figures 3–5 in [33]). Moreover, these data demonstrate a linear fit between intracellular and extracellular methionine concentrations in the range from 18 µM to 10 mM (Figure 5 in [33]). In hepatocytes, methionine is transported via a Na+-dependent and a Na+-independent transporter, with the latter accounting for ∼90% of the transport at normal methionine concentrations, and the activity of transport is high enough to provide equilibration of extracellular and intracellular methionine within a few minutes (Table 2 in [32]), which confirms the data published in [33]. The kinetics of methionine transport in hepatocytes shows that at physiologically relevant concentration of methionine (100 µM) and at 37°C the equilibrium is achieved in 1 min (Figure 3 in [30]). The distribution of methionine between hepatocytes and the medium can be estimated using the data presented in Figure 3 of [30] as follows: at 100 µM [Met] in the medium, the equilibrium [Met] inside hepatocytes was 600 pmol/106 cells. Based on the size of hepatocytes (see below in Metabolite Analysis), 106 cells are estimated to have a total volume of 10 µl. This yields an intracellular [Met] of 60 µmol/l cells. Recalculating per volume of intracellular water, the intracellular [Met] is estimated to be ∼80 µM yielding a ratio between intra- and extra-cellular [Met] of 0.8. Similar calculations for the data presented in Figure 5 of [33] reveal that the amino acid remains equally distributed between the extracellular medium and inside cells at methionine concentrations ranging from 18 µM to 10 mM. It has also been shown in vivo that [Met] in blood plasma and in normal liver are fairly similar (Table 2 in [31]). Additionally, we note that preliminary experiments based on direct methionine measurements confirm fast and uniform distribution of methionine between hepatocytes and the incubation medium in concentration range from 40 to 400 µM. So we assume that extracellular and intracellular methionine concentrations are equal.
To incorporate the kinetics of folate metabolites into the model, specifically 5,10-methylenetetrahydrofolate (5,10-CH2-THF) and MTHF, which regulate MS, GNMT, and MTHFR activities, we take into account that most folate derivatives are interconnected via highly active and reversible enzymatic reactions [34],[35]. This allows us to consider a general folate pool with equilibrium ratios between its components for all folates except dihydrofolate (DHF) and MTHF. The last two metabolites are produced from 5,10-CH2-THF in irreversible reactions with low activity [36],[37]. Because of the equilibrium between components of folate pool we assume that concentration of 5,10-CH2-THF always constitutes 20% of the total pool as seen under normal physiological conditions [38],[39]. The total intracellular concentration of all folates is assumed to be constant, and DHF is not taken into account because we assume that its concentration is constant. Texts S1, S2, and S3 and Tables S1, S2, and S3 contain a detailed description of the model and its analysis.
The experimental protocol was approved by the Scientific Council of the National Research Center for Hematology (NRCH). Female CBF1 mice (20–22 g) were obtained from the laboratory animal nursery in Stolbovaya, Moscow region, Russia. They were housed in a vivarium at the NRCH and provided with food and water ad libitum. Hepatocytes were isolated under sodium thiopental narcosis (i.p. injection of 5 mg/mouse). Liver was perfused for 5–10 min through the vena porta with a solution containing 115 mM NaCl, 5 mM KCl, 1 mM KH2PO4, 0.5 mM EGTA, 10 mM glucose, and 25 mM HEPES buffer, pH 7.5, saturated with 95% O2 and 5% CO2 at 37°C, followed by 7–8 min perfusion with the same solution lacking EGTA and containing 2.5 mM CaCl2 and 20–30 µg/ml of a collagenase-protease mixture (Liberase Blendzyme 3, Roche). The rate of perfusion was 7 ml/min. Then, the liver was dispersed at room temperature in the initial perfusing solution lacking EGTA. Hepatocytes were filtered through a 50 µm nylon mesh, sedimented by centrifugation and washed 3 times in 45 ml dispersing solution supplemented with 1 mM CaCl2, 1 mM MgCl2, 40 µM methionine and 2% BSA. Cell concentration and viability were determined using a hemocytometer after staining with 0.4% trypan blue. Cell preparations with viability ≥85% were used in all experiments. Hepatocytes were resuspended in the washing solution to a concentration of 1·106 viable cells/ml, then 5 ml suspension aliquots were placed in 50 ml Ehrlenmeyer flasks and agitated at the rate of 100 RPM under an atmosphere of humidified 95% O2 and 5% CO2 at 37°C. At the beginning of the incubation, methionine was added to flasks to a final concentration of 40–400 µM. Cells prepared from one animal were incubated simultaneously in several flasks at different initial methionine concentrations for ∼3 h for a single experiment. Typically, the initial methionine concentrations were 40 µM, 400 µM, and two intermediate values.
At the desired time intervals aliquots of cell suspension were collected from flasks and mixed with 0.2 volumes of 30% trichloroacetic acid. After centrifugation the supernatant was used for analysis of AdoMet, AdoHcy and methionine [18],[30].
The concentration of AdoMet and AdoHcy were measured by HPLC using an 8 µm Chromasil-100 C18 column, 250×4.6 mm, (Elsico, Russia) under isocratic conditions at a flow rate of 1 ml/min with monitoring at 254 nm. The mobile phase contained 40 mM NaH2PO4, 6 mM heptanesulfonic acid sodium salt (Sigma), and 15% methanol, pH 4. AdoHcy and AdoMet eluted as single peaks with retention times of 7 and 13 min respectively. The concentrations of AdoMet and AdoHcy in the samples were determined using calibration curves generated for each compound and normalized to the cell count in the suspension (i.e., expressed as µmoles/l cells). Based on an estimated cell diameter of 25–30 µm, the calculated value for the cell volume is 1·10−11 l.
To measure methionine concentration, the supernatant obtained after protein precipitation was derivatized with 2,4-dinitrofluorobenzene [30] and analyzed by HPLC using a 5 µm Diaspher-110 ODS column, 250×4 mm (BioChemMack, Russia) under isocratic conditions at a flow rate of 1 ml/min with monitoring at 355 nm. The mobile phase contained 45% acetonitrile in 1% acetic acid in water. The retention time of derivatized methionine was 12 min. The concentration of methionine in the samples was determined using a calibration curve and expressed as µmoles/l of suspension. The change in methionine concentration, due to intracellular metabolism in cell suspension over time was approximated by linear kinetics. The slope of the corresponding line and its deviation were determined using the Microcal Origin 6.0 software (Microcal Software, Inc.). To determine the rate of methionine consumption in hepatocytes, the slope was normalized to the total cell volume in suspension as described above.
As it was mentioned above, methionine transport in hepatocytes is passive, reversible and very fast over a wide range of [Met] (from 18 µM to 10 mM) and equilibrium between extracellular and intracellular methionine concentrations can be established within 1 min [30],[32],[33]. The equilibrium extracellular and intracellular methionine concentrations are similar in hepatocyte suspensions and in normal liver [30],[31],[33]. Thus, under our experimental conditions the total methionine concentration in cell suspension represents both medium and intracellular methionine concentration, and the rate of decrease in methionine concentration in suspension represents the rate of metabolic methionine consumption (catabolism) by the cells.
The accession numbers of enzymes used in the model are defined in Text S1. |
10.1371/journal.pgen.1006900 | Replication stress affects the fidelity of nucleosome-mediated epigenetic inheritance | The fidelity of epigenetic inheritance or, the precision by which epigenetic information is passed along, is an essential parameter for measuring the effectiveness of the process. How the precision of the process is achieved or modulated, however, remains largely elusive. We have performed quantitative measurement of epigenetic fidelity, using position effect variegation (PEV) in Schizosaccharomyces pombe as readout, to explore whether replication perturbation affects nucleosome-mediated epigenetic inheritance. We show that replication stresses, due to either hydroxyurea treatment or various forms of genetic lesions of the replication machinery, reduce the inheritance accuracy of CENP-A/Cnp1 nucleosome positioning within centromere. Mechanistically, we demonstrate that excessive formation of single-stranded DNA, a common molecular abnormality under these conditions, might have correlation with the reduction in fidelity of centromeric chromatin duplication. Furthermore, we show that replication stress broadly changes chromatin structure at various loci in the genome, such as telomere heterochromatin expanding and mating type locus heterochromatin spreading out of the boundaries. Interestingly, the levels of inheritable expanding at sub-telomeric heterochromatin regions are highly variable among independent cell populations. Finally, we show that HU treatment of the multi-cellular organisms C. elegans and D. melanogaster affects epigenetically programmed development and PEV, illustrating the evolutionary conservation of the phenomenon. Replication stress, in addition to its demonstrated role in genetic instability, promotes variable epigenetic instability throughout the epigenome.
| In this study, we found replication stresses reduce the fidelity of nucleosome-mediated epigenetic inheritance. Using Position Effect Variegation (PEV) in centromere as an indicator of chromatin epigenetic stability, we quantified the precision of nucleosomal inheritance and found replication stresses reduce the fidelity of nucleosome-mediated epigenetic inheritance. Further analysis of genome-wide heterochromatin distribution showed that replication stresses affect chromatin structure by expanding of heterochromatin with locus specificity. Mechanistically, we provide evidence suggesting that excessive formation of single-stranded DNA might have correlation with the reduction in fidelity of centromeric chromatin duplication. Finally, we demonstrated replication stress perturb the development process by reducing the fidelity of chromatin organization duplication in fruit fly and worm, illustrating the broadness and the evolutionary conservation of the phenomenon. Together, our results shed light on the importance of replication stresses cause epigenetic instability in addition to genetic stability.
| In eukaryotic cells, genomic DNA are packaged into arrays of nucleosomes [1], each comprised of a 147bp DNA fragment wrapped around a histone octamer core. The combination of histone variants and the large repertoire of covalent modifications on histones result in a highly complex biochemical signature of the nucleosome, which encodes important epigenetic information [2,3]. Overall, the nucleosomal organization of chromatin–including the positions of nucleosomes relative to the underlying DNA sequence and the biochemical signatures that they carry–has a profound impact on the functional state of the genome. In order to maintain the identity of the cell, nucleosomal organization must be preserved through cell divisions. On the other hand, it is conceivable that controlled alteration in cell type, such as cell differentiation during development, would require nucleosomal organization amendable for reprogramming. Despite its profound biological significance, the mechanisms on regulating or influencing the precision of chromosomal epigenetic inheritance are not well understood. In this study, we examine the effect of replication perturbation on the fidelity of chromatin duplication and epigenetic inheritance and explore the underlying mechanisms.
During cell division, chromatin is duplicated in conjunction with DNA synthesis at the replication fork, through a process called replication coupled (RC) nucleosome assembly [4,5]. The process can be divided into three major steps. First, pre-existing nucleosomes (also known as parental nucleosomes) immediately in front of the replication fork are disrupted so that the template DNA is accessible to the replication machinery. Second, shortly after replication fork passage, the core components of the parental nucleosomes–the (H3-H4)2 tetramers in specific–are recycled to assemble nucleosomes on one of the daughter strands behind the replication fork. Finally, newly synthesized histones are incorporated on the other daughter strand to form nucleosomes de novo. Although the two daughter strands can be distinguished based on whether they originate from the leading or lagging strand in replication, in general, there appears to be no strand-specificity for histone recycling and de novo nucleosome assembly.
To achieve precision in duplication of epigenetic markers on histones, the recycled parental H3-H4 molecules need to be incorporated at their original loci on one of the daughter strands. Furthermore, the nucleosomes assembled de novo on the other strand need to be positioned at the corresponding sites. The newly incorporated histone molecules should also be of the proper variant type and obtain the biochemical modifications matching that of the parental histones. Currently, little is known on how the precision is achieved for any of the above steps. An important aspect of chromatin duplication is that the parental H3-H4 histones are transferred from the template chromatin to the replicated strands as intact (H3-H4)2 tetramers [6–8]. An alternative mode of inheritance–splitting the (H3-H4)2 tetramer into two dimers and passing them equally to the two replicated strands–may occur in the minor sub-population of nucleosomes that contain the H3.3 variant but not in the majority that contain the canonical H3.1 variant [8].
The DNA replication machinery is directly implicated in RC-nucleosome assembly by interacting with, and potentially coordinating the actions of, histones, histone chaperones, histone modifying enzymes, and chromatin remodeling factors[9,10]. Among the replication proteins, the helicase MCM2-7 is thought to play critical roles in evicting the nucleosomes ahead of the replication fork as well as assembling nucleosomes behind the fork. MCM2-7 forms a stable complex with the histone chaperone Asf1 that is bridged by a histone H3-H4 dimer [11,12]. The Asf1/H3-H4/MCM complex may represent an intermediate for parental histone recycling or new nucleosome assembly. However, it remains unclear how each (H3-H4)2 tetramer is integrated into the proper site on one of the daughter strands. In the budding yeast S. cerevisiae, genome-wide tracking of the parental histone H3 molecules through several generations and quantitative modeling of experimental data lead to a model that the parental (H3-H4)2 tetramers are re-incorporated within the distance of one to two nucleosomes (~400bp) of the original site. Thus, nucleosomal inheritance may be somewhat “sloppy” [13]. Such sloppiness, if confirmed, would preclude single or small number of nucleosomes as efficient carrier of epigenetic information. Further experimental evidence is needed to directly test this model.
We used Position Effect Variegation (PEV) as an indicator of chromatin epigenetic stability to quantify the precision of nucleosomal inheritance. PEV, referring to variable expression patterns in a gene due to its translocation to a specific position in the genome, was broadly observed [14,15]. PEV phenomenon was originally discovered in specific fruit fly strains, associated with the white gene translocation adjacent to centromeric heterochromatin. There, the variegated expression states of the translocated white gene propagate clonally in adult eyes, causing mosaic eye coloration patterns [14]. In fission yeast, heterochromatin spreading is also responsible for PEV associated with the mating type locus. Within the centromeric core region, PEV is also observed but due to a different mechanism [16,17]. Here, the positioning of nucleosomes containing the specific histone H3 variant CENP-A/Cnp1 is variable within the centromere, and Cnp1 occupancy inversely correlates with the expression levels of the underlying reporter genes.
Regardless of the biochemical characteristics of local chromatin region responsible for underlying gene silencing, one common property of PEV is that the variegated gene expression states are inherited in a clonal fashion. In both budding yeast and fission yeast, inheritance of variegated gene expression is vividly demonstrated by sectoring in yeast colony coloration [16,18,19] and by tracking the gene expression status through the cell generations directly at single cell level [16]. Changes in the coloration patterns of the colonies serves as a convenient indicator of changes in the epigenetic marker underlying gene silencing.
In addition to using PEV as readout for epigenetic stability at specific loci, we wish to assess the impact of replication stresses at the whole epi-genome level by determining the genome-wide heterochromatin distribution in cells under stresses. Finally, we sought to explore the possible evolutionary conservation and the physiological significance of reduced epigenetic fidelity due to replication stress by testing the effects of perturbing replication on the development process in fly and worm.
In fission yeast, variegated expression of ade6+ inserted in the centromere (cnt2::ade6+) is readily visualized: the ON or OFF states of ade6+ correspond to white or red color, respectively, of the colonies grown with low supply of adenine. We previously have shown that cnt2::ade6+ expression inversely correlates with Cnp1 occupancy on ade6+. Furthermore, by tracking the colony coloration through cell lineages, we have demonstrated that the state of ade6+expression, and thereby, Cnp1 occupancy on ade6+, is inherited through cell generations, but can change abruptly within one generation at low rates [16].
We explored the possible association between the fidelity of centromeric Cnp1 nucleosome position inheritance and the progression of DNA replication. Hydroxyurea (HU) is broadly used to study DNA damage-independent replication fork arrest [20–22]. HU is an inhibitor of ribonucleotide reductase (RNR), the enzyme responsible for the synthesis of dNTPs. Depletion of dNTP pools through HU treatment leads to replication fork arrest and subsequent genomic instability[23,24]. We first tested whether HU treatment enhances the rate of switching in the expression state of cnt2::ade6+. Yeast cells originated from a red colony (i.e., cnt2::ade6+ silenced, with minor white sectors) were treated transiently (for four to six generation times) with low concentrations of HU, prior to plating on the media for characterization of the progeny colonies coloration. Consistent with the epigenetic inheritance of the cnt2::ade6+ silent status, in the absence of HU, most of the progeny colonies were red (with minor white sectors). However, with increasing concentrations of HU, more white (with or without minor red sectors) colonies were formed (Fig 1A, upper). Similarly, HU treatment of cells that originated from a white colony also gave rise to increased switching in colony coloration, albeit from white to red (Fig 1A, lower).
We previously have established a pedigree analysis assay to quantify the rate of switching in cnt2::ade6+ expression states per cell division [16]. Using this assay, we measured the switching rate of cnt2::ade6+ (OFF to ON) in wild type cells is 6.5%, and is increased to 8.5% in wild type cells treated with 1mM HU (Fig 1B lower panel). These results show that HU treatment enhances switching in cnt2::ade6+ expression status, suggesting that replication perturbation due to HU treatment (which inhibits ribonucleotide reductase and causes depletion of dNTP pools) could reduce the fidelity of centromeric epigenetic inheritance. To test whether depletion of dNTP pools by genetic perturbation would have the same effect, we also measured the CEN-PEV switching rate in cdc22-3, a mutation in the large subunit of ribonucleotide reductase[25]. The result shows that the switching rate of cnt2::ade6+ (OFF to ON) in cdc22-3 at 25°C is increased to 9.7% comparing with 6.5% in wild type cells (Fig 1B lower panel), suggesting that replication perturbation caused by suboptimal dNTP levels reduces the fidelity of centromeric epigenetic inheritance.
Likewise, colonies formed at the constant presence of HU exhibit higher degrees of sectoring than those formed in the absence of HU. This reflects enhanced, continually ongoing switching in cnt2::ade6+ expression state throughout the time course of colony formation (Fig 1B). Noticeably, colonies that exhibited high degree sectoring when grown in the presence of HU, once re-plated on media without HU, reverted to wild type degree sectoring (Fig 1C). This suggests that change in the degree of sectoring directly correlates with HU treatment and that such changes are not genetic. To further confirm this notion, we examined the ade6 gene in four of these red colonies by PCR amplification and DNA sequencing and found no mutation. Together, these results indicate that HU-induced perturbation of DNA replication promotes switching in cnt2::ade6+ expression status, and once switched, the expression states are inherited.
HU treatment disturbs replication progression as well as RC nucleosome assembly. Specifically, continuing unwinding template DNA combined with pausing in DNA synthesis leads to excessive formation of ssDNA on template and concurrent accumulation of parental H3-H4 histones evicted from template chromatin [26]. ssDNA activates the S phase checkpoint, which in turn, halts the MCM helicase, preventing further ssDNA formation and histone eviction [27,28]. When replication resumes, incorporation of the accumulated histones onto the daughter strands would be disordered due to the loss of their original positioning on the template chromatin [26], thus contributing to the enhanced variegation in the centromere.
According to this model, there would be a correlation between excessive ssDNA formation and the reduction in the fidelity of nucleosome inheritance. In particular, mutations that cause prolonged unwinding of the template DNA and excessive formation of ssDNA should also reduce the fidelity of nucleosome inheritance. We chose two mutants to test this prediction: deletion of the S phase checkpoint gene cds1 (rad53/chk2)—cds1-D and a c-terminal truncation of the Mcm4 subunit of the MCM helicase -mcm4-84c. cds1-D inactivates the replication checkpoint [29,30]; whereas mcm4-84c renders the MCM helicase unresponsive to inhibition by an activated checkpoint without apparent compromising of its helicase activity [28]. Both mutants are hypersensitive to HU (S1 Fig), and are shown to form ssDNA excessively upon replication pausing as evidenced by enhanced chromatin association of single strand binding protein, Ssb2/Rfa2 [28,31]. We confirmed this by quantifying the Ssb2-GFP foci signal in S phase cells (recognized by the presence of a medial septum–a morphological signature of S phase cells). The Ssb2-GFP foci signal is significantly increased in both cds1-D and mcm4-84c mutants compared with wild type at the same HU doses (Fig 2A). In order to test whether the increased Ssb2-GFP signal results from the change in the Ssb2-GFP protein level, we have performed the western blotting assay, and found the similar protein level between wild type and mutant cells with or without HU treatment (S2 Fig).
We then compared the rate of switching in cnt2::ade6+ expression states upon HU treatment in wild type and mutant cells. In wild type cells, HU treatment enhances switching rate in cnt2::ade6+ expression status using the pedigree analysis assay [16]. Meanwhile, we measured significant increases in the rate of switching in cds1-D cells (9.1% without HU treatment to 10.1% treated with 0.5mM HU) and mcm4-84c cells (13.5% without HU treatment to 14.9% treated with 1mM HU) respectively (Fig 2B). Both mutants also exhibited higher rates of switching than wild type cells under all tested conditions. Consistently, we also found that mutant colonies exhibited more complex sectoring patterns with HU treatment compared to no HU treatment, and much more complex sectoring patterns in comparison to wild type at all conditions (S1 Fig). Together, these results suggest that excessive ssDNA formation caused by HU treatment is correlated with the reduction in the fidelity of epigenetic inheritance.
Conversely, if a mutation causes DNA replication stalling without excessive formation of ssDNA, it should not affect the fidelity of nucleosome inheritance. MCM helicase unwinds the template DNA and is postulated to drive the eviction of the parental histones. We thus reasoned that perturbation of the MCM helicase function might cause replication perturbation without causing excessive unwinding of template DNA or accumulation of parental histones. To test this, we examined the effect of a temperature sensitive mutation in the MCM helicase (mcm4-M68ts) that conditionally disrupts replication initiation [32]. At a semi-permissive temperature (29°C), the biological activity of MCM is compromised so that the survival of mcm4-M68 cells is strictly dependent on the Chk1-dependent DNA damage checkpoint that is non-essential in wild type cells. Interestingly, the Cds1p-dependent intra-S phase checkpoint is not activated or required for cell survival under this stress condition [33]. Indeed, we found only insignificant level of Ssb2-GFP foci signal in S phase mcm4-M68 cells at 29°C comparable to wild type (Fig 3A), but observed a significant cell cycle delay (elongated cell morphology) and reduction in cell viability (S3 Fig), confirming that MCM activity was compromised. Consistent with minor increase in ssDNA formation, the rate of switching in cnt2::ade6+ expression state increased insignificant in mcm4-M68 cells compared to wild type at 29°C (Fig 3B). Mutant colony morphology also exhibited wild type level complexity in sectoring (S1 Fig).
Together, these results suggest a correlation between excessive ssDNA formation and the enhanced switching of centromeric PEV, indicating a reduction of fidelity in Cnp1 nucleosome position inheritance.
Given that HU treatment enhancing centromeric PEV correlates with excessive ssDNA formation, we sought to test whether other replication stresses that cause increased ssDNA formation should also cause enhanced centromeric PEV. We tested genetic perturbations in three distinct complexes of the replication machinery: deletion of ctf8, a non-essential subunit of the Ctf18-RFC clamp-loader complex [34]; and deletion of ssb3, a non-essential subunit of the ssDNA binding protein complex RPA [31], partial inactivation of Psf1, a subunit of the GINS complex essential for replication initiation and elongation [35]. Conditional inactivation of Psf1, an essential protein, is achieved by fusing Psf1 to a steroid hormone-binding domain (HBD) tag that is tightly associated with the protein chaperone Hsp90, rendering the fusion protein inactive by steric hindrance. The HBD fusion protein can be kept active by the addition of β-estradiol, which binds to HBD and displaces Hsp90. psf1-HBD cells depend on the presence of β-estradiol for viability [36].
Microscopic examination of S phase cells reveal increased Ssb2-GFP foci signal in all three mutants compared to wild type (Fig 3A). Noticeably, psf1-HBD cells with a high level of β-estradiol, which fully supported cell viability (“proficient” condition), still exhibit an elevated switching rate (10.4% in comparison to 6.5% in wild type) and enhanced ssDNA levels, indicating that the HBD tag alone may quantitatively disturb the GINS complex function. Reducing the level of β-estradiol (“deficient” condition) further exacerbate the defects (the switching rate increases to 13.9%). Consistently, quantification of the switching rate (cnt2::ade6+ OFF to ON) by pedigree analysis show higher switching rate in all mutants compared to wild type cells (Fig 3B), mutant colony morphology also exhibits more complex sectoring patterns (S1 Fig).
We further wish to test whether reduced epigenetic inheritance stability caused by replication stress is not only reflected by the enhanced switching rate of cnt2::ade6+ OFF to ON, but also the changed switching rate of cnt2::ade6+ ON to OFF. Consistently, the switching rate of cnt2::ade6+ in wild type cells is 6.5% (OFF to ON) and 3.4% (ON to OFF), and increased to 8.5% (OFF to ON) and 5.6% (ON to OFF) with 1mM HU treatment. And we also quantified the two switching rates of cnt2::ade6+ simultaneously in two replication mutants, ssb3D and ctf8D. While significant increases in switching rate of cnt2::ade6+OFF to ON were detected in both mutants (9.8% in ssb3D and 10.6% in ctf8D), no increase or slight reduction in ON to OFF rate was found (2.55% in ssb3D and 3% in ctf8D, respectively). We are unclear about the discrepancy between the measurements of the two switching rates. One possible reason may be that, the pedigree analysis is less suited to capture rare switching events (ON to OFF) quantitatively.
All together, in all three mutants tested above in which various parts of the replication machinery is perturbed genetically and with excessive accumulation of ssDNA, centromeric PEV is enhanced.
We further reasoned that replication stresses might affect the inheritance of other chromatin features in addition to Cnp1/CENP-A nucleosome occupancy within the centromeres. To test this idea, we examined two PEV systems associated with the mating type region caused by stochastic heterochromatin (histone H3K9me2/3 modifications) spreading.
The mating type region of fission yeast contains three gene loci, among which only mat1 is actively transcribed and determines the cell mating type. mat2 and mat3 act as genetic information donors for a gene conversion process at mat1 that causes mating type switching [37]. mat2-mat3 region, in wild type cells, is silenced tightly by heterochromatin formation via histone H3 lysine 9 methylation [38,39]. Furthermore, histone hypoacetylation also contributes to its silencing [40]. ade6+ inserted in the silencing domain between the boundary and mat2 (L(BglII)::ade6+) exhibits the typical variegated expression pattern [17]. Alternatively, a cis-DNA element–cenH–within the mating type region is sufficient to initiate heterochromatin formation at an ectopic site in the genome. ade6+ juxtaposed to cenH at an ectopic site (ura4::cenH-ade6+) also exhibits variegated expression [39].
We tested replication stress on PEV of L(BglII)::ade6+and ura4::cenH-ade6+, and found that HU increased ade6+ silencing in a dosage—dependent manner for both reporter constructs (Fig 4). Regardless cells originated from red or white colonies, when treated with HU, ade6+ silencing state was promptly established and persisted, resulting in red colonies with little or no white sectors. Between these two reporter constructs, L(BglII)::ade6+ exhibited a dramatic change, resulting in predominant or nearly all red colonies with HU treatment (Fig 4A–4C). In comparison, ura4::cenH-ade6+ exhibited a moderate but clear, unilateral increase in ade6+ silencing (Fig 4D–4F, the number of red colony is increased upon HU pulse treatment from 85% to 93% (start from a red colony), and the number of white colony is decreased from 97% to 90% (start from a white colony)). Such unilateral switching to silencing state is in sharp contrast to the observation in centromeric PEV using the same test, which resulted in colonies with enhanced bi-lateral switching. Upon re-plating to HU-absence media, white colonies re-emerged from L(BglII)::ade6+and ura4::cenH-ade6+ red colonies, suggesting that ade6+ silencing is due to epigenetic instead of genetic changes. To further confirm this notion, we examined the ade6 gene in four of these red colonies by PCR amplification and DNA sequencing and found no mutation (S4 Fig).
Studies in the budding yeast have raised the concern that in certain experimental settings, reporter genes (URA3 and ADE2) may exhibit gene-specific transcription responses, rendering them unsuitable for characterizing the heterochromatin-induced silencing effects [41,42] (also see comments in [43]). To assess whether or not enhanced switching in ade6+ expression status in our experiments is reporter gene-specific, we tested the effect of replication stress on a native gene, mat2-P, at its endogenous locus. Mutations in specific genes (e.g., clr1, a zinc finger protein gene) partially compromise transcriptional silencing at the mating type region [37]. The leaky expression of mat2-P in stable M cells (Mat1-Mmst0—a genetic modification at mat1 that locks the cell in the h- mating type) leads to haploid meiosis and sporulation, producing aneuploid, non-viable spores. Spore formation is detected by iodine vapor staining of the colonies or by microscopic examination of the cells. Importantly, iodine vapor staining reveals sectoring patterns, indicating that the silencing/leaky expression states of mat2-P are clonally inherited [44,45].
We examined mat2-P leaky expression in clr1-deletion (clr1-D) [46,47] colonies under replication stress conditions. The number of iodine staining patches and the intensity of staining diminished in an HU dose-dependent manner. Quantification of haploid meiosis (H.M. phenotype) within the colonies confirms a reduction in meiosis upon HU treatment (S5 Fig). h-/h+ diploid colonies under the same conditions are stained strongly by iodine vapor, suggesting that the low level of HU treatment used in this study does not inhibit meiosis or sporulation per se. These results suggest that low level HU treatment suppresses the leaky expression of mat2-P in clr1-D cells. To verify that this silencing effect is caused by heterochromatin on mat2-P, we further tested this notion by anti-H3K9me2 ChIP, and found heterochromatin is compromised in clr1-D strain (S5 Fig). HU treatment enhances H3K9me2 enrichment at mat2-P, demonstrating that heterochromatin underlies the gene silencing here. Thus, replication stress perturbs the inheritance of the expression states of a native gene similar to that of ade6+ reporter at the mating type region, arguing against the possibility of a gene-specific response to HU treatment.
In all, these results show that, unlike centromeric PEV in which the variation is enhanced, replication stress stimulates enhanced silencing unilaterally on two independent PEV systems mediated by H3K9me2/3, and the phenomenon is not reporter gene-specific.
Net enhanced silencing of heterochromatin-associated PEV by replication stress at multiple loci in the genome may be explained by a possible mechanism that the heterochromatin domains are expanded. Consistent with this hypothesis, Singh and Klar previously have shown that cdc22-3 causes heterochromatin silencing and H3K9 methylation spreading at the silent mat locus[25]. To test whether such effect is broadly seen throughout the genome, we compared heterochromatin distribution on whole genome-wide in wild type cells with or without HU treatment and cdc22-3 mutant cells by ChIP-Seq. In specific, chromatin immunoprecipitation was performed in these cells using an antibody against histone H3 dimethylated at lysine 9 (H3K9me2). And the immunoprecipitated DNA was then subjected to high throughput sequencing to determine the specific location and relative abundance of H3K9me2 throughout the genome. The result showed there is no appreciable difference in wild type cells with or without HU treatment. However, comparing between wild type and cdc22-3 mutant cells, significant difference can be seen at mating type locus and sub-telomeric regions (Fig 5A and 5B. Please see below for more in-detailed analysis of the ChIP-seq results). We are unclear why short-term HU treatment in wild type cells didn’t cause significant changes in heterochromatin distribution whereas genetic perturbation of cdc22-3 mutation did. We speculate that this may be because cells within a culture treated with HU temporarily are highly heterogeneous in terms of epigenome perturbation. Any changes in epigenome at the specific locus in a small percentage of cells may not be detected readily by ChIP-seq.
Throughout the fission yeast genome, two types of heterochromatin have been found–constitutive heterochromatin domains and facultative heterochromatin islands [48,49]. The former are strong, persistent heterochromatin domains, including peri-centromeric, sub-telomeric regions and the mating-type locus. The latter encompasses ~30 loci, the majority of which are meiosis genes [49].
In wild type cells, we have detected H3K9me2 mainly at the constitutive heterochromatin domains, identical to previous reports [49] (S6 Fig). We have also detected a number of heterochromatic islands with relatively low levels of H3K9me2, most in agreement with previous reports. A few heterochromatic islands are different from those reported in recent studies [49,50]. The discrepancies might be caused by the difference in sensitivity of the experimental tools (ChIP-chip vs ChIP-Seq), or different data processing methods.
In cdc22-3 mutant, we detected alterations in some of the heterochromatin domains in comparison to wild type (Fig 5A and 5B). Specifically, H3K9me2 enrichment at the sub-telomeric regions is expanded by 5-20kb. And the heterochromatin at the mating type locus spreads beyond the normal boundaries to the neighboring genes, just as described in Singh and Klar’s work[25].
To confirm the heterochromatin expansion, we tested the effect of HU on mating type locus using reporter genes inserted outside the wild type mating type boundaries [25]. Consistently, we have observed enhanced reporter gene silencing under HU treatment (S7 Fig, ura4 gene silencing cells exhibit resistance at 5-FOA plate), similar to what was previously reported [25].
Expansion of the heterochromatin domains is specific to the regions described above. No change was detected in the peri-centromeric regions of any chromosomes. This suggests that alteration in heterochromatin expansion induced by replication stress is locus-specific.
Noticeably, we observed that the detected alterations of heterochromatin in mutant cells were variable among biological repeat samples (Fig 5C), whereas the positions of heterochromatin domains are highly consistent among wild type biological repeats (S6 Fig). In mutant Sample 1 (cdc22-3_1), the heterochromatin of Tel1 L and Tel2 L was shorter than the parallel biological repeats (cdc22-3_2 and cdc22-3_3), whereas the heterochromatin of Tel2 R, Tel3 L, and Tel3 R was longer than the parallel biological repeats (Fig 5C). The variation among the samples is unlikely due to technical reasons, as such variation is seen only in the sub-telomeric regions, while other regions of the genome are highly consistent.
Such sample-specific chromatin structure changes may indicate they are sporadic events among genetically identical cells/cultures and are epigenetically inheritable. To validate this notion, we examined the variation of heterochromatin spreading using ChIP-PCR among seven independent yeast colonies derived from the same ancestor cells (Fig 5D). Using primers to amplify a fragment nearby tel2R (SPBPB2B2.08, with low level of H3K9me2 enrichment in wild type cells) along with primers set that amplified a fragment of tel2R (with high level of H3K9me2 enrichment in wild type cells) as a control, we found that three to ten folds enrichment of SPBPB2B2.08 gene fragment among different samples, whereas constant levels of mei4 and dg are detected throughout all the colonies (Fig 5D). Thus, varied expansion of heterochromatin at sub-telomeric regions from genetically identical cells (cdc22-3 mutation in this case) occurs presumably randomly and once established, is relatively stable.
Chromatin organization and its inheritance through the cell linage are crucial for cell fate specification and cell identity maintenance during development in metazoans. We postulated that replication stress might induce chromatin change and alter gene expression and therefore, perturb the development process.
To test this notion, we first ask whether HU treatment may affect the inheritance pattern of epigenetic marks in Drosophila, using PEV of the white gene expression as the readout. The X chromosome inversion In(1)wm4 brings the white locus near to the heterochromatin. Spreading of the silent chromatin marks causes a highly variable mosaic pattern in eye coloration.
Fly larvae are treated with HU and the eye coloration patterns are quantified in adult flies. Flies are sorted into five bins based on the degree of pigmentation (Fig 6A). Bin 1 contains flies with nearly pure white eyes (silenced w locus) with only a few pigmented omatidia. The eyes of flies in bin 3 contain roughly equal sectors of red and white tissue. Bin 5 flies have eyes that were solid red (fully active w locus). When parallel cultures of flies are fed with 6 mM HU throughout larval development, roughly half the animals that form pupae die at that stage, but those that reach adulthood are placed in the same five bins based on eye pigmentation. We have found a strong shift in the distribution towards whiter eyes (Kolmogorov-Smirnov comparison p< 0.001), indicating greater probability that the w locus is silenced (Fig 6A). The same result is obtained for females carrying two copies of the wm4 locus and males with only one. This suggests that replication fork pausing in Drosophila reduces the fidelity in chromatin organization duplication.
We also test whether HU treatment affects cell fate specification during vulva development in the nematode C. elegans. A number of genes implicated in vulva development encode chromatin factors, including heterochromatin binding protein (hpl-2), histone methyl-transferase (met-2) and nucleosome remodeling protein (ssl-1), supporting a prominent role of chromatin structure in vulva development [51]. These genes and others constitute three redundant genetic pathways–SynMuv (Synthetic Multi-vulva) A, B and C–that control cell fate during vulva development. Simultaneous inactivation of two pathways results in abnormal induction of vulva cell fate and creates multiple vulva-like structures in contrast to the normal, single vulva [51].
We examined whether HU treatment would enhance the synthetic multi-vulva (SynMuv) phenotype of hpl-2 (Fig 6B and 6C). The hpl-2 (ok916ts) allele, a truncation of hpl-2 encoding only the N-terminal one third of HPL-2, displays a SynMuv Phenotype at 25°C (restrictive temperature) but not 20°C (permissive temperature) together with a SynMuv A mutation [52]. No Muv phenotype is observed in hpl-2 single mutants at either temperature ([52] and Fig 6D–6F). Synchronized L1 larvae of wild type and hpl-2 worms are placed on plates with HU, and the SynMuv phenotype is scored when worms reach adulthood. None of the HU concentrations we use stopped worm development at either 20°C or 25°C (Fig 6D–6F), indicating that these dosages do not completely block DNA replication or cell cycle progression. When worms are raised at 20°C and treated with HU, only the highest HU dosage (10 mM) results in a significant percentage of SynMuv animals in hpl-2 mutants but not in wild type (Fig 6D). When samples are raised at 25°C, a much more dramatic, HU dosage-dependent increase in the percentage of SynMuv animals is observed (Fig 6E). The most obvious induction of the SynMuv phenotype is observed from 25°C-raised hpl-2 mutant animals that are descendants of mothers that have been raised at 25°C. The HU-induced SynMuv phenotype associates with hpl-2 mutation indicates that chromatin duplication and/or remodeling is mis-regulated by HU treatment in C. elegans.
In conclusions, these results demonstrate that reduced fidelity in chromatin duplication induced by replication stress is an evolutionarily conserved phenomenon and that, in a suitable cellular context, it causes defects in cell fate specification and development.
Various external or intracellular factors such as metabolic stresses, genotoxic insults, deregulation of replication and oncogene activation could induce replication stress, which is a major source of genome instability and a hallmark of most cancer types.
In addition to genetic instability, changes in chromatin structure have also been associated with replication stress[9,10,25], suggesting that replication stress may also affect epigenetic inheritance. Here, we present a series of in vivo evidence in three organisms that replication stress disturbs chromatin duplication, causing epigenetic instability. Our results highlight the importance of coordination between replication and nucleosome assembly to ensure accurate chromatin duplication.
Previous study shows that failure to maintain processive DNA replication at G4 DNA in REV1-deficient cells leads to uncoupling of DNA synthesis from histone recycling, resulting in a local tract of chromatin lacking the parental epigenetic marks [53]. Chromatin alterations can arise as a consequence of perturbed histone dynamics in response to replication stress, which may facilitate stochastic epigenetic silencing by laying down repressive histone marks at sites of fork stalling[54]. Our in-detailed study in the fission yeast S.pombe expands this notion. Using reporter gene expression states as the readout, we demonstrated that replication stress affects the precision of inheritance of an epigenetic trait on chromatin—Cnp1 occupancy in the centromere. We further demonstrated that replication stress affects heterochromatin distribution at multiple loci in the genome. Together, these results show that replication stress broadly affects chromatin-mediated epigenetic inheritance. Our results in the fruit fly and worm studies further illustrated the broadness and the evolutionary conservation of the phenomenon.
Finding that replication stress affects epigenetic stability is potentially important for elucidating the mechanisms of its functional role, for instance, in tumorigenesis. In cancer development induced by aberrant activation of the Rb-E2F pathway, the primary cause of replication stress is the reduced nucleotide pool in cells [55], a condition similar to the HU-induced or cdc22-3 mutation-dependent replication stress tested here and in a previous study[25]. Our current work and others [9,10,25,26,56] support the model that, in addition to its well-established role in genetic instability, replication stress may also lead to epigenetic instability, providing an additional mechanism for cancer initiation and progression. Furthermore, our analysis of worm vulva’ development showcases the impact of replication stress in perturbing embryonic development via epigenetic instability.
Based on the biochemical studies in human cells, Jasencakova et al has proposed the model that replication stress may promote chromatin structure change by perturbing parental histone recycling in replication-coupled nucleosome assembly [26]. Their results have illustrated that with replication stalling, a complex comprised of the histone chaperone Asf1, histone H3-H4 dimer and the MCM helicase accumulates in nucleus, with the histones carrying the parental molecule signatures. Upon recovery from replication stalling, the accumulated histones are recycled and incorporated in the replicated chromatin [26]. It is possible that the excessive amount of ssDNA binding with RPA, formed in the presence of replication stresses, may bind the evicted parental histones and new histone H3-H4. This speculation is supporting by a recent study in budding yeast, demonstrating that during normal replication, ssDNA binding protein RPA binds histone H3-H4 and multiple histone H3-H4 chaperones and promotes the deposition of H3-H4 onto adjacent dsDNA[56]. If validated under the conditions of replication, this would provide a molecular mechanism for how replication stress may affect the chromatin epigenetic inheritance stability, causing broad alteration in the nucleosomal organization pattern [57,58] (Fig 7). At this stage, we cannot rule out the possibility of global perturbations to the physiological state of the cells due to HU treatment and/or genetic mutations may underscore the epigenetic stability perturbation. An alternative model for the above observation may be that the epigenetic instability induced by replication stress is a consequence of changes in cellular homeostasis and physiological state, for example, altered histone protein supplying due to S phase delay/arrest. However, the fact that checkpoint inactivation, which prevents the delay/arrest of S phase but exacerbates ssDNA formation, enhances the epigenetic instability in centromere (Fig 2) argues against this notion.
Studies in chick DT40 cells lacking the translesion synthesis polymerase REV1 showed that non-Watson-Crick G-quadruplex (G4) DNA causes persistent replication stalling, leaving gaps in DNA that are filled in a post-replication manner [59]. Such severe perturbation in replication leads to loss of local histone modification signature and changes in gene expression state. In this case, it is postulated that a failure to recycle the parental histones at all during post-replicative filling-in of the gaps [53], providing an alternative mechanism for epigenetic instability induced by DNA damage repair. Interestingly, in fission yeast mcm4-M68 cells at the semi-permissive condition, although there was no excessive ssDNA formation in S phase, strong Ssb2-GFP signal accumulation was detected in G2 phase (S3 Fig). This is consistent with the previous finding that mcm4-M68 cells delay cell cycle progression in G2 phase and require the Chk1-dependent DNA damage checkpoint for survival [33]. Nonetheless, these cells exhibited near-wild type levels of centromeric PEV, indicating that in contrast to REV1-deficient chick DT40 cells, here, post-replication repair of DNA damage does not induce epigenetic instability in centromere. The reason underlying this discrepancy is unclear. One possible explanation is that DNA damage due to failed replication initiation could be specifically localized at the sites of replication origins [33], thus, its repair may only affect chromatin structure locally (such as what was observed in REV1-deficient chick DT40 cells [53], [58]). It is also noteworthy that neither the wild type cells nor the mcm4-M68 mutant cells show an accumulation of ssDNA at 29°C as well as 25°C (Fig 3 and S2 Fig). However, the switch frequency is increased appreciably (although quite low) in 29 degree compared with that in 25 degree in wild type cells. This may be because the current method of Ssb2-GFP foci measurement is not as sensitive (and accurate) as needed to detect a possible minor ssDNA accumulation. Or, the high temperature may affect the epigenetic inheritance stability via a mechanism other than ssDNA accumulation.
The above mechanistic model highlights the contribution by impaired parental histone recycling to epigenetic instability. It is conceivable that such effect is broadly applicable throughout the epigenome. Although, this does not exclude the possible contribution of defects in other steps of RC nucleosome assembly in association with replication stalling, such as the maturation of newly assembled nucleosomes in which new histones obtain proper post-translational modification marks.
Specifically for heterochromatin organization in fission yeast genome, in addition to covalent modification signature carried by the parental histone molecules (H3K9me2), multiple mechanisms, such as RNAi, specific DNA element-binding proteins and HP1-mediated local spreading of histone H3 modification, are at play to establish/maintain heterochromatin in a locus specific manner[9]. These mechanisms may also be affected by replication stresses. For example, Singh and Klar found that mutations in cdc22 cause heterochromatin spreading beyond the silent mat locus[25], may via an indirect mechanism of increased recruitment of Swi6 and Clr4 by the stress-induced transcription factors Atf1/Pcr1. Consistently, we found that Atf1 mRNA expression level is increased significantly in cdc22-3 mutant cells (S8A Fig). Furthermore, ChIP analysis indicates that levels of Swi6 at the heterochromatin region and on the reporter genes outside the mating type boundaries (S8D Fig) are increased in cdcd22-3 mutant cells, supporting the notion that the enhanced spreading of heterochromatin at mating type locus and sub-telomeric region under replication stress may also via an indirect mechanism of increased Atf1-mediated recruitment of Swi6. Alternatively, removal of the jmjC protein Epe1 that antagonizes heterochromatin spreading beyond its normal borders by Cul4-Ddb1Cdt2, may also lead to expansion of the heterochromatin domain[60–62]. However, we found the levels of Epe1 are increased at all of the investigated heterochromatin sites and decreased at the ade6 gene outside the mating type boundary when under replication stress (S8E Fig). The reason underlying this discrepancy is unclear. The increased binding of Epe1 in heterochromatin region might be a result from the up-regulated bindings of Swi6 when under replication stress. Considering that the process of Pol II-dependent transcription through heterochromatin and siRNA formation are restricted to S phase [63,64]and are also affected by Epe1[60], it is possible that in prolonged S phase with replication stress, increased binding of Epe1 would stimulate the binding of Pol II to heterochromatin and thus the formation of siRNAs during S phase; subsequent removal of Epe1 by Cul4-Ddb1Cdt2 would then allow assembly of heterochromatin. Further experiments are needed to test this in future.
Comparing between mst0 clr1D cells with or without HU treatment, we also detected a slight increase in atf1 mRNA expression. However, no significant difference was found in terms of Swi6-9myc or Epe1-9myc association at the mating type locus. More evidence is needed to support that HU promote heterochromatin silencing in clr1D cells via Atf1-mediated recruitment of Swi6 as well as reduced level of Epe1 at mating type locus boundary.
To some extent, these results suggest that the direct impact of replication stress on epigenetic inheritance stability may vary, at least for heterochromatin, at various loci within the genome. Further supporting this notion, we have found sample-specific chromatin structure changes at sub-telomeric regions in three biological mutants with good correlations (Fig 5C and 5D), which suggest they are sporadic events and are epigenetically inheritable. Thus, while replication stress can alter epigenetic stability and chromatin structure at multiple loci (and perhaps throughout the whole genome), so far, there lacks a common pattern on how epigenetic stability and chromatin structure may be changed. Such change may likely vary depending on the local chromatin context, the biochemical nature of the epigenetic marker. And in some cases (such as the heterochromatin distribution at the sub-telomeric regions), the changes appear to be sporadic and variable among individual cells. Perhaps as an indication of additional levels of complications in the mode of heterochromatin change due to replication perturbation, earlier studies demonstrate that gene silencing due to heterochromatin decreased in certain replication mutations (swi7 and mcl1, defective in DNA polymerase alpha [10,65]), in contrast to a unilateral increase as demonstrated in these studies. It remains to be tested whether replication perturbation due to mutations in mcl1 and swi7 causes accumulation of ssDNA. Clearly, further studies are needed to better understand how the replication stresses affect epigenetic stability of the whole epigenome in general as well as distinct domains of the chromatin in a context-specific manner.
S.pombe strains used in this study are listed in S1 Table. Yeast strains are constructed by either random spore method or by tetra analysis. Yeast cells are grown on YE+5S medium (add with 5 supplements including histidine, uracil, lysine, leucine and adenine) or YE+4S medium (add with histidine, uracil, lysine and leucine, except adenine is provided by yeast extract, it is sufficient to provide the growth for cells with ade6+ transgene inserted in the centromere or mating type locus and for colony color differentiation). Solid malt extract (ME) medium or synthetic medium without nitrogen (EMM-N) is used for mating and sporulation. β-estradiol (Sigma) at 100nM in YES is used for growth of psf1-HBD mutants, with β-estradiol at 0.1nM in YES for psf1-HBD depletion.
Pedigree analysis is performed as described [16] with modifications to suit specific experimental conditions. To track the cell lineage for each three-generation family, cell suspension is spread in a line near the top of a thin YE+4S plate. Eight parallel lines are drawn below the line of cell spreading on the back of the plate using a marker pen. Using a tetrad dissection microscope (Nikon), one cell is picked and moved onto the first line and the plate is incubated at 25°C, allowing the cell divide into two cells. One of the daughter cells is then moved onto the fifth lines. In the ensuing generations, always leave one daughter at the original position and move the other to a designated position. After three generations, eight progeny cells are placed at specific positions, and allowed to grow into colonies by incubation at 25°C for 5 days.
To test whether HU treatment affects the switch frequency, a “hybrid” plate with a HU-free zone and a HU-containing zone is used. Cell suspension is spread in a line near the top of HU containing zone. One cell is picked and moved to the nearby location, allowing the cell divide into two cells. Then separate these two cells, waiting their division. Keep doing this step until the cell finished third generation. Move these eight progeny cells into HU-free zone at the designated position, incubate at 25°C for 5 days.
To test the effect of partial inactivation of mcm4-M68 at 29°C, three-generation tracking is performed at 29°C. The plates are then incubated at 25°C for 5 days before scoring.
Logarithmic growth cells are treated with specified concentrations of HU for 6 hours, and the Ssb2-GFP signals are measured under the microscope. Photomicrographs are obtained using a Delta vision core and personal DV. First, use the “Deconvolve” function to deconvolve all z-stacks. Then, use the “Quick projection” function to combine all z-stacks, choosing the “maximum intensity”. In “Edit Polygon” function, use the freehand polygon to draw a circle around the bright dot, reading the raw number. And put this same size of circle at the middle of cells, reading the number as background. Ssb2-GFP signal of this dot should be the subtraction of these two numbers. The total Ssb2-GFP signals in each nucleus equal the sum of all the detectable dots within the nucleus. Image processing and analysis are carried out using Soft WORX 3.2.2 software and Adobe Photoshop. Ssb2-GFP signals in wild type cells are normalized as 1X.
Individual colonies grown on sporulation (EMM-N) medium for 5 days at 30°C are exposed to iodine vapors to stain a starch-like compound produced by sporulating cells. In mating-type switching-defective strains, such as mat1-Msmt0, the intensity of dark brown/black staining indicates the level of sporulation.
ChIP analyses with H3K9me2 antibody (Abcam, ab1220) and anti-myc antibody (Abcam, ab9132) are performed as described previously [16,66]. Log-phase cells of wild type and cdc22-3 mutant are grown at 30°C in YE5S media, as described in Singh and Klar’s work [25]. The 109 cells were harvested and digested by Zymolyase 20T with final concentration of 0.25mg/ml for almost one hour at 37°C. For H3K9me2 ChIP analysis, the chromatin was pre-warmed to 37°C for 5min, followed by the addition of micrococcal nuclease (Thermo EN0181) to a final concentration of 240U/ml for 30min. The digestion reactions were incubated at 37°C with gentle rolling, and were immediately stopped by the addition of EDTA to a final concentration of 2mM. After centrifugation, the ChIP reaction was performed as before. For anti-myc ChIP analysis, 2*108 cells were harvested. And the chromatin was cross-linked by 1% final formaldehyde and stopped by adding 125mM final glycine. The chromatin was sheared by sonication with 30S ON and 30S OFF for 30 cycles.
Multiplexed libraries are prepared at the same time using the library preparation kit (KK8301) from kapa biosystems, and the barcode adapters from Life Technologies (014D01-14). All the libraries are sequenced on Ion torrent_PGM (200bp sequencing kit) with one 318 chip, and approximately 0.5–0.8 million aligned reads per sample are taken.
Portions of immunoprecipitated DNAs and whole cell extract from seven independent colonies of cdc22-3 mutant are used as the PCR template. The primers are set to amplify a fragment from telomere nearby locus (SPBPB2B2.08) along with that amplified a product from sub-telomere region (tel2R). Primers amplify meiosis gene mei4 as well as dg fragment of centromere are set as controls. The locations of the primer sets are shown in Fig 5D. The numbers shown in Fig 5D are calculated by the gray value of PCR product from ChIP samples divided by the gray value of that from WCE samples.
All real-time PCR were done with a Bio-Rad CFX96 Touch. All samples were run in triplicate to ensure accuracy of the data, and their average was calculated. PCR of 45 cycles was done using SYBR Green qPCR kit (Bio-Rad172-5120). Primers were used at 0.3uM for each experiment. The PCR product length was about 120bp. ChIP analysis was similarly performed by quantifying the amount of DNA in ChIP samples without antibodies (beads only), using the same apparatus and reagents. ChIP/WCE was determined by calculating their Ct value difference, followed by subtracting the Ct value difference between beads only and WCE. The primer sequences for qPCR are available upon request.
Instant Drosophila food (Carolina Biological) was prepared with either water or water containing 2, 5, 7, or 10 mM hydroxyurea and supplemented with a few grains of dry yeast on the surface. In(1) wm4 parents laid eggs for three days and were moved to new food vials. Few progeny were recovered from the 10 mM HU vials, but approximately half of the third instar larvae growing on 7 mM HU successfully completed pupation allowing adult eye pigmentation to be scored. The adults were sorted by sex and then by the degree of eye pigmentation (Bin 1: pure white-~15% pigmented, Bin 2: ~15–40%, Bin 3: ~40–60%, Bin 4: ~60–85%, Bin 5: 85%- solid red). In most cases both eyes of a single fly had comparable pigmentation, but when there was an obvious difference, the fly was placed in the bin for the darker eye. The experiment was performed in four trials for the–HU control (total 726 adults) and five times for the 7 mM HU assay (total 527 adults).
HU were added to NGM plates at indicated concentrations. Before E. coli (worm food) was seeded on the NGM plates, HU was also added to the E. coli suspension so that the HU concentration in the E. coli suspension matches that in NGM plates. Gravid mothers that were raised at either 20°C or 25°C were bleached and eggs collected on NGM plates that were without E. coli. After 24 hours incubation at 20°C or 25°C, synchronized L1 larvae were placed on NGM plates containing different concentrations of HU and incubated at 20°C or 25°C. Adults that were 24–48 hours post-L4 stages were scored for the SynMuv phenotype.
Data are presented as the mean±SD. Statistical analysis was made for multiple comparisons using analysis of variance and Student’s t test or Fisher exact test. A p value <0.05 was considered to be statistically significant.
Reads are mapped to S. pombe ASM294v2 assembly using “bwa” with default parameters, and only the uniquely mapped reads are obtained for further analysis. Nucleosome position and occupancy are calculated by DANPOS [67], which normalize data by random sample and adjust clonal signal to 1 read. Reads are adjusted to 73 bases with 5’ fixed when occupancy is counted. IGV is chosen for data visualization, and the annotation file (which is updated to May 2015) for S. pombe ASM294v2 assembly is downloaded from website (http://www.pombase.org).
All ChIP-Seq data have been submitted to GEO Datasets under accession numbers [GSE89816].
|
10.1371/journal.pmed.1002775 | Incidence of eclampsia and related complications across 10 low- and middle-resource geographical regions: Secondary analysis of a cluster randomised controlled trial | In 2015, approximately 42,000 women died as a result of hypertensive disorders of pregnancy worldwide; over 99% of these deaths occurred in low- and middle-income countries. The aim of this paper is to describe the incidence and characteristics of eclampsia and related complications from hypertensive disorders of pregnancy across 10 low- and middle-income geographical regions in 8 countries, in relation to magnesium sulfate availability.
This is a secondary analysis of a stepped-wedge cluster randomised controlled trial undertaken in sub-Saharan Africa, India, and Haiti. This trial implemented a novel vital sign device and training package in routine maternity care with the aim of reducing a composite outcome of maternal mortality and morbidity. Institutional-level consent was obtained, and all women presenting for maternity care were eligible for inclusion. Data on eclampsia, stroke, admission to intensive care with a hypertensive disorder of pregnancy, and maternal death from a hypertensive disorder of pregnancy were prospectively collected from routine data sources and active case finding, together with data on perinatal outcomes in women with these outcomes. In 536,233 deliveries between 1 April 2016 and 30 November 2017, there were 2,692 women with eclampsia (0.5%). In total 6.9% (n = 186; 3.47/10,000 deliveries) of women with eclampsia died, and a further 51 died from other complications of hypertensive disorders of pregnancy (0.95/10,000). After planned adjustments, the implementation of the CRADLE intervention was not associated with any significant change in the rates of eclampsia, stroke, or maternal death or intensive care admission with a hypertensive disorder of pregnancy. Nearly 1 in 5 (17.9%) women with eclampsia, stroke, or a hypertensive disorder of pregnancy causing intensive care admission or maternal death experienced a stillbirth or neonatal death. A third of eclampsia cases (33.2%; n = 894) occurred in women under 20 years of age, 60.0% in women aged 20–34 years (n = 1,616), and 6.8% (n = 182) in women aged 35 years or over. Rates of eclampsia varied approximately 7-fold between sites (range 19.6/10,000 in Zambia Centre 1 to 142.0/10,000 in Sierra Leone). Over half (55.1%) of first eclamptic fits occurred in a health-care facility, with the remainder in the community. Place of first fit varied substantially between sites (from 5.9% in the central referral facility in Sierra Leone to 85% in Uganda Centre 2). On average, magnesium sulfate was available in 74.7% of facilities (range 25% in Haiti to 100% in Sierra Leone and Zimbabwe). There was no detectable association between magnesium sulfate availability and the rate of eclampsia across sites (p = 0.12). This analysis may have been influenced by the selection of predominantly urban and peri-urban settings, and by collection of only monthly data on availability of magnesium sulfate, and is limited by the lack of demographic data in the population of women delivering in the trial areas.
The large variation in eclampsia and maternal and neonatal fatality from hypertensive disorders of pregnancy between countries emphasises that inequality and inequity persist in healthcare for women with hypertensive disorders of pregnancy. Alongside the growing interest in improving community detection and health education for these disorders, efforts to improve quality of care within healthcare facilities are key. Strategies to prevent eclampsia should be informed by local data.
ISRCTN: 41244132.
| High blood pressure in pregnancy affects nearly 1 in 7 pregnant women worldwide.
Failure to recognise and treat this condition with antihypertensives, magnesium sulfate, and delivery of the baby can result in serious complications for the mother and baby such as eclampsia, intensive care admission, and death.
In high-resource settings, severe complications are rare. In low-resource settings, they are known to be more common, but their frequency and impact are not as well reported.
We undertook a trial introducing a novel device that measures heart rate and blood pressure into routine maternity care in 10 geographical areas in 8 countries in sub-Saharan Africa, Haiti, and India, with the aim of reducing severe pregnancy complications and death.
In this study, we analysed all the women who experienced eclampsia or stroke or were admitted to intensive care or had a maternal death as a result of high blood pressure in pregnancy. In total, 0.5% of women experienced eclampsia (n = 2,692), 6.9% of these women died (n = 186), and 15.9% of their babies died (n = 397). There was substantial variation in eclampsia across different geographical regions, ranging from 0.20% to 1.42%.
Nearly a third of eclampsia occurred in women aged under 20 years, but this proportion differed between geographical regions, ranging from 9.9% to 50.1%.
More than half (55.1%) of women experienced their first eclamptic fit in a healthcare facility, but this proportion also differed between regions, with 6% to 85% of first eclamptic fits occurring in a central referral facility.
This is one of the largest studies of eclampsia to date and confirms that high blood pressure in pregnancy remains an important cause of morbidity and mortality in low-resource countries.
The majority of our geographical areas were urban or peri-urban and contained a large tertiary hospital. This means that more women with high blood pressure in pregnancy may be referred to these sites, so the rates may be higher than in the country as a whole.
Understanding place of first eclamptic fit (in the community or in small or large tertiary facilities) may be useful in targeting interventions to reduce eclampsia, focusing either on community interventions or quality of facility care.
| Hypertensive disorders of pregnancy cause 14% of all maternal deaths globally, approximately 42,000 each year [1,2]. Nearly all of these deaths occur in low-resource settings (99%), with death in high-income settings being very rare [3]. Hypertensive disorders of pregnancy encompass chronic hypertension, gestational hypertension (new hypertension without proteinuria), pre-eclampsia (new hypertension with proteinuria or end-organ damage after 20 weeks of gestation [4]), and eclampsia. The majority of morbidity and mortality is associated with pre-eclampsia and eclampsia.
It is estimated that the prevalence of pre-eclampsia globally is 4.6% (95% CI 2.7%–8.2%) [5]. The prevalence of eclampsia globally is reported to be 0.3% [6]. This is based on secondary analysis of a World Health Organization (WHO) multi-country survey that included 875 cases of eclampsia, collected over a short duration from only secondary or tertiary hospitals [6]. Women under 20 years of age, women with low levels of education, and women in their first pregnancy are all reported to be at higher risk [6]. Reliable data reporting the prevalence of maternal deaths related to eclampsia globally are scarce. Estimates from 16 datasets reported the case fatality rate to be 8.3% [5], whereas the WHO survey reported 32 maternal deaths, 3.7% of women with eclampsia [6]. Data from individual countries suggest that prevalence and mortality risk vary depending on region and socio-economic status [7].
Administration of magnesium sulfate more than halves the risk of eclampsia in women with pre-eclampsia [8]. It is considered an essential drug by WHO [9], but data on its availability in relation to prevalence of eclampsia are scarce [5]. Planned delivery after 36 weeks of gestation is effective at preventing maternal morbidity in women with pre-eclampsia [10]. Evidence for other interventions effective at reducing morbidity and mortality of pre-eclampsia is mixed [11], and research is generally undertaken in high-income settings, where the burden of illness is small. There is a lack of understanding around modifiable risk factors and availability of life-saving interventions, both vital in reducing the high number of deaths from this treatable cause.
The aim of this paper is to describe the incidence (per pregnancy) and characteristics of eclampsia, stroke, maternal death from hypertensive disorders of pregnancy, and intensive care unit (ICU) admission from hypertensive disorders of pregnancy across 10 geographical regions in 8 low- and middle-resource countries. The secondary aim is to describe the effect of a novel vital sign device and educational package on eclampsia, stroke, maternal death from a hypertensive disorder of pregnancy, or ICU admission with a hypertensive disorder of pregnancy.
This is a secondary analysis of a pragmatic, stepped-wedge cluster randomised controlled trial of the introduction of the CRADLE intervention (described below) into routine maternity care in 10 sites across Zimbabwe, Zambia, Sierra Leone, Malawi, Ethiopia, Uganda, Haiti, and India over 20 months from 1 April 2016 to 30 November 2017 (ISCRTN41244132) [12,13]. The stepped-wedge design means that at the trial start all clusters commenced data collection; all clusters then crossed from control to intervention at a randomly allocated time point, at 2-monthly intervals, until all had received the intervention. The intervention effect was then determined by comparing the data points after delivery of the intervention with those in the control period. Each site comprised at least 1 secondary or tertiary hospital that provided comprehensive emergency obstetric care (central referral facility) and the main peripheral facilities that referred to the central referral facility. All secondary and tertiary hospitals were urban or peri-urban, but the geographical regions of peripheral facilities covered a range of settings, with the mean distance to the central referral facility varying from 3.3 km to 74 km. The intervention was delivered to all healthcare professionals working in the site facilities. Community healthcare providers received the intervention where they were formally involved in routine maternity care provision and supported at the district level [12].
The CRADLE intervention consisted of the CRADLE Vital Sign Alert, an accurate, easy-to-use device that measures maternal blood pressure and heart rate and calculates shock index. Results are displayed via a traffic light early warning system [14,15]. The devices were delivered with a one-off interactive training session of CRADLE Champions, who then provided ongoing training and support for use of the device in their clinical area. Further details of the development of the CRADLE intervention have previously been described [12,16]. This intervention was compared to routine maternity care using local management guidelines.
The primary outcome of the overall trial was a composite outcome of maternal mortality and morbidity (at least 1 of eclampsia, emergency hysterectomy, and maternal death) per 10,000 deliveries. In spite of a reduction in the primary outcome over time, this trial was unable to demonstrate an effect of the intervention. For the purpose of the analysis reported here, all women who presented to maternity care at any gestational age or up to 42 days after delivery and were diagnosed with eclampsia or experienced a complication of a hypertensive disorder of pregnancy (stroke, or being admitted to an ICU or dying as a result of a hypertensive disorder of pregnancy), between 1 April 2016 and 31 November 2017 were eligible for inclusion. The denominator was all deliveries in the trial area in the same period. Eclampsia was defined as convulsions with raised blood pressure in the absence of a known neurological cause during pregnancy or within 42 days after delivery. Other data collected included maternal age at eclamptic fit, timing of eclampsia (antenatal, including day of delivery, or postnatal), and the place of first eclamptic fit (community, peripheral facility, or central referral facility). The number of stillbirths and neonatal deaths up to 28 days was recorded for all women who had antenatal eclampsia, had a stroke, or were admitted to ICU or died as a result of hypertensive disorders of pregnancy.
Sites were described by the number of deliveries, number of ICU beds per 1,000 deliveries, and the proportion of facilities (central referral and peripheral) where magnesium sulfate was available. Availability of magnesium sulfate was recorded on a monthly basis. Details on the quantity available daily or individual-level prescriptions were not collected. Methods of data collection were discussed and optimised based on the existing resources available in each site. All data collectors were given detailed training to ensure comparability of results. Outcomes were triangulated across multiple sources (including referral registers, ward registers, patient records, local mortality and morbidity records, and active case finding) to ensure data completeness, and all outcomes were checked to avoid double counting.
Ethical approval was granted by the Biomedical Sciences, Dentistry, Medicine and Natural and Mathematical Sciences Research Ethics Subcommittee at King’s College London (LRS-14/15-1484). This and all local ethical approvals were in place prior to the study start. Institutional-level consent on behalf of the cluster was obtained.
Statistical analyses were undertaken in Stata version 13.1. The main analysis used logistic regression with generalised estimating equations and a population-averaged model, with fixed centre effects and separate fixed linear trends in each site for changes in outcome over time [17]. Results are reported as odds ratios with 95% confidence intervals. The trial protocol stated that individual components of the primary outcome, including eclampsia, ICU admissions, and maternal deaths, and place of eclamptic fit would be analysed. However, there was no predefined analysis plan for this secondary analysis [12]. To describe the association between eclampsia and magnesium sulfate availability, the eclampsia rates for each site time period (month), and place of eclamptic fit were calculated. Eclamptic fits in the community were excluded from these analyses as magnesium sulfate was not available in the community. The analysis of the association between magnesium sulfate availability and total eclampsia by site used linear regression of the log of eclampsia rate with robust standard errors. The analysis of the association between magnesium sulfate availability and place of eclamptic fit used logistic regression with robust standard errors. Adjustments were made for time period (linear) and centre (categorical) to account for trends over time. Individual patient data were collected only for known cases.
In this cohort of 536,233 deliveries there were 2,692 cases of eclampsia over 20 months. This gives an incidence of eclampsia of 0.5%, as shown in Table 1 (50.2/10,000 deliveries). In total, 6.9% (n = 186; 3.47/10,000 deliveries) of women with eclampsia died (sepsis [n = 4], stroke [n = 4], haemorrhage [n = 18], hypertensive disorders of pregnancy [n = 150], and other causes [n = 10]), and a further 51 women died from other complications of hypertensive disorders of pregnancy without having had eclampsia (0.95/10,000). Eight of the 10 sites had capacity for ICU admission, although availability of beds varied between sites (S1 Table). In total, 1,322 women were admitted to ICU with hypertensive disorders of pregnancy, 27.8% of these with eclampsia (n = 367) and 72.2% (n = 955) with other complications of hypertensive disorders of pregnancy. After planned adjustments for clustering and time trends in each site, the implementation of the CRADLE intervention was not associated with any significant change in the rates of eclampsia, stroke, maternal death from a hypertensive disorder of pregnancy, or ICU admission with a hypertensive disorder of pregnancy.
Rates of eclampsia varied between sites from 19.6 per 10,000 deliveries in Zambia Centre 1 (Lusaka) to 142.0 per 10,000 deliveries in Sierra Leone (Fig 1; S2 Table). When comparing the effect of the intervention across individual sites, further consideration is required as these are non-randomised data and are vulnerable to external influences such as seasonal trends. After planned adjustment, there was a significant reduction in eclampsia in Haiti and Zambia Centre 1, a significant increase in Malawi and Uganda Centre 2, and no significant changes in other sites (S2 Table). The rate of maternal death from eclampsia in each site largely reflected the incidence of eclampsia (from 0.4 per 10,000 deliveries in Zambia Centre 1 to 15.5 per 10,000 deliveries in Sierra Leone; S3 Table). The range of case fatality rate for women with eclampsia was from 2.1% (5/242) in Zambia Centre 1 to 14.4% (18/125) in Haiti. Only 33 in the cohort of 536,233 women were diagnosed with a stroke. ICU admission from hypertensive disorders of pregnancy also varied between sites as shown in Fig 1 (S3 Table).
Across all sites, 92.7% (n = 2,495) of eclampsia cases occurred in the antenatal period and 7.3% (n = 197) in the postnatal period. The proportion of eclampsia cases occurring in the antenatal period was similar across sites (range 88.9% in Uganda Centre 1 to 98.2% in Uganda Centre 2) (Fig 2; S4 Table). Approximately a third of eclampsia cases (33.2%; n = 894) occurred in women aged under 20 years. This proportion varied between sites from 10% in Ethiopia to 51% in Malawi (Fig 2; S4 Table). The majority of eclampsia cases occurred in women aged 20–34 years (60.0%; n = 1,616); 6.8% (n = 182) occurred in women aged 35 years or over.
In total, there were 10 central referral facilities and 268 peripheral facilities. Nearly half of all first eclamptic fits occurred in the community (44.9%; range 30.8% in Malawi to 66.0% in Freetown; Table 2), with 31.2% occurring for the first time in central referral facilities (range 5.9% in Sierra Leone to 85.0% in Uganda Centre 2) and 23.9% in peripheral facilities (range 4.7% in India to 33.0% in Ethiopia). On average, magnesium sulfate was available in 74.7% of facilities (range 25% in Haiti to 100% in Sierra Leone and Zimbabwe). Availability of magnesium sulfate did not significantly change during the trial period. There was no significant association between the overall availability of magnesium sulfate in central referral and peripheral facilities and the proportion of eclampsia cases that occurred in each (central referral facilities: p = 0.42; peripheral facilities: p = 0.13; Table 2). There was also no detectable association between the proportion of facilities with magnesium sulfate available across the sites and the rate of eclampsia in each site (p = 0.12). Of the 1,210 women who had their first eclamptic fit in the community, 7.5% (91/1,210) died; of the 1,482 who had their first eclamptic fit in a facility, 6.4% (95/1,482) died.
In the group of 3,493 women with antenatal eclampsia (n = 2,495), stroke, or hypertensive disorders of pregnancy causing ICU admission or maternal death (n = 998), the rate of stillbirth or neonatal mortality was very high (17.9%; n = 625). The rate of stillbirth or neonatal mortality was higher in women with hypertensive disorders of pregnancy (i.e., resulting in stroke, ICU admission, or maternal death) without eclampsia than in women with eclampsia (stillbirth or neonatal death: 22.8% [n = 228] in women with hypertensive disorders of pregnancy without eclampsia and 15.9% [n = 397] in women with eclampsia; Table 3). Overall rate of stillbirth or neonatal mortality in women with eclampsia varied between sites from 4.1% in Malawi to 23.1% in Uganda Centre 1 (S5 Table).
Overall, we report that 0.5% of women in our sites experienced eclampsia, 57.2% of women with eclampsia were admitted to ICU, and 6.9% died. Our individual site analysis shows large variation both in the rate of eclampsia and in the rates of maternal death and ICU admission from hypertensive disorders of pregnancy. Stroke was a rare outcome in all of our sites. The majority of eclampsia cases across all sites first occurred in the community (44.9%), in the antenatal period (92.6%), and in women aged 20–34 years (60.0%). Overall, the implementation of the CRADLE intervention was not associated with any significant change in the rate of eclampsia, stroke, or maternal death or ICU admission with hypertensive disorders of pregnancy, but the effect in individual sites varied.
To our knowledge, this is the largest prospectively collected dataset on women with eclampsia. The strengths of these data are the rigorous method of prospective data collection, verified from multiple sources, and inclusion of multiple countries and settings. The majority of existing data have focused on eclampsia presenting to secondary or tertiary hospitals [6]. The data presented here improve the accuracy of incidence estimates by including cases across the health system, including cases from primary healthcare facilities and community cases.
Although the geographical settings of this study varied, it is a limitation of this study that the majority of sites were urban or peri-urban. These settings were selected as a substantial proportion of births occur in them. This approach, in addition to the inclusion of the national referral hospital in many of our sites, means the incidence of complications from hypertensive disorders of pregnancy in our sites may be higher than country-wide levels. Due to the size of the study, it was not feasible to collect demographic data in the population of women who delivered in the trial area. Therefore, the proportion of eclampsia cases in different age groups and perinatal outcomes cannot be presented at the population level. Effects of the intervention on perinatal outcomes or by age or place of eclampsia are therefore not presented. The effect of the intervention in individual sites needs further consideration as these are non-randomised analyses. The numbers of cases of eclampsia and hypertensive disorders of pregnancy were based on the data reported by attending clinicians in the health facilities; it was not feasible to undertake additional searching in all sites to identify cases not reported. However, the inclusion of maternal deaths and ICU admissions only from hypertensive disorders of pregnancy means that misdiagnosis is less likely. Whilst it is a strength that this paper reports magnesium sulfate availability, this variable was collected on a monthly basis at the level of the facility. As daily fluctuations in the quantity available or the number of doses prescribed remain unknown, it is possible that supply may not have been adequate to meet demand.
In the post–Millennium Development Goal era, the focus of global health is on not just reducing mortality but also reducing morbidity [18]. Yet, in this study, the large variation in fatality from eclampsia between countries emphasises that both inequality and inequity in management of hypertensive disorders of pregnancy persist. It has been previously reported that organ dysfunction is up to 60 times more frequent in women with eclampsia compared to women without eclampsia [6]. Therefore, the very high rates of maternal death in some countries compared to previously reported estimates [6,7] highlight that the true burden of disease in these countries may be even greater than previously recognised and that hypertensive disorders of pregnancy should remain firmly on the global agenda.
This study showed no effect of the CRADLE intervention on eclampsia, stroke, or maternal death or ICU admission from hypertensive disorders of pregnancy. It is possible that the intervention increased detection but without the capacity to improve management. The primary purpose of the study was to collect accurate incidence data, and therefore detailed case information on clinical management was not routinely collected, given the resource constraints of the trial. It is challenging therefore to draw conclusions on differences in management that may also have contributed to variations in the rate of eclampsia and resulting morbidity seen. However, it is evident that Zambia Centre 1 (Lusaka) had the lowest rate of eclampsia and case fatality, and admitted substantially more patients to ICU than any other site. This was possible as they have a specialist unit specifically for women with hypertensive disorders of pregnancy that provides continuous monitoring and close observation by trained staff. In comparison, the site in Freetown in Sierra Leone, which had the highest rate of eclampsia and case fatality, has no higher-level care available. The availability of monitoring to rapidly detect deteriorations—and initiate treatment such as antihypertensives, magnesium sulfate, and timely delivery—is likely to be important. This idea is in keeping with reports that the largest reductions in maternal mortality from hypertensive disorders of pregnancy in England and Wales were achieved with improved surveillance, diagnosis, and timely delivery, with further benefit from fluid-restriction management protocols and increased use of anticonvulsant therapies in more recent decades [19].
In this study, nearly a third of eclampsia cases occurred in women aged under 20 years. This proportion varied greatly between sites, with the Malawi site reporting that half of eclampsia cases occurred in women aged under 20 years. Other studies have reported rates of 26% [6] to 55% [20]. Whilst this study did not collect demographic data in all deliveries, nationwide demographic data show that 15% of births in Malawi in 2015–2016 occurred in women aged under 20 years [21]. Existing literature suggests that teenage pregnant women are at greater risk of eclampsia [22], and their care should be prioritised in clinical practice. Interventions aiming to overcome the complex socio-cultural needs of this group to improve access to healthcare and prevent eclampsia warrant further research.
This study also presents novel data on the place of eclamptic fit, previously only reported in smaller cohort studies, where 74.5% (n = 142) were reported to occur before hospital admission in Latin America [20]. Our data demonstrate that over half of women experience their first eclamptic fit in a healthcare facility, despite the relatively good availability of magnesium sulfate in these settings. However, the proportion of eclampsia cases that first occurred in healthcare facilities compared to the community varied substantially between sites. This suggests that the most appropriate interventions and strategies to reduce eclampsia should be informed by local incidence data. For example, in Sierra Leone, Zambia Centre 1, and India, where a high proportion of eclampsia cases occurred in the community, interventions aiming to improve community monitoring and overcome barriers to accessing care, including health education [23], may be the most appropriate use of resources. In contrast, in Uganda Centre 2 and Haiti, targeting the quality of care within facilities may be a more effective strategy for preventing eclampsia. Therefore, when vital actions such as treating severe hypertension with magnesium sulfate to prevent eclampsia [8] and timely delivery of the baby [10] are recommended, national and international policy makers may also recommend collection of regional data to identify how these interventions should be delivered to achieve the greatest benefit locally, thus maximising their impact and identifying the most appropriate use of resources.
In conclusion, this analysis provides accurate contemporaneous estimates of incidence of eclampsia and hypertensive disorders of pregnancy from the largest known prospective dataset across 8 low- and middle-resource settings. These data highlight that mortality (for the woman and baby) from eclampsia remains high, and higher-risk groups exist that should be prioritised in research and policy. Use of magnesium sulfate to prevent eclampsia and timely delivery after diagnosis remain important strategies to reduce maternal and perinatal mortality from hypertensive disorders of pregnancy at the facility level, but interventions should also be targeted to meet the needs of specific regions.
|
10.1371/journal.pcbi.1002335 | Protein Design Using Continuous Rotamers | Optimizing amino acid conformation and identity is a central problem in computational protein design. Protein design algorithms must allow realistic protein flexibility to occur during this optimization, or they may fail to find the best sequence with the lowest energy. Most design algorithms implement side-chain flexibility by allowing the side chains to move between a small set of discrete, low-energy states, which we call rigid rotamers. In this work we show that allowing continuous side-chain flexibility (which we call continuous rotamers) greatly improves protein flexibility modeling. We present a large-scale study that compares the sequences and best energy conformations in 69 protein-core redesigns using a rigid-rotamer model versus a continuous-rotamer model. We show that in nearly all of our redesigns the sequence found by the continuous-rotamer model is different and has a lower energy than the one found by the rigid-rotamer model. Moreover, the sequences found by the continuous-rotamer model are more similar to the native sequences. We then show that the seemingly easy solution of sampling more rigid rotamers within the continuous region is not a practical alternative to a continuous-rotamer model: at computationally feasible resolutions, using more rigid rotamers was never better than a continuous-rotamer model and almost always resulted in higher energies. Finally, we present a new protein design algorithm based on the dead-end elimination (DEE) algorithm, which we call iMinDEE, that makes the use of continuous rotamers feasible in larger systems. iMinDEE guarantees finding the optimal answer while pruning the search space with close to the same efficiency of DEE. Availability: Software is available under the Lesser GNU Public License v3. Contact the authors for source code.
| Computational protein design is a promising field with many biomedical applications, such as drug design, or the redesign of new enzymes to perform nonnatural chemical reactions. An essential feature of any protein design algorithm is the ability to accurately model the flexibility that occurs in real proteins. In enzyme design, for example, an algorithm must predict how the designed protein will change during binding and catalysis. In this work we present a large-scale study of 69 protein redesigns that shows the necessity of modeling more realistic protein flexibility. Specifically, we model the continuous space around low-energy conformations of amino acid side chains, and compare it against the standard rigid approach of modeling only a small discrete set of low-energy conformations. We show that by allowing the side chains to move in the continuous space around low energy conformations during the protein design search, we obtain very different sequences that better match real protein sequences. Moreover, we propose a new protein design algorithm that, contrary to conventional wisdom, shows that we can search the continuous space around side chains with close to the same efficiency as algorithms that model only a discrete set of conformations.
| Computational structure-based protein redesign is a promising field with applications for drug design [1], biosynthesis [2], protein∶peptide design [3], and predicting drug resistance [4]. The goal of a structure-based protein redesign algorithm is to search over protein conformations and find the global minimum energy conformation, or GMEC, with respect to a given protein design model. The protein design model defines both the input to the algorithm and how the redesigned protein can move (the flexible space). As input the algorithm takes one or several starting protein structures, an energy function to score the designed proteins, and whether the design search allows amino acid type mutations (a mutation search). If mutations are allowed, the protein design algorithm searches protein conformations from multiple sequences to find the amino acid sequence of the GMEC.
Most protein design models limit the flexible space during the search in the interest of computational feasibility. A common protein design model assumes a fixed backbone and only allows the side chains to move among a set of discrete conformations called rotamers. Rotamers are determined using theoretical calculations and the empirical observation that the side chains of amino acids in protein structures avoid most of the available conformational space and appear frequently as clusters in -angle space [5] (Figure 1A).
Traditionally, a rigid-rotamer model is used for protein design. The rigid-rotamer model represents each empirically-determined side-chain cluster as a single discrete rotamer (usually the modal or mean value of the cluster's distribution is chosen for the rotamer conformation (Figure 1B)). However, protein energetics are sensitive to small changes in atom coordinates, so a single discrete conformation cannot fully describe a continuous region of side-chain conformation space. On the other hand, the continuous-rotamer model allows each rotamer to represent a region in -angle space in order to more accurately reflect the empirically-discovered side-chain clusters (Figure 1C). Because both methods use different rotamer models, they obtain different GMECs; we refer to the GMEC when using a rigid-rotamer model, and the continuous-rotamer model, respectively, as the rigid GMEC and the minGMEC.
Many protein design algorithms focus on finding the rigid GMEC instead of the minGMEC. These algorithms often try to account for this simplification by allowing side-chain angles to rotate slightly after the rigid search to optimize energy interactions, a process known as post hoc energy minimization. This is dangerous because rigid rotamers will often score poorly during a search and be discarded, even though they can potentially minimize to lower energies than the rigid GMEC. The toy example in Figure 2 illustrates how rotamers that are part of a well-packed structure would be discarded by a rigid-rotamer search. Even though a post hoc energy minimization of the rigid-rotamer model in this example would result in a low-energy structure, the pre-minimization energy would be so high that this conformation would not be considered for minimization. Thus, rigid-rotamer methods are likely to not even consider the minGMEC as a good candidate structure.
Previous work has shown the benefit of continuously minimizing rotamers [6], [7]. For example, the method described in [7] extends post-hoc energy minimization by allowing rotamers to change during the minimization step. First, a Monte Carlo, rigid-rotamer based algorithm finds a low-energy structure. Next, one residue position at a time, rotamers for that position are continuously minimized, and the lowest energy rotamer is chosen. Thus, the method in [7] is (a) dependent on the rigid-rotamer solution, (b) dependent on the order residue positions are minimized, and (c) does not explicitly allow concerted side-chain movements. In contrast, we use continuous rotamers instead of relying on a rigid-rotamer search. The new design search is no longer over discrete side-chain conformations. Instead, each side-chain rotamer is a continuous region of -angle space. Therefore, our method is independent of the order in which rotamers are minimized, and allows for coordinated side-chain movements. The use of continuous rotamers guarantees that our protein design search, (i) can find the global minimum energy sequence for continuously minimized side chains, and (ii) never gets stuck in local minima. Our results show the benefits of using continuous rotamers over rigid-rotamer-based models.
In this work we focus on the protein design method dead-end elimination (DEE) because it provably finds the globally optimal solution according to the protein design model. Many protein designs, however, use heuristic algorithms instead of provable algorithms. Heuristic algorithms make no guarantees on the optimality of the solution, but they are popular because of their speed. Our results are relevant to these methods as well because the optimal solution computed by DEE provides a bound on the accuracy of all possible heuristic methods. We can therefore measure precisely the limitations of any rigid-rotamer algorithm. The original DEE algorithm (referred to in this paper as rigid DEE) finds the GMEC with respect to the discrete rigid-rotamer model by pruning rotamers that provably cannot be part of the rigid GMEC [8]. An advancement of rigid DEE, the MinDEE algorithm [9], [10], addresses the problem of finding the minGMEC by computing an upper and lower bound on the continuous energies of each rotamer and each pairwise rotamer interaction.
In addition to finding lower bounds for each rotamer individually, MinDEE also finds energy bounds for the possible change in energetics that might occur during minimization across the entire protein. The MinDEE pruning criterion prevents the algorithm from using a rotamer to prune a rotamer if could potentially perturb the other minimizing side chains during its minimization to make it a higher energy rotamer than (Figure 2). Even though MinDEE is a powerful technique that prunes the design conformation space by orders of magnitude, the range of potential minimization perturbations that MinDEE considers results in unrealistically loose bounds that bracket each energy interaction. These bounds represent theoretical worst cases which reduce MinDEE's capacity to prune. Therefore, MinDEE's pruning power is significantly weaker than rigid DEE.
MinDEE is an integral part of the algorithm [2], [9], [11], an ensemble-based algorithm that estimates the binding constant of a protein-ligand complex through a provably-accurate approximation of the partition function. was used prospectively in drug design [1], enzyme redesign [2], protein∶peptide design [3], and drug resistance prediction [4], all with experimental validation. approximates the partition function by evaluating only the low energy conformations that carry the largest weight in the Boltzmann-weighted partition function. The MinDEE algorithm is essential for , since MinDEE prunes the majority of conformations that cannot minimize into low energy conformations, and therefore need not be considered by . Therefore, improvements to the MinDEE criterion and algorithm directly improve the efficiency of MinDEE/A* and the algorithm.
In this work we show that when a protein design algorithm uses a continuous-rotamer model, the algorithm is able to find the minGMEC, which is often a much lower energy sequence than the rigid GMEC. Specifically, we show that the MinDEE algorithm is able to find lower energy sequences than those found by rigid DEE in 66 out of 69 proteins from the PDB. We also show that trying to find the minGMEC by increasing the number of rotamers in the rigid-rotamer model (Figure 1D) is often impractical, and still fails to find the minGMEC in most cases. In addition, we propose a simplified and improved alternative to MinDEE, which we call iMinDEE. iMinDEE uses a new technique that we call Greedy Estimation of Minimization (GEM), which allows iMinDEE to reduce the search space by orders of magnitude when compared to MinDEE. iMinDEE and MinDEE are mathematically guaranteed to compute the same results, and to check this is true, we ran both algorithms and obtained identical results. Finally, we used native sequence recovery, a commonly used metric to evaluate protein design algorithms, to show that continuous rotamers result in more biologically accurate protein redesigns. We tested how well the sequences of both the minGMEC and the rigid GMEC recapitulated the native protein sequence and found that iMinDEE significantly improves native sequence recovery over rigidDEE.
In this work we focus on the importance of using continuous rotamers instead of rigid rotamers in protein design. First, we establish that protein design searches that use continuous rotamers find sequences lower in energy than those using rigid rotamers. Next, we present an improved and simplified DEE pruning criterion that makes continuous-rotamer protein design more computationally feasible.
In this section we first describe the original rigid DEE [8] and MinDEE criteria [9], and then show an experimental comparison of the two methods. This comparison shows that MinDEE provides a substantial advantage over rigid DEE in computing low-energy sequences. Finally, we compare a rigid-rotamer protein design search using an expanded rotamer library against MinDEE with a standard rotamer library.
The MinDEE algorithm is guaranteed to find the GMEC when searching over continuous rotamers, which we call the minGMEC. To efficiently prune and search over continuous-rotamer conformations, the MinDEE algorithm computes lower and upper bounds on the pairwise energies of continuous rotamers ( and , as defined above). In practice, however, these maximum and minimum bounds can be very loose. This results in a large gap between the maximum and minimum terms, which consequently makes the terms in the MinDEE pruning criterion (Eq. 5) very large. For example, a pair of neighboring tryptophan rotamers might have a maximum energy within a voxel of , but these can minimize and form favorable stacking to an energy of . These large terms make it difficult to prune rotamers, resulting in much less pruning than rigid DEE.
In this section we present a new criterion and algorithm, iMinDEE, which can prune rotamers much more efficiently than MinDEE and is still guaranteed to find the minGMEC. iMinDEE obtains improved pruning by removing the need to define maximum bounds on continuous-rotamer energies, which eliminates the large terms from the pruning criterion. Remember that the terms from the MinDEE criteria were needed to account for all possible side-chain rearrangements that could occur during protein minimization. Instead of accounting for all potential side-chain rearrangements, iMinDEE greedily estimates how much minimization can actually occur.
We refer to the overall technique that iMinDEE uses to prune rotamers as Greedy Estimation of Minimization (GEM). The basis behind GEM is to greedily assume that protein minimization occurs independently for each rotamer pair. Rotamers are initially pruned based on this assumption, and the A* algorithm finds the best conformation in the remaining (unpruned) conformational search space. After this first run, we can check whether the assumption was wrong and if the minGMEC was pruned. Remarkably, if the minGMEC was pruned, we can provably refine our initial assumption to obtain a new pruning criterion that is guaranteed to recover the minGMEC, and the algorithm will run at most one more time.
iMinDEE is mathematically guaranteed to compute the same result as the original MinDEE, but can do so much more efficiently. To show the benefit of our approach, we implemented iMinDEE and applied it to the 69 protein core redesigns. We show that iMinDEE significantly reduces the conformation search space compared to the original MinDEE criterion. In many cases iMinDEE is nearly as efficient as rigid DEE, while still guaranteeing to compute the minGMEC. Finally, we analyze the meaning and impact of the interval term, , in the iMinDEE criterion.
We show here and in previous work [9], [18], [19] that rotamers pruned by rigid DEE can often minimize below the rigid GMEC. Specifically, in 68 of our test systems (Figure 3), MinDEE finds different rotamers for the minGMEC than for the rigid GMEC, as well as different amino acid sequences (in some cases differing in over half of the amino acids) in 66 of the designed protein cores. This demonstrates the importance of using continuous rotamers to find the true minimum energy conformation given the input energy function. In addition, we have developed a new algorithm, iMinDEE, which greatly increases the efficiency of searching over continuous rotamers during protein design.
Stable wild-type proteins have well-packed cores, and mutations that decrease core packing can result in unstable or misfolded proteins [20]–[22]. This is important for our designs because all of the residues that we selected are part of the protein core and have low solvent accessibility (see Materials and Methods). In nearly all of our designs the mutated side chains of the minGMEC have a larger volume than those of the rigid GMEC ( on average, as high as ). In an average example, 1ZZK with 12 redesigned amino acids and a volume difference of , the rigid GMEC and the minGMEC differ in four amino acids: three residues are larger in the minGMEC (M20, M47, I70 in the minGMEC vs. V20, T47, T70 in the rigid GMEC ), and just one residue is smaller (A73 in the minGMEC vs. S73 in the rigid GMEC). Rigid DEE selects a sequence with much smaller amino acid side chains because it cannot find a low energy conformation for the minGMEC sequence. Since overpacking of the minGMEC is unlikely because all of the minGMEC conformations have good vdW potential energies, this increase in volume supports better packing of the minGMEC with respect to the rigid GMEC. Therefore, we believe that modeling continuous rotamers in protein design will reduce the misfolding and increase the stability of predicted proteins.
To further evaluate the biological relevance of our results we performed native sequence recovery with rigid DEE and iMinDEE. iMinDEE obtained significant improvements over rigid DEE in sequence recovery. This shows the importance of fully exploring the protein structural landscape to find the lowest energy structures according to the energy function. Previously, sequence recovery has been used to demonstrate the importance of incorporating desolvation penalties into a protein design energy function [23]. These penalties are usually considered essential for protein design because they account for the hydrophobic effect, which drives protein folding [24]. Interestingly, our results show that the increase in sequence recovery obtained using continuous rotamers is comparable to the increase in sequence recovery obtained by incorporating implicit solvation [23]. This implies that accurately modeling continuous rotamers is as vital to computing accurate designs as incorporating sophisticated energy terms.
It is informative to categorize our sequence recovery results by amino acid mass: (i) small-mass amino acids with a mass less than 100 Da (Val, Ala, Gly, and Ser); (ii) medium-mass amino acids, with a mass between 100 Da and 130 Da (Asp, Lys, Ile, Gln, Asn, Leu, Glu, Thr); and large-mass amino acids, with a mass over 130 Da (Trp, Phe, Tyr, Arg, Met, His). Our results show that, in a rigid-rotamer model, the large-mass residues are recovered significantly less frequently than the small-mass residues. We show that rigid DEE recovered 83.55% of the small-mass residues, but only 45.56% of the large-mass residues. By using a continuous-rotamer model the difference in native sequence recovery of the large-mass residues vs. the small-mass residues is much smaller. iMinDEE recovered 86.54% of the small-mass amino acids and 71.11% of the large-mass amino acids. This further demonstrates that continuous rotamers are necessary to model large amino acids because they are sensitive to small changes in angles.
One might think that increasing the size and resolution of the rotamer library would allow rigid DEE to find the minGMEC. Although this is true in the limit, it is impractical to systematically increase the size of the rotamer library because the side chains of amino acids have many degrees of freedom. If flexibility is handled through more sampling, the protein designer must determine on an ad hoc basis what additional sampling should be done within the limits of computational feasibility to allow an angle to deviate from ideal rotamer values. We show in this work that increasing the rotamer library by diversifying the , or and dihedrals still fails to find sequences identical to the minGMEC, and in many cases causes the search to become intractable.
With the introduction of iMinDEE we show that continuous rotamers can efficiently be searched to find the minGMEC. Our pruning results (Figure 7) show that iMinDEE always prunes significantly more rotamers than MinDEE. This increase in pruning greatly reduces the number of protein conformations that A* must search through to find the minGMEC. Remarkably, iMinDEE often prunes close to as many rotamers as rigid DEE. The comparison between iMinDEE/MinDEE and rigid DEE pruning is somewhat complex to interpret since rigid DEE pruning is often incorrect relative to the MinDEE criterion, and the minGMEC is in most cases pruned by rigid DEE. It could also be argued that MinDEE intrinsically should not prune as much as rigid DEE, because its correctness criterion is more stringent (i.e. minimization-aware). Nevertheless, we show that the pruning of MinDEE can be greatly increased while still maintaining correctness. Both MinDEE and iMinDEE have identical outputs, and both guarantee not to prune the minGMEC, and yet iMinDEE prunes orders of magnitude more conformations in all cases.
Pruning with iMinDEE for each design system is greatly affected by the value for that system. The results in Figure 8 show that the performance of iMinDEE can be improved by reducing the value of . is defined as the difference between and (Eq. (8)). Hence, can potentially be reduced either by finding a conformation with a lower energy, or by improving the lower bound on the energy of (see Figure 6). First, to find a low-energy conformation for , the parameter of the iMinDEE algorithm must be chosen with care. While a large can lead to very little pruning during the first iMinDEE pruning step, a very small could prevent a low-energy minimized conformation (i.e. a low energy conformation , see Figure 6) from being found. This would cause to have a high energy and make needlessly large. Second, to improve the lower bound on the energy of requires improving all of the rotamer energy bounds. The example in Figure 9 shows a case where a poor lower bound on the energy of can arise because iMinDEE decomposes the system into rotamer pairs and uses bounds on these pairs to compute the total lower energy bound. One way to prevent this would be to compute lower bounds in a four-wise manner (Arg44 would compute the lower bound with all combinations of neighbors), but this would increase the complexity of the problem by forcing bounds computations (where is the number of rotamers per residue, and the number of mutable residues). If a four-wise bounds computation solved this specific case, there might be other cases where a higher-order, -wise computation might be necessary. However, is most likely effectively bounded by a small constant. Improving these bounds as well as choosing an optimal for each design system represents an interesting future research direction.
Our results suggest that the optimal value of (Eq. (6)) measures the difficulty of accurately designing a given protein system for any pairwise-energy based design algorithm. First, we observed that larger values resulted in less iMinDEE pruning (Figure 8). We also found that rigid DEE with RL2 fails to complete the design search for proteins where iMinDEE computed a large value. These results suggest that large -value systems represent difficult design problems for any pairwise-energy based design algorithm. However, since the value computed for is dependent on the value of chosen in the iMinDEE algorithm (as described above), it is likely that the optimal value of , which is approximated by , reflects the intrinsic difficulty of a design problem. Therefore, we believe that , which can be approximated by , measures an intrinsic degree of difficulty of any design run.
Our previous work, the Backbone DEE (BD) [18] and Backrub DEE (BRDEE) [19] algorithms, showed that we can provably incorporate backbone flexibility into protein design, similar to how MinDEE incorporates side-chain flexibility. Therefore, we can expect an analysis of continuous versus rigid backbone flexibility to yield similar results to those presented here, and that the iMinDEE algorithm presented here can be extended to improve the pruning efficiency of the BD and BRDEE algorithms.
In this work we show that incorporating continuous rotamers into protein design algorithms can lead to substantially improved design predictions. We used the DEE/A* framework to demonstrate these gains, but our results are applicable to any design method that uses a similar protein design model. As defined in the Introduction, the protein design model defines both the input to the algorithm (i.e. energy function and rotamer library) and how the redesigned protein can move (i.e. rigid rotamers or continuous rotamers). Imagine we use the same protein design model, but use different algorithms. Because rigid DEE/A* is guaranteed to find the best sequence according to the protein design model, any design method that uses rigid rotamers, such as Faster [25], Monte Carlo [26], or simulated annealing [27], will never find a lower energy sequence than the rigid GMEC found by DEE/A*. Therefore, the energies of the conformations computed by DEE/A* will always be as low or lower than those computed by non-DEE/A*-based methods using the same protein design model. Hence, our DEE-based results provide a bound on the performance of the other methods. Similarly, the iMinDEE/A* algorithms provide a bound on how well any algorithm based on continuous rotamers can perform. By using these bounds, we can precisely measure the consequences of using rigid rotamers to approximate continuous rotamers, and obtain general results that are applicable to all other algorithms using either rigid or continuous rotamers. We can therefore guarantee that the limitations of rigid rotamers are as important for other protein design methods as they are for rigid DEE/A*.
The main consequence of using rigid rotamers in the design search is that the search for side-chain conformations that result in low energy protein structures will not be accurate. Our results show that improving the accuracy and realism of the modeled protein flexibility can greatly improve the results of the design search. In our work we used a simple energy function in which every term can be related to physical phenomena, and found that by switching from rigid to continuous rotamers we could discover lower energy sequences and observe large gains in sequence recovery. This demonstrates that if all sequences and structures are not adequately searched to find the lowest energy ones, the most biologically-relevant results are missed. Unfortunately, the importance of accurately searching for the true lowest energy structure and sequence is sometimes overlooked and the inaccuracies are attributed instead to the energy function. Protein design energy functions are constantly improved through careful crafting to better correlate designs with retrospective biological results. Many improvements to energy functions are made through the introduction of complex statistical terms based on structural bioinformatics data and other additional parameters[28], [29]. If the rigid-rotamer search inaccuracies are wrongly attributed to imperfections in the energy function, the results will be used to incorrectly modify the energy function. Therefore, to avoid over-fitting the energy function, accurate flexibility, such as continuous rotamers, should be used during the design process.
It is often assumed in the protein design field that even if the minGMEC and the rigid GMEC are different, minimizing and reranking the top results from a rigid approach can lead to finding the minGMEC [30]. Several of our results suggest that this is very likely to not be the case, and the minGMEC would never even be considered by any rigid-rotamer method. First, the enormous difference in sequence and amino acid composition between the rigid GMEC and the minGMEC is striking: in some cases the difference is over 60%. Second, the side chains of the amino acids in the rigid GMEC tend to have a smaller volume than the side chains of the minGMEC, suggesting that unavoidable clashes in a rigid-rotamer model would make the rotamers of the minGMEC unable to sterically fit in a rigid-rotamer environment. We analyzed the conformations of the minGMEC in all of our 69 designs and found that if the continuous rotamers were replaced by their closest (i.e. in -angle space) rigid-rotamer counterpart at each position, most of the designs would obtain high-energy steric clashes (up to 1000 kcal/mol higher than the rigid GMEC). Even when the rigid-rotamer library was expanded, the new library could not capture the low-energy sequences of the continuous rotamers. Thus, contrary to conventional wisdom, rigid rotamers are always a severely limited approximation to continuous rotamers.
Any protein design algorithm that switches from using rigid rotamers to continuous rotamers will expand the side-chain search space it explores. As the sequence and conformation space increases, it is always desirable to quickly and efficiently reduce the space to make the search more tractable. In this work we presented the novel iMinDEE pruning condition which can reduce the conformational space by many orders of magnitude. After iMinDEE pruning we search the remaining conformational space with the A* search algorithm. We use A* as the search algorithm because it is guaranteed to find the optimal answer, but any search algorithm can be used in combination with iMinDEE. In fact, an approach analogous to using iMinDEE with a different continuous-rotamer search algorithm is frequently used in rigid-rotamer protein design protocols. Rigid DEE was used as a filter for Monte Carlo searches [31] or for the FASTER algorithm [25]. iMinDEE can therefore have considerable impact for any protein design algorithm that uses continuous rotamers.
Crystal structures of protein chains with a maximum percentage sequence identity of 10% and a maximum resolution of 1.3 Å were chosen using the PISCES protein culling server [32]. In addition, the protein chains were restricted to have a maximum length of 100 residues. The protein crystal structures were gathered from the PDB and further curated by adding hydrogens [33] and removing waters and ions. Residues with missing side chains were either removed entirely or the missing atoms were added using the King software package [34]. In total, 69 protein structures were selected for the test set.
For each protein in the test set, a redesign to find low energy sequences for the initial backbone (a mutation search) was conducted. Each mutation search was designed so that approximately 12–15 core residues of the protein would be mutable. Core residues were chosen by finding all residues with a side-chain relative solvent accessible surface area (SASA) less than either 5%, 10%, or 20%. SASA values were determined with the program NACCESS [35]. If a protein had less than 12 residues with 20% SASA, only these residues were allowed to mutate. Each mutable residue was allowed to take on its wild-type identity and several other amino acid types. The mutant amino acid types were determined by finding the 5–7 most likely amino acid type substitutions based on the BLOSUM62 matrix [36]. The AMBER [37] energy function and the Richardson's Penultimate Rotamer Library [38] were used as input to the algorithm. Each design run consisted of three steps: (1) A pairwise energy matrix precomputation between all pairs of side chains [9], and a minimum energy bound matrix precomputation for MinDEE [9] and iMinDEE; (2) Several rounds of DEE/MinDEE/iMinDEE pruning to reduce the search space; and (3) An A* conformational search [9], [13] of the remaining space. Each design was run in an Intel Xeon machine with at least 4 GB of dedicated RAM and at least 2.50 Ghz of processor speed.
The protein design runs were done using rigid DEE, MinDEE, and iMinDEE. All three algorithms performed an initial steric filter to prune rotamers that could not minimize away from a clash with the template. Implementations of Goldstein DEE [39], Goldstein Pairs, and Split Flags [40] were used for all three algorithms, while Bounds Pruning [9], [41] was used for rigid DEE and MinBounds Pruning for MinDEE and iMinDEE [9]. iMinDEE was run with an initial interval value for all the mutation searches. was chosen based on the minimum difference between the lowest-energy bound and the lowest minimized energy found in the first run.
To evaluate molecular energetics we used an energy function very similar to the energy function used for our previously described, empirically successful protein designs [2]–[4]. The energy function is composed of the following energy terms: (1) attractive-repulsive van der Waals forces, and coulombic electrostatics with a distance-dependent dielectric from the AMBER energy function [37]; (2) implicit solvation terms from the Lazardis Karplus EEF1 solvation model to account for the hydrophobic effect [24]; and (3) entropic penalties [10], [42] and reference energies [15] to account for entropy and energetics of the unfolded protein state. The total energy for a protein structure was calculated by computing a linear combination of all the energy terms, using weightings for the terms as described below.
The weighting of each energy term is important for accurate results and most successful protein designs perform some training of the energy parameters [3], [16], [29]. We trained our energy function by performing protein core redesigns on 9 structures from the PDB database that were not in the set of 69 structures used in this study. The structures for the training set (PDB ids: 1fus, 1ifc, 1lkk, 1plc, 1poa, 1rro, 1whi, 2rhe, and 2trx) were selected from the Richardson's Top 100 database of high-quality curated protein structures [43]. All of them were reprotonated according to the PDB v3 [33] standard and energy minimized with Sander [37]. Residues with less than 20% SASA were selected to mutate; the low-SASA residues were split into groups of 10–15 highly-interacting residues each.
Training was performed by redesigning each group of low-SASA residues with rigid DEE/A* and allowing each amino acid to be mutated to the same 5–7 amino acids allowed in the design runs, which were based on the BLOSUM62 matrix [36]. In addition, each wild-type rotamer was added to the rotamer library. Each redesign was first run using 21 different coarse parameter combinations of solvation and dielectric constant defined by a grid with solvation and dielectric constant. The optimal value found was solvation and dielectric. We then set solvation to 0.5 and dielectric to and performed a local minimization by scaling atom radii. Scaling down the radii of atoms decreases the effect of the repulsive term in the van der Waals energy term. We used scales. The optimal atom radii scaling factor was determined to be .
Each of the 69 protein systems used in our runs was manually analyzed for ligands or co-factors that appeared close to core-residues. Structures with ligands or co-factors in close contact to the mutable design residues were not considered, because functional residues tend to be optimized for functionality and not to stabilize the monomeric structure [17]. 43 protein structures remained after removing those with interacting ligands or co-factors. Each mutation search was set up so that approximately 12–15 core residues of the protein would be mutable. Core residues were chosen by finding all residues with a side-chain relative solvent accessible surface area (SASA) less than either 5%, 10%, or 20%. SASA values were determined with the program NACCESS [35]. If a protein had less than 12 residues with SASA, only these residues were allowed to mutate. Each mutable residue was allowed to take on its wild-type identity and 5–7 other amino acid types. The mutant amino acid types were determined by finding the 5–7 most likely amino acid type substitutions based on the BLOSUM62 matrix [36]. The native rotamers were not included in the native sequence recovery experiments. Native sequence recovery was then performed on the 43 proteins with PDB ids: 1lni, 1ok0, 1psr, 1t8k, 1u2h, 1usm, 1wxc, 1zzk, 2cov, 2fhz, 2hs1, 2r2z, 3d3b, 3dnj, 1l9l, 1r6j, 1u07, 1ucs, 1vbw, 1y6x, 2hin, 2j8b, 2p5k, 2wj5, 3g21, 3hfo, 3jtz, 1aho, 1f94, 1oai, 1vfy, 2b97, 2cc6, 2cg7, 2dsx, 2fma, 2gom, 2hba, 2hlr, 2ic6, 3g36, 3i2z, and 1i27.
|
10.1371/journal.pcbi.1000672 | Conditions for the Evolution of Gene Clusters in Bacterial Genomes | Genes encoding proteins in a common pathway are often found near each other along bacterial chromosomes. Several explanations have been proposed to account for the evolution of these structures. For instance, natural selection may directly favour gene clusters through a variety of mechanisms, such as increased efficiency of coregulation. An alternative and controversial hypothesis is the selfish operon model, which asserts that clustered arrangements of genes are more easily transferred to other species, thus improving the prospects for survival of the cluster. According to another hypothesis (the persistence model), genes that are in close proximity are less likely to be disrupted by deletions. Here we develop computational models to study the conditions under which gene clusters can evolve and persist. First, we examine the selfish operon model by re-implementing the simulation and running it under a wide range of conditions. Second, we introduce and study a Moran process in which there is natural selection for gene clustering and rearrangement occurs by genome inversion events. Finally, we develop and study a model that includes selection and inversion, which tracks the occurrence and fixation of rearrangements. Surprisingly, gene clusters fail to evolve under a wide range of conditions. Factors that promote the evolution of gene clusters include a low number of genes in the pathway, a high population size, and in the case of the selfish operon model, a high horizontal transfer rate. The computational analysis here has shown that the evolution of gene clusters can occur under both direct and indirect selection as long as certain conditions hold. Under these conditions the selfish operon model is still viable as an explanation for the evolution of gene clusters.
| Genes involved in a common pathway or function are frequently found near each other on bacterial chromosomes. A number of hypotheses have been previously presented to explain this observation. A particularly influential theory is the selfish operon model, which posits that horizontal transfer could promote gene clustering by favouring transfer of arrangements of genes that are close together. Subsequent theoretical development and analysis of genomic data have contributed to the debate about the plausibility of this model. Here, by re-examining the evolutionary dynamics of gene clusters, we provide and discuss conditions under which gene clusters can evolve. We find that first, some form of bias for clustering is required for clusters to evolve. This bias can be in the form of bias in horizontal transfer towards genes that are close together, or direct natural selection for gene proximity. Our computational work does not present a theoretical obstacle to the selfish operon model as a possible explanation for the evolution of gene clusters.
| A conspicuous feature of bacterial genomes is the grouping of genes involved in a metabolic pathway into functional units on the chromosome. Early linkage studies of Escherichia coli and Salmonella typhimurium showed that genes in the biosynthetic pathways of tryptophan and histidine occur on a contiguous region of the genome [1],[2]. Furthermore, genes are often found in their biochemical reaction order [3]. Gene clustering has since become recognized as a widespread feature of bacterial genomes. Grouped genes are sometimes transcribed together as an operon, with shared promoter and operator sequences (for example the galactose operon galETK [4],[5]). Regulatory genes have also been found close to the genes they regulate. A classic example is the lacI repressor gene, which resides near but not within the lacZYA operon in Escherichia coli. The extent of gene clustering is variable – a given set of related genes may be clustered in one species but unclustered and/or reordered in another [6],[7]. Interestingly, most clusters do not contain much intergenic DNA, and in some cases genes even overlap [8],[9].
A number of explanations for clustering have been considered over the years. The most controversial and influential hypothesis has been the selfish operon model, which offers a mechanism for the evolution of clustering without needing to invoke the action of natural selection [10],[11]. In this model, gene clusters persist because the proximity of the genes in question facilitates their collective transfer between species. It applies to genes encoding accessory functions rather than essential genes.
Another model that does not require direct selection to explain clustering is the persistence model [12]. Unlike the selfish operon model, this model has been proposed to explain the clustering of essential genes – genes that are evolutionarily persistent. The hypothesis here is that by occupying less space, clustered genes are less likely to be disrupted by the deletion or insertion of DNA. In other words, an individual with clustered genes is more “resilient” to the lethal or deleterious effects of mutation. This hypothesis is similar to the idea that genes sharing regulatory sequences by residing in a single operon present a smaller target for deleterious mutation than scattered genes with individual control elements [13].
Hypotheses involving direct selection have also been examined. Here, clustering of genes confers a direct fitness advantage to the organism. For example, a scenario in which selection directly favouring the co-regulation of genes can lead to the evolution of operons has been outlined [14]. Apart from efficiently regulated transcription, a fitness advantage may arise through shorter diffusion times for proteins finding their targets when the genes encoding them are clustered. Thermodynamic models have been developed to apply this idea to enzymes and transcription factors [15],[16]. The efficiency gained from shorter diffusion times is assumed to translate into a reproductive fitness advantage [17]. Another mechanism conferring advantage to gene clustering is gene amplification [18]. In this model, gene dosage is rapidly and reversibly increased by tandem duplication of the genes in question. The closer the genes are, the greater the probability of coamplification. The increased dosage is assumed to contribute to elevated fitness. Other models for the evolution of gene clusters based on metabolic arguments have also been studied [19],[20].
Other hypotheses have been considered but rejected [10],[17, for example]. A hypothesis now called the natal model suggests that clusters arose by gene duplication and divergence such that the newly formed genes participate in a common pathway. However, the lack of sequence homology for most genes within clusters undermines this explanation [10]. Fisher's theory of the evolution of linkage and recombination has been suggested to apply to bacteria [1],[21]. Under this theory natural selection favours increased linkage among co-adapted genes – genes whose products work well together – because recombination (chromosomal crossover during meiosis) breaks up combinations of alleles with high fitness. However, it has been pointed out that this requires high recombination rates, which are typical for eukaryotes, to work [10]. Although recombination rates are found to be high in some species [22],[23], they are not high enough relative to the cellular generation rate to support an account of clustering based on Fisher's theory.
The debate on the origins and maintenance of gene clusters continues, with recent genomic studies casting doubt on the selfish operon hypothesis. First, the prediction that non-essential genes are clustered while essential genes are not has been tested and rejected [24]. Second, if horizontal gene transfer is an important source of gene clusters, then horizontally transferred sequences should be associated with operons. Genomic data, however, do not support such an association [14]. On the other hand, they do support the possibility that genes and their regulators may have evolved close proximity via horizontal transfer [25]. Third, the selfish operon model is unable to explain the observation that genes in clusters are sometimes arranged in the order of biochemical reactions. A resolution may involve multiple mechanisms, of which horizontal transfer of selfish operons is one [12].
Here, we re-examine the theoretical basis for explaining the origins and maintenance of gene clusters. By studying a number of distinct models, we provide and discuss conditions under which clustering can evolve.
We describe three kinds of models for gene clustering in this article. First, we revisit the selfish operon model [10]. We seek to explore the parameter space and understand in more detail when and why it produces gene clusters. Second, we propose a model based on the Moran process, which tracks individual bacterial cells and in which the total population size is constant. Third, we develop a further model that tracks the substitution of new arrangements, making the assumption that populations are monomorphic. By running computer simulations of these three systems we consider the factors that lead to the evolution of gene clusters.
The assumptions common to all models are as follows. Genomes are made up of circular chromosomes divided into regions; we let kilobases (kb). This genome size is constant over time. There are genes in the pathway of interest. Only a single gene can occupy any given position. The units of reproduction are either species or individual bacteria depending on the model. A genome can undergo rearrangement with probability per step or generation. We explore two processes: first, translocation of a random gene to a random position and second, inversion by which two breakpoints are chosen randomly uniformly and the intervening segment inverted. If the resulting arrangement moves the terminus or origin more than kb the new arrangement is regarded as lethal [26],[27]. Both translocation and inversion are used within the selfish operon framework of Lawrence and Roth 1996, while only inversion is considered for the Moran model and the rearrangement substitution model.
In their influential paper, Lawrence and Roth describe a simulation model that produces gene clusters through a horizontal gene transfer process that is biased towards genes that are physically closer on a chromosome [10]. This is called the selfish operon model. In this model, species in which individuals carry all the genes needed for the function are called “positive” species. Each species is assumed to be monomorphically composed of genomes with a particular arrangement of genes on the chromosome, and fixation is assumed to occur instantaneously. That is, each species is associated with a single arrangement of genes. We are interested in the minimum arc distance along the chromosome that contains all genes, which is equivalently the genome length minus the longest interval between pairs of neighbouring genes. The simulation is initialised with 100 species, with each species given a random set of gene positions. Lawrence and Roth kept the number of species between 10 and 900 [10]. We have implemented this by switching off the horizontal transfer process when the number of species reaches 900 and re-instating it when the size goes under 900. We ran our simulations for 50,000 time steps.
Horizontal transfer leads to a species that lacks the function (a “negative” species) acquiring the function along with the arrangement of gene positions of the donor genome. The probability of horizontal transfer decreases with distance . Although its form is not given in [10], we will assume it is exponential with a decay parameter . That is,(1)The exponential distribution is a natural choice for the size distribution of transferred DNA among bacteria, and has been empirically tested for homologous recombination [28],[29]. Some support for a skewed distribution of gene transfer fragment lengths is found in Ochman and Jones 2000 [30].
At each time step, each species or individual can undergo loss of the function with probability . Following Lawrence and Roth, we set the loss probability to 0.001 per genome per time step and the maximum probability of horizontal transfer , occurring when the genes are located in the same minute of the chromosome, to 0.01 per genome per time step [10]. We set by default, under which a 50 kb fragment is 6 times more likely to transfer than one of 500 kb. Because the probability of rearrangement is likely to be very low in nature [31], we set per genome per time step by default. Lawrence and Roth 1996 used a much higher value of and we investigate the effect of lowering this parameter from this high value. We studied the effect of varying and by varying parameters one at a time as well as using latin hypercube sampling [32],[33] to explore the parameter space. Under this methodology, each parameter is divided into equiprobable regions in the area of interest, and parameter sets are constructed by selecting values randomly from the resulting grid without replacement. A uniform distribution was used for each parameter.
The algorithm we used for the dynamic is as follows.
One problem we have noticed with this model is that given a rearrangement event, the genes in question are always affected. A more natural assumption would be that the genes in question are affected with probability , which is the proportion of the genome occupied by the genes assuming that genes are 1 kilobase in length. Thus, we have also run the simulations using this corrected translocation process, replacing step 2(a) in the above algorithm with
This correction effectively lowers the rearrangement probability by a few orders of magnitude.
We have also implemented a version of the model in which rearrangement occurs by inversion instead of translocation. Here, we replace step 2(a) in the algorithm with
If ( and ) or ( and ) or () then the inversion is viable. (Recall is the tolerance to imbalance between origin and terminus.) For each gene whose location is between and , move it to its new location given by .
We construct a model in which the population evolves according to a Moran process [34],[35] combined with a process of genome inversion. Here, we track a population of bacterial cells. As with the selfish operon model, we consider a pathway involving 3 or more genes. A population is initialised with all bacteria carrying the same genome with genes placed randomly uniformly on the chromosome. The population size is . Let represent the relative fitness of cells with the genes at minimum arc distance . Genomes with the genes closer together have a reproductive or survival advantage over those with the genes further apart. We use the function to describe this relationship. Because this relative fitness function is analogous to , we use the same symbol () to describe the decay in fitness with respect to distance . An alternative function was also used to ascertain the effect of using a steep sigmoidal relationship. Selection for clustering here can be due to any of the mechanisms discussed in the Introduction.
The algorithm is as follows.
Following the classical definition of the Moran process, a single generation is time steps.
This process is very slow with high population sizes, particularly when the rearrangement probability is low. The computational demands of running these simulations precluded the possibility of systematically analysing sensitivity to parameters. This motivated us to develop a further model, which tracks the mutation and fixation process without following details at the population level. This model is described in the next subsection.
Here, the population is monomorphic (except during periods of substitution of new arrangements) and so only a single genome arrangement is tracked. Again, the genes in the pathway in question can occupy positions, represents the population size and is the rearrangement probability. The assumption that the population is monomorphic implies that must not be too large. In each generation the probability of a rearrangement occurring in at least one individual is which can be approximated with since is small. The time until the next rearrangement event is distributed geometrically with parameter . We use inversion rather than translocation as the source of rearrangements.
As above we specify selection through an exponential decay in fitness as a function of the minimum arc distance , so that the relative fitness of a new genome with distance is , and the selective coefficient is . A new arrangement fixes in a population with probability(2)and the time it takes to reach fixation is(3)These quantities have been derived from diffusion theory in population genetics (for details see [36]). We use in place of to apply the theory to haploids, where is the effective population size of a diploid population.
The algorithm for the rearrangement substitution model is therefore as follows.
was set at 50,000 generations. We investigated this model by varying one parameter at a time as well as using latin hypercube sampling to explore the parameter space.
When three genes are placed randomly around a chromosome with a uniform distribution, the average minimum arc distance between them is around 1900 kb. When the rearrangement probability is or , the selfish operon model [10] produces an initial wave of gene clustering down to around 600–800kb ((Figure 1A), also reflected in the rise of the proportion of genomes that are clustered under a threshold (Figure 1B). The maximum population size of 900 is reached quickly (Figure 1C) and the dynamics of clustering undergo a change as a new population dynamic regime sets in. When the rearrangement probability is high, clusters break up until the average minimum arc distance settles on high values (Figure 1A). In these cases, the selfish operon model fails to maintain tight clustering in the long term. In particular, gene clusters do not evolve under the parameter values used by Lawrence and Roth [10].
To determine if there are conditions under which the selfish operon model does produce clustering, we re-examined this model by exploring its parameter space. Figure 2 reveals the effect of varying the parameters in this model on the average minimum arc distance. It shows that under the original model clustering is only produced when the rearrangement probability is low, the number of genes is small, and the maximum transfer probability is sufficiently high. Under the corrected translocation process, the effective rearrangement probability is lowered by a factor and the probability itself has no apparent effect on clustering. The decay in transfer probability (see Equation 1) must take intermediate values of around for clustering to evolve. If is too low, selection is too weak to promote clustering while if it is too high, the probability of transfer is depressed for most minimum arc distances, preventing selection from acting effectively.
Very similar results are observed when translocation is replaced by inversion, as shown by varying one parameter at a time as well as by latin hypercube sampling analysis (Figure 3). The major difference is that a high probability of inversion does not prevent the evolution of clusters to the same extent as observed in the uncorrected translocation process of Figure 2A.
We further explored the evolution of clustering using the Moran model with selection for gene clusters. By holding the population size constant this model also allows us to disentagle the effects of population dynamics from those of rearrangement and selection. Figure 4 shows simulation runs of the process for progressively lower values of : . It was not computationally feasible to run the simulation under even lower, and more realistic, values. The general pattern emerging from these sample trajectories is that the minimum arc distance is reduced through a series of selective sweeps. The time taken until the appearance of a rearranged genome that reaches fixation is long and depends on the rearrangement probability and the population size . The reduction of minimum arc distance is a slow process made even slower by lowering . Using a steep sigmoidal function for selection instead of exponential decay (Figure 4D) gave qualitatively similar results.
The rearrangement substitution model, which “compresses” time by tracking fixation events, is amenable to sensitivity analysis. Figure 5 demonstrates that a low rearrangement probability of is able to produce clustering in 50,000 generations. Even lower probabilities lead to weak or no clustering because successful rearrangements that reduce the distance between genes are too rare. Increasing the population size improves the efficiency of selection and leads to clustering. Similarly, increasing the decay in fitness improves clustering. Gene clusters are also more readily formed for pathways with a smaller number of genes .
Similar results are produced when the parameter space is explored using latin hypercube sampling (Figure 6). Minimum arc distance decreases with and and increases with . Distance also decreases with , though this effect is subtle. For (panel B) and (panel C) the correlation with distance is statistically detectable using a non-parametric method (Kendall's tau), with -values of and 0.0148 respectively. The corresponding -values for (panel A) and (panel D) were both less than . Note that each factor on its own does not explain much of the variation in distance.
This study presents new computational models showing that direct natural selection can lead to the formation of gene clusters under appropriate conditions. We have also re-examined an existing simulation model involving indirect selection – the selfish operon model. By exploring these models under many conditions, we have identified the regions in parameter space that produce gene clustering. In the following, we will discuss parameters as rates rather than probabilities per time step.
|
10.1371/journal.pgen.1000155 | Double Strand Breaks Can Initiate Gene Silencing and SIRT1-Dependent Onset of DNA Methylation in an Exogenous Promoter CpG Island | Chronic exposure to inducers of DNA base oxidation and single and double strand breaks contribute to tumorigenesis. In addition to the genetic changes caused by this DNA damage, such tumors often contain epigenetically silenced genes with aberrant promoter region CpG island DNA hypermethylation. We herein explore the relationships between such DNA damage and epigenetic gene silencing using an experimental model in which we induce a defined double strand break in an exogenous promoter construct of the E-cadherin CpG island, which is frequently aberrantly DNA hypermethylated in epithelial cancers. Following the onset of repair of the break, we observe recruitment to the site of damage of key proteins involved in establishing and maintaining transcriptional repression, namely SIRT1, EZH2, DNMT1, and DNMT3B, and the appearance of the silencing histone modifications, hypoacetyl H4K16, H3K9me2 and me3, and H3K27me3. Although in most cells selected after the break, DNA repair occurs faithfully with preservation of activity of the promoter, a small percentage of the plated cells demonstrate induction of heritable silencing. The chromatin around the break site in such a silent clone is enriched for most of the above silent chromatin proteins and histone marks, and the region harbors the appearance of increasing DNA methylation in the CpG island of the promoter. During the acute break, SIRT1 appears to be required for the transient recruitment of DNMT3B and subsequent methylation of the promoter in the silent clones. Taken together, our data suggest that normal repair of a DNA break can occasionally cause heritable silencing of a CpG island–containing promoter by recruitment of proteins involved in silencing. Furthermore, with contribution of the stress-related protein SIRT1, the break can lead to the onset of aberrant CpG island DNA methylation, which is frequently associated with tight gene silencing in cancer.
| Human cancers contain epigenetic changes as well as DNA mutations that play a role in abnormal silencing of tumor suppressor genes. In contrast to DNA mutations that change the sequence of DNA, epigenetic changes cause abnormal silencing of genes through DNA methylation via the addition of methyl groups to DNA and through modifications to the associated chromatin proteins. One important event in tumor initiation and progression is the exposure of cells to DNA damage during events such as chronic inflammation and carcinogen exposure. We hypothesized that such damage may play a role in producing chromatin alterations, which could initiate epigenetic silencing of tumor suppressor genes. Here we show, using an exogenous gene promoter model, that key proteins involved in epigenetic silencing are recruited to the DNA near a double strand break. Occasionally, sustained localization of these proteins to the gene promoter leads to silencing of the associated gene and to the seeding and spreading of DNA methylation within the promoter that further stabilizes the silencing. This finding suggests that DNA damage may directly contribute to the large number of epigenetically silenced genes in tumors.
| Chronic inflammation along with aging causes an increase in reactive oxygen species that induces DNA damage in the form of base oxidation, single stand breaks, and double strand breaks (DSBs) [1]. Errors in DSB repair can cause mutations and chromosome instability that lead to cancer or cell death [2]. In response to DSBs, cells undergo cell cycle arrest or apoptosis. Cell cycle arrest gives the cell time to repair the damage utilizing repair proteins that are recruited to the site of damage and activated. DSBs are repaired by either homologous recombination (HR) or nonhomologous end joining (NHEJ) [3]. The pathway followed to repair DSBs is determined by the location in the cell cycle and the type of cell [4].
The above repair processes occur in DNA that is often packaged in highly organized, mostly condensed chromatin, which also consists of histones and histone-associated proteins. Chromatin structure and dynamics regulate the genome such that non-desirable transcription is repressed [5]. This chromatin structure is determined by modifications of histone tails by acetylation, methylation, and phosphorylation in patterns which have been termed the histone code [6]. In general, acetylation of lysine residues induces an open chromatin configuration associated with gene expression, whereas deacetylation induces closed, compact chromatin associated with transcriptional repression. The amino-terminal tails of both histones H3 and H4 contain several lysine residues that can be acetylated by histone acetyl transferases (HATs) and deacetylated by histone deacetylases (HDACs) [7],[8]. Acetylation neutralizes the positive charge of the lysine residues and changes the structure of the histone, likely affecting the interaction of these histones with both proteins and DNA [9]. Specifically, mutational studies have indicated that lysine 16 of histone H4 (H4K16) and lysines 9, 14, and 18 of H3 are critical in silencing and are all acetylated in active chromatin and hypoacetlyated in transcriptionally-repressive chromatin [9],[10]. Histone methylation also plays a role in chromatin dynamics with mono-, di-, and tri-methylation of H3K4 being associated with active chromatin, and alternatively with mono-, di-, and tri-methylation of H3K9 and H4K20 and di- and tri-methylation of H3K27 being associated with closed chromatin and gene silencing [11]–[13].
It has become increasingly apparent that DNA repair must be intimately involved with regulation of chromatin. For the repair of DSBs there is an access, repair, restore (ARR) hypothesis wherein chromatin after a DSB is first modified to generate an open chromatin structure allowing access to the DNA by repair proteins [14]. Additionally, specific modifications of chromatin may be necessary for components of the DNA repair or checkpoint machinery to recognize damaged DNA [15]. In support of this step, in yeast, both HATs and chromatin remodeling complexes are recruited to DSBs [16]–[19]. Repair of the break then occurs, followed by the need to restore the chromatin back to its original, more condensed state [14],[20]. Condensed chromatin might also prevent transcription and/or replication machinery from accessing the DNA and/or interfering with the repair process [21],[22]. Additionally, condensed chromatin may play a role in ending the DNA damage response signaling cascade [23]. Restoration of the chromatin around a break suggests that silencing factors such as HDACs and histone methyl transferases (HMTs) might be recruited to the area of DSBs [19]. Additionally, DNA methylation patterns also need to be restored, suggesting a possible role for DNA methyltransferases (DNMTs) in DSB repair [24].
One example of a histone-modifying protein involved in DNA repair in the yeast S. cervisiae is Sir2, a NAD+ dependent protein and histone deacetylase [25]. The family of yeast sirtuins (Sir2-4) has been shown to be involved in telomeric silencing, silencing at the mating-type locus, and DSB repair [26],[27]. In telomeres Sir2-4, with the help of Rap1, a telomere DNA-binding protein, polymerize across nucleosomes by binding to the histone tails of H3 and H4 to create an inactive heterochromatin state causing silencing of the region [28]. In response to activation of the DNA damage checkpoint pathway Sir2-4 are recruited from the telomeres to the DSB where they facilitate end-joining to such an extent that yeast with mutations in any of the Sir proteins have a defect in NHEJ [29],[30]. Sir2 specifically modifies chromatin by deacetylating H4K16 and H3K9 [31]. In the area of a defined DSB in yeast there is an increase followed by a decrease in H4K16 acetylation [19]. Localization of Sir2 to the region occurs in the same time frame as the decrease in acetylation, suggesting that Sir2 is responsible for the deacetylation [19]. Additionally, mutations of four lysine residues on the histone H4 tail increase the sensitivity of yeast to DSB-inducing agents [32]. In mammalian cells, acetylation of H4 also seems to play a role in DSB repair because TRRAP (Transactivation-transformation domain-associated protein)/TIP60 (HIV Tat-Interacting Protein, 60 kDa) dependent acetylation of H4 occurs immediately after a DSB [33]. TIP60 binds to the chromatin around the DSB and plays a role in chromatin relaxation required for the efficient recruitment of repair factors as well as repair of the DSB [33]. Mutants that lack this ability accumulate DSBs following exposure to gamma-irradiation [34]. After repair of the DSB, there may be a need to deacetylate H4 to return the acetylation levels back to normal.
SIRT1, the mammalian homolog of Sir2, mediates transcriptional repression, heterochromatin formation, heritable gene silencing, p53 function, and lifespan [35]–[39]. SIRT1 has been shown to be involved in the maintenance of silencing associated with abnormal promoter region CpG island DNA methylation in tumor suppressor genes [39]. SIRT1 is localized to the promoters of these methylated and silenced tumor suppressor genes, but not to promoters of the same genes in cell lines where they are normally maintained in an unmethylated, open chromatin state facilitating gene expression [39]. Inhibition of SIRT1 caused re-expression of these genes along with a corresponding increase in H4K16 and H3K9 acetylation and SIRT1 recruitment to an artificial promoter via a gal4 DNA binding site mediates transcriptional repression, H4K16 deacetylation, and an increase in H3K9me3 [40]. Additionally SIRT1 has been found in a stem/precursor cell and/or “transformation specific” polycomb group (PcG) complex (PRC4) containing Enhancer of Zeste Homologue 2 (EZH2), the enzyme catalyzing H3K27me3 and H1K26me [41],[42], and Eed2 [43]. Previously, EZH2 had been identified as part of the PRC2/3 complex that plays a role in the initiation of chromatin silencing during development [41],[42]. SIRT1 is also linked to the increased methylation of H3K9 because SIRT1 has been shown to bind to, and increase the activity of, the suppressor of variegation 3–9 homologue (SUV39H1), a HMT that tri-methylates H3K9 [44]. These findings suggest that the recruitment of SIRT1 (and possibly EZH2) to the promoter of a gene can induce gene silencing via closed chromatin and that the continual presence of SIRT1 helps maintain the silencing.
In this study, we demonstrate the recruitment of silencing factors to a DSB induced in a model exogenous construct containing the CpG island region of the E-cadherin (E-cad) promoter, which is often aberrantly silenced and DNA hypermethylated in human cancer [45]. After an induced break, both SIRT1 and EZH2 are transiently recruited to the area surrounding the break. Their recruitment corresponds, following an initial increase, to a decrease in H4K16ac and an increase in H3K27me3. Additionally, DNMT1 and DNMT3B are also transiently recruited to the break site. By inducing DNA damage and then selecting for silencing of the HSVTK gene, driven by the E-cad promoter in our system, we demonstrate occasional gene silencing and onset of DNA methylation in the CpG island area. Moreover, the induced DNA methylation and recruitment of DNMT3B appear to be dependent on the presence of SIRT1 during the initial break and repair cycle.
In order to induce a defined DSB in mammalian cells, we utilized the homing endonuclease I-SceI that has an 18 base pair recognition sequence [46]. MB-MDA-231 cells were first transfected with a tetracycline (tet) repressor plasmid and a tet operon plasmid that drives the expression of the hemagglutinin (HA)-tagged I-SceI endonuclease (Figure 1A). A single clone was selected that had no basal level of HA-I-SceI expression but had a high level of tet-induced expression. This clone was stably transfected with a plasmid that contains a consensus I-SceI cut site inserted into a copy of the E-cad promoter, containing a CpG island often DNA hypermethylated in multiple human tumor types including the MB-MDA-231 cell line [45]. The promoter drives the expression of the herpes simplex virus gene, thymidine kinase (HSVTK). A single copy clone was then tested for inducible expression of the enzyme by adding tet for 4 hours followed by washing out the tet and collecting cells at indicated intervals (Figure 1B). By RT-PCR analyses, HA-I-SceI was induced after 4 hours of tet treatment, and this expression was maintained after a 4 hour wash (Figure 1C). By immunofluorescence each cell was shown to express high levels of nuclear HA-I-SceI protein at the 4+4 hour time point (Figure 1D).
To determine the timing of the DSB formation and repair, the HA-I-SceI induced breaks were monitored by a PCR assay with primers spanning the cut site. Using this PCR, only uncut or repaired DNA will result in a PCR product. The PCR product was slightly decreased at the 4 hour time point, followed by a more substantial decrease at the 4+4 hour time point, which corresponds to the induction of the enzyme by RT-PCR (Figure 1E). The PCR product level increased at the 4+24 hour time point suggesting that a significant portion of the cells repair the DNA break during this time frame. Within minutes of damage, H2AX is phosphorylated on its C-terminal residue serine 139 at the site of DNA damage [47]. Phospho-H2AX plays a role in stabilizing repair foci containing DNA repair factors, and the mark is maintained at the break site until the break is repaired [47],[48]. Therefore, phospho-H2AX foci are a way to monitor DNA damage and repair. By chromatin immunoprecipitation (ChIP) following induction of the cutting enzyme, phospho-H2AX was localized to the DNA near the DSB at the 4+4 hour time point (Figure 1F). The presence of phospho-H2AX also suggests that not only did the break occur but that the chromatin around the break was modified as expected. Also, because both the greatest phospho-H2AX localization and the greatest amount of cut DNA as determined by PCR occur at the same time point, 4+4 hours, it suggests that this time point represents the immediate response to double strand breaks in contrast to a response to unrepaired persistent lesions that might be present at later time points [49].
As part of the ARR model of DSB repair, the restoration phase may require the recruitment of histone marks indicative of closed chromatin, as well as the proteins responsible for establishing these histone marks [14],[19],[20]. We examined the enrichment of histone marks and the recruitment of chromatin-binding proteins after inducing the DSB to determine if chromatin takes on characteristics of closed chromatin after DNA damage. Using protein from the sonicated material for ChIP, we confirmed that the HA-I-SceI enzyme was expressed at the 4 hour and 4+4 hour time point, corresponding to induction of phospho-H2AX at the 4+4 hour time point (Figure 2A). As previously introduced, SIRT1 is a stress response protein associated with DNA repair in yeast [19],[50] and transcriptional repression [39], and is a component, in drosophila and mammalian cells, of the PcG silencing complex PRC4 [43]. SIRT1 recruited transiently to the DNA near the break (SCE PCR) increased from the 4 hour time point to the 4+16 hour time point. The highest recruitment levels correspond to when the DNA begins to be repaired in our experimental design (Figure 2B – right lower panel). A lesser and earlier increase occurred at the downstream TK gene site which persisted to a varying degree over 24 hours (Figure 2B – right lower panel). Importantly, H4K16ac, the residue that SIRT1 is known to deacetylate, shows an early increase in enrichment at 4 hours followed by a decrease at the 4+4 hour time point, particularly at the SCE site (Figure 2B – upper right panel). The most significant decrease in the enrichment of H4K16ac corresponds to the sharp increase in SIRT1 recruitment at the 4+16 hour time point (Figure 2B – compare right upper and lower panels). We also looked for other silent chromatin marks at the break site. Importantly with respect to participation of SIRT1, we demonstrated a strong enrichment of H3K27me3, the mark catalyzed by the EZH2 enzyme in PRC4 in the absence of histone H1, again primarily to the area near the break site. There was also a less substantial enrichment of the repressive mark K9H3me2 at the SCE region at the 4+16 hour time point and in the TK region at the 4 hour time point of I-SceI induction. In addition, H3K9me3 increased sharply in the same TK region at the 4+4 hour time point (Figure 2C).
After DNA repair, in addition to changes in and restoration of histone modifications, we were particularly interested in the possible recruitment of DNMTs to the promoter after DNA damage because DNA methylation is abnormally increased at the E-cad promoter in many cancers [45]. In previous studies, using a model of UVA laser microirradiation, the DNA methylation catalyzing enzyme DNMT1 has been shown to be localized grossly to the regions irradiated immediately following damage [24]. Therefore, we looked for localization of this maintenance DNA methylation enzyme plus the de novo DNMT, 3B, to the break site. DNMT1 was localized to the break, in modest increases, mostly at the 4+4 hour time point for both the SCE and TK regions and interestingly this enrichment is maintained at the SCE site only at later time points (Figure 2D). On the other hand, DNMT3B was localized to the break site only early in the time course. Enrichment was demonstrated at the 4 hour time point only at the SCE region and at both the SCE and TK regions at the 4+4 hour time point when the I-SceI enzyme expression is the highest and the DNA is undergoing cutting.
We next looked to further understand the potential interactive roles of the demonstrated recruitment of SIRT1, the DNMTs, and histone modifications to the promoter in the function and DNA methylation of the promoter. We initially focused on the role of SIRT1 in the kinetics of break repair by knocking down levels of this protein. By western blot, overall levels of cellular SIRT1 were significantly knocked down in SIRT1 small interfering RNA (siRNA) treated cells versus the non-target (NT) treated cells (Figure 3A). In addition to its role in deacetylating histones SIRT1 can also deacetylate p53, and we used this latter modification to further monitor the efficacy of our knockdown. In the SIRT1 knockdown cells there is a significant increase in acetyl lysine 382 p53 consistent with SIRT1 depletion. Importantly, in MDA-MB-231 cells, p53 contains a point mutation that makes the protein non-functional, so this increase in acetyl p53 has no functional consequence [51].
In our SIRT1 knockdown studies, it is first important to note that the kinetics of SIRT1 recruitment to the DSB is somewhat different from those shown in Figure 2, possibly because the rounds of transfection necessary for the siRNA knockdown additionally stress the cells. Thus, in the non-target control (NT) cells, SIRT1 recruitment is seen at the SCE and TK sites (Figures 3B and C) earlier than in the studies in Figure 2B, peaking at 4 hours, and being maintained over 48 hours. Even though, in the SIRT1 knock down studies, there is a striking reduction of overall levels of the cellular SIRT1 protein (Figure 3A), the reduction at the SCE and TK sites, relative to that in the control NT cells, was less severe (Figure 3B and C). However, this reduction did correlate with changes in levels of H4K16ac. Thus, overall, SIRT1 localization appeared delayed at the promoter region in the SIRT1 knockdown cells as compared to the NT cells, and this correlated with sustained enrichment of the H4K16ac mark early after I-SceI induction and lasting through the 4+16 hour time point (Figure 3B). Importantly, at the 4+24 hour time point where we see late enrichment of SIRT1 at the SCE site in the knockdown cells, H4K16ac levels are again reduced, suggesting that the level of acetylation of H4K16 is dependent on SIRT1 recruitment to the break site and the SIRT1 knockdown has a functional consequence. In addition, there was a modest early decrease (4 and 4+4 hour time points) of SIRT1 at the TK site, and this correlated with increased H4K16ac at the 4+4 hour and 4+16 hour time points (Figure 3C). The most informative local result for SIRT1 knockdown appears to be the levels of H4K16ac as discussed above. The persistent high levels of H4K16ac recruitment through the 4+16 time point in the knockdown cells demonstrate the effect of the SIRT1 knockdown at the chromatin near the break site.
We performed ChIP for phospho-H2AX in knockdown cells to see if the kinetics of phospho-H2AX recruitment to, and removal from, the break site were altered (Figure 3D). In the SIRT1 knockdown cells, the levels of phospho-H2AX recruitment at the SCE site were distinctly diminished by ChIP but had the same overall time frame of recruitment and loss as in the NT cells. Importantly, for the levels of knock down of SIRT1 achieved, and with the delayed recruitment of SIRT1 to the break site, there did not appear to be a significant effect on the kinetics of repair as analyzed by our PCR assay using primers that are on either side of the cut site (Figure 3E).
We next examined the potential role of SIRT1 in recruitment of the PcG mark H3K27me3 and the recruitment of EZH2, the enzyme responsible for catalyzing the mark [42]. In the NT cells, EZH2 was enriched in the promoter and in the body of the gene at the 4 hour time point and, to a greater extent, at the 4+4 hour time point in the promoter (Figure 4A and 4B). H3K27me3, correspondingly, was enriched in the promoter and the body of the gene at these time points, directly corresponding to the localization of EZH2. Interestingly, in the SIRT1 knockdown cells there was an increased enrichment of EZH2 in the promoter at 4 hours as compared to NT knockdown cells, but similar levels of H3K27me3 over the entire time course (Figure 4A and 4B). In contrast, in the TK gene, EZH2 enrichment was sharply increased at the 4 and 4+4 hour time points in the SIRT1 knockdown as compared to the NT cells and, correspondingly, the H3K27me3 mark was greater at all time points in the body of the gene. Thus, even modest SIRT1 knockdown appears to increase the magnitude of recruitment of EZH2 to the break region downstream to the actual cut site. Recruitment of H3K27me3 in the promoter of both the NT and SIRT1 knockdown cells decreased at 4+16 hours but increased again at 4+24 hours. This later increase may be indicative of some persistent double strand breaks. Alternatively, this increase may reflect altered kinetics observed selectively in the knockdown experiments. Both the NT and SIRT1 knockdown cells have similar H3K27me3 enrichment whereas the non-siRNA treated cells (Figure 2) show no enrichment for this modification at this later time point.
To determine whether there is any dependence, on SIRT1, of DNMT1 and DNMT3B recruitment to the DSB, we examined localization in the NT and SIRT1 siRNA treated cells. In the SIRT1 knockdown cells, there was a much increased enrichment of DNMT1 which occurred earlier than in the NT treated cells—at the 4 and 4+4 hour time point—and persisted through the 4+24 hour time point in the promoter and mostly at the 4 hour time point in the gene (Figure 4C and 4D). These data suggest that DNMT1 can be recruited, possibly to an increased degree and with slightly different timing, to the area around the break when SIRT1 levels are reduced. The most striking change, however, was that DNMT3B recruitment, even with the modest change in SIRT1 knockdown, was virtually absent in the SIRT1 knock down cells. These data suggest that SIRT1 may play a role in early recruitment of DNMT3B to the DNA around the DSB.
We next sought to further place the above findings for changes in chromatin surrounding an induced DSB into the context of genes that are DNA hypermethylated and heritably silenced in cancer—and for which our engineered E-cad promoter region provides a model. Despite the dynamic chromatin changes and DNMT recruitment we have outlined above, we saw no evidence during the acute period of DSB repair of any induction of the cancer-related gene silencing events (i.e. loss of TK expression-Figure 1C and/or DNA methylation-data not shown). However, models for how these events may take place in native cancer evolution [52],[53], and experimental models for acute, transient silencing of genes [40],[54],[55] suggest that the transient state of silent chromatin in the gene promoter region results in rare instances wherein the silenced chromatin is maintained to produce seeding of DNA methylation and permanent silencing of the downstream gene. We thus tested this hypothesis for our model.
To look for selection of long term silencing events, we induced our DSB by treating the cells with tet for either 4 hours or 24 hours and then negatively selected the cells for HSVTK silencing via treatment with ganciclovir. Cells that silence the E-cad promoter do not express HSVTK and therefore are not sensitive to ganciclovir, unlike the parental, uncut cell line. After selecting one thousand cells that were either uncut or cut with ganciclovir, no clones from the uncut cells survived whereas ten clones from the 1000 cells plated from the cut cells survived. As an additional control experiment, a cell line containing the inducible I-SceI enzyme and an E-cad promoter without an I-SceI consensus cut site driving the expression of the HSVTK gene was treated with tetracycline followed by ganciclovir as above. No clones survived from this cell line (data not shown). By RTPCR expression levels, HSVTK was transcriptionally silent in the ganciclovir resistant clones from above (Figure 5A). Interestingly, one out of the ten silent clones had a portion of the promoter or gene deleted (data not shown), indicating that improper repair may result in deletion events. The rest of the silent clones appeared to have intact sequences as examined by PCR. Therefore, the frequency of silencing without deletion of HSVTK in our system is 0.9%. We examined the promoters in two of these silent clones by ChIP at passage 5 after ganciclovir selection. There was no enrichment of SIRT1 or change in H4K16ac in the clones as compared to uncut, unselected cells (P), but there was a significant enrichment of DNMT1, DNMT3B and EZH2, along with the silent chromatin marks dimethyl and trimethyl H3K9 and H3K27me3 (Figure 5B). Interestingly, when the promoters were examined by ChIP at a later passage (p34 to p36 after ganciclovir selection) DNMT1 was still enriched at the promoter, but localization of DNMT3B was lost (Figure 5C). These findings suggest that the chromatin in the promoter is indeed in a silent state and that this silent state is accompanied in later passages by the presence DNMT1 but not DNMT3B.
Using a PCR that was specific for the I-SceI containing E-cad promoter, we bisulfite sequenced the promoter to examine the DNA methylation status of the HSVTK silent clones. The parental cell line was almost completely unmethylated (Figure 5D). The HSVTK silent clones showed a varying degree of methylation. HSVTK silent clones originally treated with tet for 4 hours showed very little CpG methylation (data not shown), however, the majority of those treated with tet for 24 hours showed an increase in CpG methylation 3′ to the break site (Figure 5D and 6B). To examine how methylation might change with time in the silenced clones, we bisulfite sequenced increasing passages of two HSVTK silent clones, one without and one with initial DNA methylation (Figure 5D). Clone 1B, which initially had very little methylation continued to have only a scattered change in methylation with passage. Clone 8B had initial methylation just 3′ to the break site and methylation spread with passage towards the actual break site and became quite prominent by passage 30 (Figure 5D). Interestingly, this methylation occurs in the region that is flanked by the ChIP PCR primers in the promoter (SCE), demonstrating that the DNA methylation enzymes are recruited to the region where the methylation is occurring (Figure 5D). To further demonstrate how the DNA methylation changes with cell passage, we calculated the mean number of methylated CpGs per bisulfite-sequenced clone per passage of selected HSVTK silent clones (Figure 5E). Clones with little initial methylation (clones 1B and 6B) showed almost no increase in the mean number of methylated CpGs per bisulfite sequenced clone (Figure 5E). However the HSVTK silent clones with an initial methylation of 3–4 CpGs (clones 3B and 8B) gained methylation with increasing passage (Figure 5E).
To further look at the nature of the relationships between silencing and DNA methylation, we treated one unmethylated and one DNA methylated clone with the DNA demethylation agent 5-deoxy-azacytidine (DAC) or with Trichostatin A (TSA), a type I/II histone deacetylase inhibitor. DAC treatment inhibits DNMT activity and causes re-expression of genes silenced with DNA methylation [56],[57]. It will sometimes cause this response in low expression genes which have no proximal promoter DNA methylation [57]. However, we and others have previously shown that TSA treatment is generally ineffective for re-expression of such silent genes, particularly when the CpG island is densely DNA methylated [56]. TSA can be more effective when the DNA methylation in such genes is partial or minimal [56]. Interestingly, when the HSVTK silent clones are treated with TSA or DAC, the clone that has silent chromatin but no methylation (clone 1B) has re-expression of HSVTK by either treatment (Figure 5F). However, clone 8B, which has silent chromatin and increased DNA methylation, has HSVTK re-expressed by DAC treatment but to a much lesser degree with TSA. These findings suggest that at this later passage of clone 8B the partial DNA methylation plays at least some role in maintaining the silencing of the HSVTK gene.
As demonstrated above, a DSB in the promoter of a gene that is associated with transient recruitment of silencing proteins can, in occasional cells, cause long term silencing of the involved gene. Some partial DNA methylation can also be associated with such silencing and is possibly maintained by the persistence of the maintenance DNMT, DNMT1, in the region. Because we observed in the acute DSB induction studies a transient recruitment of the de novo DNMT, 3B—which would be the best candidate to initiate any DNA methylation—and evidence that SIRT1 may play a role in this recruitment, we sought to determine the significance of these dynamics in long term silencing. To study this, we performed our siRNA knock down of SIRT1 by treating cells with NT or SIRT1 siRNA, followed by treatment with tet for 24 hours, then selection with ganciclovir. Global SIRT1 knockdown levels were similar after tet treatment to those in the studies described earlier (Figure 3A). Both NT and SIRT1 siRNA treated cells had similar numbers of surviving, silenced clones (9 and 8 out of approximately 1000 cells selected, respectively) suggesting that the amount of reduction achieved for SIRT1 recruited to the promoter during DNA damage did not alter silencing of the promoter. Next, the DNA from clones that survived ganciclovir treatment was bisulfite treated and sequenced as above. NT treated HSVTK silent clones showed a similar pattern to non-siRNA treated cells (untreated) both in terms of how many bases were DNA methylated per clone and the position where the methylation occurred (Figure 6A and 6B). Thus, CpGs 3′ to the cut site were methylated in 8 out of 9 of the clones with a mean of 3.1 methylated CpGs per clone (Figure 6A and 6C). In HSVTK silent clones from SIRT1 siRNA treated cells, only 2 out of 8 clones had methylated CpGs in numbers greater than those for uncut cells. This difference in the number of HSVTK silent clones with methylated CpGs versus those with unmethylated CpGs was significantly different between the SIRT1 knockdown cells and either the NT or the untreated cells (Figure 6B). Also, the mean number of methylated CpGs per HSVTK silent clone in SIRT1 knockdown cells, 1.3, is significantly different from the number in both the NT knockdown cells, 3.1 methylated CpGs, and the cells not treated with siRNA, 3.4 methylated CpGs with p<.05 by Student's T-test (Figure 6C). The above results suggest that the reduced levels of SIRT1 recruited to the break site during repair do not affect silencing of the promoter. However, possibly by playing a role in transient recruitment of DNMT3B to the DSB region, SIRT1 does appear to play a role in seeding of methylation in the promoter CpG island in occasional cells, which can then be perpetuated and expanded by the persistent presence of DNMT1.
In the present study our data emphasize, as has been shown for some chromatin constituents by others [19],[24],[29],[30], that during normal repair of a DSB, silencing proteins are recruited to the site of DNA damage along with enrichment of their corresponding histone marks. We substantially add to these previous data by showing the involvement of the principal long term silencing complex PcG. In the ARR model of DSB repair SIRT1 and the PcG protein, EZH2, most likely play a role in the restoration phase of repair by returning chromatin back to its original more condensed state or making chromatin even more condensed (Figure 6D). We hypothesize that following DNA damage in our particular model involving a gene promoter region, the EZH2 catalyzed trimethylation of H3K27, plus the enrichment of the additional silencing marks, H3K9me2 and H3K9me3, may all lead to a transient silencing of the gene in order to make sure the DNA repair is complete before transcription can resume and/or to a compaction of the chromatin that blunts the DNA damage signaling stimulated by the initial opening of the chromatin [23].
During the normal process of DSB repair the association of the above proteins and histone marks with the DSB appears to be transient for most cells, returning to low or absent baseline levels after repair has occurred. This is true in our exogenous gene promoter region, even in a tumor cell which involves a promoter sequence that frequently is DNA hypermethylated and abnormally, heritably silenced in cancer. However, and important to the model for how abnormal CpG island DNA hypermethylation and gene silencing might occur in cancer, we have demonstrated that induction of a break in the promoter of a gene can infrequently lead to long term silencing of that gene. Silencing could occur because, occasionally, there is permanent association of the silencing factors to the break or at least proteins that are important for establishing an epigenetic memory for silencing. Additionally, in cells with such retained silenced promoters, there appears to be an early seeding of CpG methylation that spreads over time and which potentially can, then, contribute to a more stable silencing of the promoter.
This work suggests that a DSB occurring in the promoter of a gene may be an initiating event for the silencing of the promoter, leading to a mechanism by which oxidative or other DNA damage can induce epigenetic silencing, including promoter CpG island DNA hypermethylation of tumor suppressor genes. In this regard, SIRT1 is a key stress response and cell survival protein [35],[38]. This protein has now been associated in stem/precursor and cancer cells with silencing chromatin [40] including PcG complexes [43],[58], with DNA damage repair in multiple settings [59],[60], and, in our own studies, with maintenance of gene silencing for DNA hypermethylated cancer genes [39]. We [61] and others [62],[63] have recently reported that there is an association of embryonic stem cell like repressive chromatin patterns for large groups of such cancer genes, and particularly PcG components and the corresponding histone modification, H3K27me3. Furthermore, we have hypothesized from studies of cancer progression, including the very early appearance of many DNA hypermethylated genes, that this PcG component is particularly important for the vulnerability of genes to the abnormal DNA methylation during cancer evolution [64]. In turn, we have wondered whether settings that are high risk for cancer development, such as chronic inflammation which exposes cells to a significant amount of DNA damage, collaborate with the PcG chromatin for such DNA methylation recruitment. Our present study further points to this possibility and links SIRT1 to the process, especially to recruitment of DNA methylation. The findings suggest that a DSB occurring in the promoter of a gene may initiate epigenetic silencing in occasional cells and this silencing, in turn, could contribute risk of tumor development.
Our present link of SIRT1 and PcG to the DNMTs during DNA damage repair brings up important issues regarding whether these proteins form a complex during DNA repair or are recruited independently to the break site. A PcG complex, termed PRC4, containing SIRT1 and EZH2, has previously been identified [43]. Additionally, SIRT1 has been found to co-localize and to be co-immunoprecipitated with DNMT1 at rRNA [65]and it has been hypothesized that the DNMTs may be recruited to DNA through interaction with PcG [62],[66]. Although SIRT1 and EZH2 appear to be recruited to the break site in the same time frame, EZH2 is still recruited, and possibly even more so, when SIRT1 is knocked down, suggesting that its recruitment is not dependent on SIRT1. There are higher levels of EZH2 enrichment in the TK gene in the SIRT1 knockdown cells in contrast to relatively low EZH2 enrichment in the TK gene in the non-target knock down cells. The presence of higher levels of EZH2 and H3K27me3 may be an attempt to further compact the DNA in the absence of high levels of SIRT1 or to turn off the DNA damage signal which may have been initiated by histone acetylation and therefore maintained in the SIRT1 knockdown cells. Intriguingly, an important and novel finding in our studies is that the seeding of the DNA methylation appears highly dependent on SIRT1 presence during the acute DNA damage and repair interval and seems likely to involve a role for SIRT1 in the transient localization of the de novo DNMT, 3B, during repair. It is unclear from this work whether this recruitment of DNMT3B is because of a direct interaction with SIRT1. After DNA damage, the earliest time points of SIRT1 recruitment (4 and 4+4 hours-Figure 2B) correspond to the time points where we demonstrate enrichment of DNMT3B (Figure 2D). However, DNMT3B is not enriched at later time points where SIRT1 recruitment is the greatest. As an alternative to a direct physical interaction between SIRT1 and DNMT3B, SIRT1 could affect DNMT3B localization to the cut site indirectly by playing a critical role in a complex that forms at the break site or by modifying another protein that plays a role in DNMT3B recruitment. For example, SIRT1 knockdown appears to have the greatest effect on the persistence of H4K16ac after damage. Acetyl H4K16 is a critical residue for chromatin formation, with deactylation of this residue being associated with tight compaction of chromatin [67]. Its status could influence other histone modifications through composition of the complex formed at the break site that could potentially contain HATs, HDACs, HMTs, and/or histone demethylases. A change in these other histone modifications could in turn influence DNMT3B recruitment. Unlike DNMT3B, DNMT1 localization is independent of SIRT1, suggesting that DNMT1 is recruited through a different mechanism. Further studies need to be performed to identify how silencing complexes formed at sites of DNA damage precisely involve interactions between SIRT1, EZH2 and other PcG components, and the DNMTs and whether such interactions are operative in other transcriptional silencing processes.
Recently Cuozzo et al also demonstrated that DNMT1 is associated with chromatin, after DNA damage, specifically after repair by HR [68]. After HR, some DNA methylation occurs that plays a role in silencing the recombined genes. This silencing is dependent on DNMT1 and reversed by treatment with DAC. Interestingly, this paper showed induced methylation localized 200 to 300 bp 3′ of the break site, similar to the degree and relative localization of DNA methylation in our model cut site [68]. We extend these findings by showing that with passage this methylation can be expanded, increasing from approximately 4 methylated CpGs to 10–13 methylated CpGs in 30 passages. We suggest that our current findings provide molecular support for our previous model [64] concerning how this expansion occurs in tumors. An event occurs in a cell that causes an initial seeding of methylation at a promoter, and over time during tumor progression, this methylation spreads and contributes to progressive stable silencing of the involved promoter. A similar model exists for transient “hit and run” silencing of a promoter construct by a transcription repression complex leading to cell clones with retained silencing, even in the absence of the original complex, and progressive spread of DNA methylation [54].
Another intriguing finding from our work is the different potential roles of DNMT1 and DNMT3B in silencing. Both are found to be transiently recruited to the site of DNA damage, albeit with slightly different timing. In terms of normal DNA repair, this transient recruitment of DNMT1 and DNMT3B to the break site may be part of a universal mechanism used during DNA damage repair to restore the correct DNA methylation code to the area around the break. However, this may be more for areas widely flanking the break in our model situation since the CpG island of gene promoters like the one we are using are generally maintained free of DNA methylation in normal cells. Alternatively, in the transient setting, these proteins could perform a silencing role without using their DNA methylating capacity since multiple studies suggest these proteins have transcriptional repression potential independent of their ability to catalyze DNA methylation [69]–[71]. Although DNMT1 is predominantly a maintenance DNMT, it has been demonstrated to have some de novo methyltransferase activity [72],[73], while DNMT3A and DNMT3B are thought to be the predominate de novo methyltransferases [74]. Our present work supports the thought that these enzymes can work together at different phases of methylation initiation, maintenance, and spreading. We only detect seeding of methylation when DNMT3B is present at the cut site and do not observe seeding when only DNMT1 is present. Additionally, at early passages of the clones that have silenced the promoter containing the cut site, both DNMT1 and DNMT3B are present. However, at later passages, enrichment of DNMT3B is lost even though the methylation is still expanding. We were not able to detect enrichment of DNMT3A either transiently or in our silent clones. It is unclear however whether these results are due to a lack of recruitment or a sensitivity issue for the antibody used for ChIP. These findings suggest that DNMT3B is important for the initial seeding of methylation, while DNMT1 is needed for de novo activity in expanding and maintaining the sites of DNA methylation.
With respect to DNA repair, chromatin modifications are important in the specific steps of repair of DSBs. Phospho-H2AX is required for the recruitment of the chromatin remodeling complex INO80 that most likely plays a role in repositioning nucleosomes around the break [75]–[77]. Phospho-H2AX is also necessary for the stable, concentrated recruitment of DNA repair proteins to the site of the break [48],[78]. In addition to H2AX modifications, it has been demonstrated by using a pan-acetyl lysine H4 antibody that lysine residues of histone H4 are acetylated by the human TIP60 histone acetyltransferase complex in response to DNA damage [33]. Our work supports these findings because at the 4 hour time point we see an increase in H4K16ac when compared to uncut cells. Adding to this process, we show that this initial acetylation of histone H4, specifically at lysine 16, is followed by the deacetylation of the same residue concomitant with recruitment of SIRT1 to the break site. We hypothesize that this deacetylation is important to return the chromatin back to its original state following DNA repair. Although we did not see a change in DNA repair following knockdown of SIRT1 and prolonged acetylation of H4K16, it has been demonstrated in yeast that a lack of deacetylation of H4K16 after DNA damage affects repair by the NHEJ pathway [79].
In mammalian cells it is hard to separate a role for SIRT1 in DNA damage repair from its role in p53 regulation. SIRT1 deacetylates p53 and therefore decreases transcriptional activity of the protein. Inhibition of SIRT1 has been shown to induce apoptosis and enhance radiation sensitization, most likely because p53 acetylation is increased [80]. Loss of SIRT1 has also been shown to allow cells to bypass senescence and allow cell division without repair of DNA [81]. While it is unclear in these previous studies how much of the effect of SIRT1 on damage sensitization and cell cycle check points is dependent on p53, our system looks at the role of this protein in DNA damage repair independent of p53. In MDA-MB-231 cells p53 is mutated so, although SIRT1 knock down causes an increase in acetyl p53, the p53 protein is non-functional [51]. Our work then directly demonstrates that SIRT1, in the absence of functional p53, is localized to the chromatin near a DSB and plays a role in recruiting DNMT3B to the vicinity.
In summary, the system of Jasin et al [46] used here uniquely allows us to determine if induction of a DSB in a promoter can lead to transcriptional silencing. In a transient setting, several factors that play a role in gene silencing are recruited to the break and, occasionally, retention of some of these factors can lead to sustained silencing, which can be associated with initiation and spreading of DNA methylation to further stabilize the silencing. Our model is important to understanding how DNA damage occurring at gene promoter sites may be one key factor in initiating abnormal epigenetic gene silencing in association with abnormal CpG island DNA methylation. In terms of cancer prevention, targeting the series of events suggested by our model at sites of chronic inflammation may be beneficial to reducing tumor formation by decreasing the silencing of the large number genes which we now know become aberrantly silenced during neoplastic progression [45],[52],[64]. Our findings suggest new molecular events to consider for cancer prevention targeting and the need for a further understanding of the complex that initiates DNA silencing and how it is recruited to promoters.
MDA-MB-231 cells (ATCC, www.atcc.org) were cultured in Dulbecco's modified Eagle's medium supplemented with 10% tetracycline-tested fetal bovine serum (Hyclone, www.hyclone.com). The homing endonuclease I-SceI, along with the NLS and HA epitope tag, was amplified from the pCMV-ISceI vector [46] (a gift from M. Jasin) and inserted into the pcDNA4-TO vector (Invitrogen, www.invitrogen.com). The pEGFP1-E-cad vector contains genomic DNA corresponding to the human E-cad promoter inserted into the EcoRI/SalI sites of the pEGFP-1 vector. The consensus I-SceI cut-site was inserted into the E-cad promoter at the unique MluI restriction site that is located at −171 in relation to the transcription start site. This insertion avoids all characterized Ebox and Sp1 elements within the E-cad promoter. The EGFP coding sequence was removed and replaced with HSV thymidine kinase sequence that was amplified from the BaculoDirect N-terminal linear DNA Gateway Cassette (Invitrogen). MDA-MB-231 cells were co-transfected with pcDNA6-TR (Invitrogen) and pcDNA4-TO-HA-I-SceI and dual integration was selected for using Zeocin and Blasticidin treatment. Stable clones were isolated and screened for tetracycline induced expression of HA-I-SceI and no background expression without tetracycline. The clone with highest inducible expression was then transfected with the pCDH1-I-SceI-HSVTK vector. Clones with stable integration were selected for with G418 treatment. Clones were screened by two PCRs with linear amplification for single copy insertion of the entire sequence. To verify the copy number in the final clone selected (ROS8) we prepared copy standards using a known amount of pCDH1-I-SceI-HSVTK plasmid DNA combined with 50 µg genomic DNA from non-transgenic MDA-MB-231 cells or pGAPDH plasmid DNA only. By realtime PCR for a primer set in the TK gene or in the GAPDH gene we used the copy standards to develop two standard curves. The TK copy number in 50 µg genomic DNA from the pCDH1-I-SceI-HSVTK containing clone was normalized so the GAPDH copy number was 2 (Figure S1) [82].
For tet-induced expression of the HA-I-SceI enzyme, tet (Sigma, www.sigmaaldrich.com) was added to the culture media to a final concentration of 1 µg/ml. After 4 hours the media was removed and the cells were washed twice with PBS. Then cells were either collected (4 hour time point) or fresh media was added and the cells were incubated at 37°C for an additional 4 hours (4+4 hour time point), 16 hours (4+16 hour time point), 24 hours (4+24 hour time point), or 48 hours (4+48 hour time point). For the generation of silent clones, cells were treated with tet for 4 hours or 24 hours followed by being sub-cultured at a density of 1000 cells per 100 mm dish (cell numbers were determined by counting with a hemocytometer). Ganciclovir (Sigma) was added to the dish at a final concentration of 50 µM. Media was changed bi-weekly until single clones were observed. Silent clones were continually grown in the presence of ganciclovir. Clones originating from cells treated with 4 hours of tet were labeled A and those originating from cells treated with 24 hours tet were labeled B.
Total RNA was extracted (Qiagen, www.qiagen.com) according to the manufacturer's instructions and subjected to reverse transcription using Superscript II RNAse H Reverse Transcriptase (Invitrogen) followed by semi-quantitative polymerase chain reaction or quantitative real-time polymerase chain reaction. For real-time analyses, the QuantiTect SYBR Green PCR kit (Qiagen) and a BioRad iCycler (Biorad, www.bio-rad.com) were used. Values reported were based on a standard curve generated by serial dilution of the untreated parental sample, and expression was reported as a fraction of the expression in the untreated parental samples. The sequences of the primers used are listed in Table S1.
Part of the sonicated samples collected for ChIP was used for western blot. Protein concentrations were measured by BCA (Pierce Biotechnology, www.piercenet.com). Protein extracts were subjected to polyacrylamide gel electrophoresis using the 4%–12% NuPAGE gel system (Invitrogen), transferred to PVDF (Millipore, www.millipore.com) membranes, and immunoblotted using antibodies that specifically recognize SIRT1 (DB083, Delta Biolabs, www.deltabiolabs.com), HA (sc-805, Santa Cruz Biotechnology, www.scbt.com, Figure 2), HA-HRP (12013819001, Roche Applied Science, www.roche-applied-science.com, Figure 3), phospho-H2AX (05-636, Millipore Corporation), and acetylated lysine 382 p53 (Cell Signaling Technology, www.cellsignal.com).
ChIP analysis was performed as described previously [83] with a few modifications. Culture medium was removed, the cells were washed once with PBS, and then an additional 10 ml of PBS was added to the plate. Proteins were cross-linked to proteins by addition of disuccinimidyl glutarate (DSG, Pierce) to the PBS to a final concentration of 0.5 mM for 30 min at room temperature. Proteins were then cross-linked to DNA by addition of formaldehyde to a final concentration of 1% for 10 min at room temperature. Antibodies to SIRT1 (05-707, Figure 2B & 5B), phospho-H2AX (07-164), and H4K16ac (07-329) were obtained from Millipore. Antibodies to SIRT1 were also obtained from Delta Biolabs (DB083, Figure 3B & C). Antibodies to EZH2 were from Cell Signaling Technologies (4905). Antibodies to DNMT1 (IMG-261A) and DNMT3B (IMG-184A) were from Imgenex (www.imgenex.com). Antibodies to H3K9me2, H3K9me3, and H3K27me3 were generous gifts from T. Jenuwein. Immune complexes were collected with 100 µl of 3∶1 Protein A and Protein G magnetic Dynabeads (Invitrogen) for 2 hours at 4°C. Primers were used to amplify the promoter region of the inserted E-cad promoter (SCE). The sense primer was specific for the E-cad promoter and the anti-sense primer was specific for the E-cad promoter containing the cut site. A different set of primers was used to amplify a region in the HSVTK gene (TK). Sequences of the primers are listed in Table S1. Ten microliters of PCR product were size fractionated by PAGE and were quantified using Kodak Digital Science 1D Image Analysis software. Enrichment was calculated by taking the ratio between the net intensity of the gene promoter PCR products from each primer set for the bound, immunoprecipitated sample and the net intensity of the PCR product for the non-immunoprecipitated input sample. Values for enrichment were calculated as the average from at least three independent PCR analyses. Each ChIP experiment was performed twice. The data presented is from one representative experiment.
Cells were transiently transfected with 25 nM non-target siRNA (D-001210-05, Dharmacon, www.dharmacon.com) or SIRT1 siRNA (L-003540-00, Dharmacon) using lipofectamine 2000 (Invitrogen) for three consecutive days following the manufacturer's suggested protocol. On the fourth day, the tet treatment schedule was started either for collection of samples at different time points or tet treatment prior to ganciclovir selection.
Cells were treated with mock, 1 µM 5-Aza-dC (Sigma) for 72 hours, or with 300 nM TSA (Wako, www.wakousa.com) for 16 hours, as described previously [56].
Bisulfite sequencing was performed as previously described [84] on DNA from parental uncut cells or clones that were ganciclovir resistant and had silenced HSVTK. Primers that are specific for bisulfite treated DNA and are methylation non-specific were used (Table S1). The sense primer is specific for the E-cad promoter. The anti-sense primer is specific for the E-cad promoter containing the cut site.
|
10.1371/journal.pbio.1001144 | Genetic Variation Shapes Protein Networks Mainly through Non-transcriptional Mechanisms | Networks of co-regulated transcripts in genetically diverse populations have been studied extensively, but little is known about the degree to which these networks cause similar co-variation at the protein level. We quantified 354 proteins in a genetically diverse population of yeast segregants, which allowed for the first time construction of a coherent protein co-variation matrix. We identified tightly co-regulated groups of 36 and 93 proteins that were made up predominantly of genes involved in ribosome biogenesis and amino acid metabolism, respectively. Even though the ribosomal genes were tightly co-regulated at both the protein and transcript levels, genetic regulation of proteins was entirely distinct from that of transcripts, and almost no genes in this network showed a significant correlation between protein and transcript levels. This result calls into question the widely held belief that in yeast, as opposed to higher eukaryotes, ribosomal protein levels are regulated primarily by regulating transcript levels. Furthermore, although genetic regulation of the amino acid network was more similar for proteins and transcripts, regression analysis demonstrated that even here, proteins vary predominantly as a result of non-transcriptional variation. We also found that cis regulation, which is common in the transcriptome, is rare at the level of the proteome. We conclude that most inter-individual variation in levels of these particular high abundance proteins in this genetically diverse population is not caused by variation of their underlying transcripts.
| The level of protein produced by each gene corresponds approximately to the level of mRNA transcript produced by that gene: so high-abundance proteins, like those involved in protein synthesis, are represented by high-abundance transcripts, whereas low-abundance proteins, like those involved in signaling pathways, are represented by low-abundance transcripts. Furthermore, genetic variation can cause variation in transcript levels for the same gene between different individuals. These two observations have led to the assumption that inter-individual variation in transcript levels for any particular gene causes corresponding variation in protein levels. However, this need not be the case, because protein levels could be controlled not only by regulating transcript levels but also by regulating protein translation and stability. Because inter-individual variation in the levels of the transcript for any particular gene is typically less than 3-fold, rather than orders of magnitude, it is possible that the predominant cause of inter-individual variation in levels of any particular protein is transcription-independent regulation of protein levels. Here, we look in a genetically diverse population of 95 yeast strains at the genetic variation that leads in turn to variation in levels of 354 proteins that function within co-regulated networks. We find that the between-strain variation predominantly reflects transcription-independent mechanisms. If this result is typical of the proteome as a whole, it suggests that protein levels in genetically diverse populations cannot be accurately inferred from levels of their underlying transcripts.
| Genetic variation leads to networks of co-regulated transcripts. The implications of these network structures have been discussed extensively, generally with the assumption that such transcriptional networks give rise to corresponding protein networks [1]–[6]. However, due to limitations in technology, these hypothesized protein networks have not been examined directly, and thus it is not known whether they are driven by underlying transcriptional networks. By measuring protein and transcript levels for 354 genes in a genetically diverse population of yeast segregants, we are now able to address the question “Are protein networks formed primarily on the basis of regulation of their underlying transcripts?” (We use the term “protein networks” to refer to groups of proteins that are co-regulated and not groups of proteins that interact physically or genetically.) Before describing our results, three points are important to consider: First, the magnitude of individual to individual variation in transcript levels for a single gene is generally far less than the magnitude of gene to gene variation in transcript levels within a single individual [7]. Second, the demonstration in multiple studies that the correlation between transcript and protein levels for different genes within a single individual is high does not imply that differences in abundance of the same transcripts between different individuals must cause corresponding variation in protein abundance [8],[9]. And third, a correlation between transcript and protein networks does not prove a causal relationship between the two.
Protein levels cannot necessarily be inferred from transcript levels because protein levels can be controlled not only by regulating transcripts but also by regulating other steps in protein metabolism, such as translation and protein stability. Thus the degree to which protein levels can be inferred from transcript levels depends on the degree to which the former mode of regulation overwhelms the latter two (Figure 1). In experimental situations when a gene is placed under a strong promoter like the CMV promoter, a transcript can be elevated 1,000-fold and this generally leads to a striking increase in protein. However, in genetically diverse populations, transcript levels generally do not vary 1,000-fold between individuals; for example, in the population of yeast described in this report, a typical transcript varies just 2.7-fold across 95 individuals. Such modest variation in transcript levels may be buffered such that it causes no variation in protein levels, or regulation of translation and/or protein stability may obscure effects of minor transcriptional variation. Under such circumstances, transcript levels should not be expected to reflect protein levels, and whether such circumstances are the norm or the exception for typical levels of inter-individual variation is not known.
Several reports have demonstrated that, when comparing different genes whose transcript levels vary over orders of magnitude, high abundance proteins are associated with high abundance transcripts and vice versa [8]–[10]. There are two important differences between this issue and the issue we address here: First, the magnitude of variation in transcript levels is vastly different in the two situations. Just as 1,000-fold overexpression of a transcript through experimental manipulation is virtually guaranteed to increase the level of the corresponding protein, transcript levels that differ by orders of magnitude between genes are virtually guaranteed to manifest themselves in differences in the corresponding proteins. For example, a 2007 study by Lu et al. of 346 genes in yeast demonstrated a high correlation (R = 0.85) between transcript and protein levels, and thus concluded that “… >70% of yeast gene expression regulation [occurs] through mRNA-directed mechanisms” [8]. However, if one calculates correlation coefficients with this same data set using a sliding window within which transcript levels vary on average just 3.5-fold, the average correlation drops from 0.85 to 0.36 with almost half of the bins showing correlation coefficients that could easily have been achieved by chance (Figure S1, Text S1). Thus, the striking correlation between transcripts and proteins all but disappears when analysis is limited to a range of transcript variation similar to that occurring between individuals. Second, as we will discuss further below, studies like that of Lu et al. involve protein and transcript measurements from a single individual under a single experimental condition, making it impossible to measure gene-specific correlation coefficients (Figure S2). The point here is not to call into question the solid conclusions of these past reports, but instead to point out that our work addresses a different issue.
In assessing the importance of transcriptional regulation in determining protein levels, it is important to distinguish correlation from causality. Biological pathways that sense physiological conditions can trigger responses that include changes in transcription, translation, and protein stability, often with the same group of genes targeted by more than one of these regulatory mechanisms [11]. For example, the TOR pathway, a highly conserved pathway named for the signaling kinase “Target of Rapamycin,” responds to changes in nutritional conditions by increasing both transcription and translation of a group of target genes [12],[13]. Thus transcript and protein levels of these genes will be correlated, but if translation has a much larger effect on protein levels than does transcription, this correlation need not reflect causality. More generally, any time a cellular response pathway has both a transcriptional and a post-transcriptional branch, the target genes are expected to show a correlation between transcript and protein levels, but it is only in those cases where the former regulatory mechanism is the dominant one in affecting protein levels that this correlation reflects a predominantly causal relationship between transcript and protein levels (Figure 1).
We have previously reported a mass spectrometry-based method for protein quantitation that relies on mathematical alignment of ion signals in mass spectra (MS1) from multiple samples [14],[15]. This algorithm rounded mass to charge measurements to integer “Dalton” values, which has the advantage of making the data sets much smaller than they would be if one made full use of the high mass accuracy of modern mass spectrometers (like that on which the data were collected) and thus avoids computational difficulties that arise with large data sets. However, such rounding sacrifices accuracy to the extent that our previous quantitation, while sufficient for obtaining a broad view of the genetic architecture of protein expression, was insufficient in terms of both accuracy and coverage to rigorously address the causal relationship between variation in transcript levels and variation in the corresponding proteins. To overcome the limitations of our previous algorithm, we used a modification of an accelerated random search [16] to solve computational challenges and developed a new protein quantitation algorithm that exploits the high mass accuracy and resolution of modern mass spectrometers (Text S1). We then used this algorithm to reanalyze our previously reported [14] mass spectrometric data, aligning 380 data sets: two technical replicates of two biological replicates for each of 95 progeny strains derived from a cross between a wild type and a laboratory strain of yeast [17]. These two strains differ at approximately 0.5% of their base pairs [18], and this cross has been studied extensively [19]–[21]. Restricting ourselves to peptides that were identified with high confidence (Text S1) and that corresponded uniquely to one protein, we quantified 354 proteins (Table S1). This is more than twice the number of unique peptides (164) we were able to quantify in our previous report [14].
If one is to assess the effect of transcriptional variation on the proteome, it is necessary to focus on transcripts that show significant individual-to-individual variation. With measurements for only 354 proteins, constituting less than 6% of the proteome, we were concerned that the corresponding transcripts might not show significant individual-to-individual variation; thus we looked at variance of the transcripts in question (transcript data previously reported [17]). The 354 genes for which we had protein measurements were all among the most highly variant ∼10% (522/6,215) of all transcripts. Thus this subset of proteins is not merely sufficiently variable for our study; it comprises almost 70% (354/522) of the ideal genes on which to focus for our purposes. We speculate that this fortuitous result reflects the fact that we are best able to measure high abundance proteins, and thus our data set is enriched for accurately measured high abundance transcripts as well. (Levels of highly abundant proteins tend to be less variable than low abundance proteins [22]; therefore it is unlikely that the high variance of this set of proteins is a reflection of their abundance.) We note, however, that this is a special set of proteins in that they are mostly high abundance proteins involved directly or indirectly with protein synthesis and thus they may not be representative of the proteome as a whole.
We next constructed a connectivity matrix between proteins on the basis of Pearson's correlation coefficient. For each pair of proteins, we used permutation testing to determine a false positive rate (FPR) cutoff, accepting only connections that were below a 1% FPR cutoff. Out of 62,481 possible protein-protein connections, we observed 7,058 connections, 91% of which were deemed genuine because we expect only 625 connections by chance. The numbers of connections for individual proteins ranged from 1 to 100, with an average of 40 and a median of 37. For transcripts, we observed 15,989 connections out of 62,481 possible. For individual genes, the transcript numbers ranged from 4 to 176, with an average of 90 and a median of 76. Among the 50 most highly connected proteins, there was a 1.9-fold enrichment for genes involved in amino acid biosynthesis, and among the 50 most highly connected transcripts, there was a 2.7-fold enrichment of genes involved in ribosomal functions, but enrichment for other functions was not obvious. Remarkably, the most highly connected genes for proteins and transcripts look entirely unrelated. For example, among the 34 most highly connected genes in the two groups, there are only two genes present in both groups (RPS7A and TEF4). A global comparison suggests that connectivity of genes at the transcriptional level is unrelated to their connectivity at the protein level (Figure S3).
In order to identify networks of co-regulated genes, we turned to a widely used “community”-based approach [23]. Cliques are groups in which each member is connected to every other member, and “communities” are simply groups of highly overlapping cliques (precise definition in legend to Figure 2); thus in our case, communities are groups of proteins or transcripts that show a high degree of co-variation. In both the protein and transcript data sets, we identified two large communities, one enriched for genes involved in amino acid metabolism and the other for genes involved in ribosome biogenesis. The amino acid and ribosomal communities in the protein data set consisted of 93 and 36 genes, respectively, and these two communities in the transcript data set consisted of 67 and 127 genes, respectively (Figure 2; Table S2). Even though the genes within each community showed a high degree of co-regulation, the two communities within each data set showed very little connection. For example, only 1.6% (54 out of 3,348 with 378 expected by chance) of possible intercommunity protein pairs were connected; thus we have two networks of highly connected proteins that vary largely independently of one another (Figure 2). (Below, unless specified otherwise, if we refer to a ribosomal network or community, we mean the protein ribosomal community and the same is true for references to an amino acid network or community.)
With two large networks of functionally related proteins, we were in a position to address our main question, namely “to what extent are protein networks shaped by regulation of their underlying transcripts?” If protein networks are shaped primarily by variation in their underlying transcripts, we would expect (1) high correlations between proteins and transcripts, (2) similar genetic regulation of proteins and transcripts, and (3) that normalization of protein levels according to variation in transcript levels should abolish linkage to genetic regulatory loci. Using these criteria, we found that both protein networks were formed primarily through non-transcriptional mechanisms.
We calculated correlation coefficients between protein and transcript levels for all 354 genes and then assigned each a binary value of “significant” or “not significant” on the basis of 1,000 permutations done separately for each gene. Only 4 out of 36 genes in the ribosomal network showed a significant correlation (p<0.05); thus clearly for the vast majority of these genes, protein levels vary without regard to the levels of their corresponding transcripts. The genes in the amino acid network showed a higher fraction of significant correlations, but even here less than half of the genes (41 out of 93) showed a significant correlation at the same 5% cutoff. These results demonstrate an important non-transcriptional component to regulation of both networks and suggest that the ribosomal network is either largely unaffected by transcriptional variation, within the range of transcriptional variation observed here, or that transcriptional regulation of protein abundance is obscured by regulation of translation and/or protein stability.
The variable correlation between transcripts and proteins for these two networks raises the broader question of how transcripts and proteins are correlated in general. As noted above, several studies in yeast including one from our laboratory have reported a wide range (0.34–0.98) of correlations between protein and transcript levels [8],[9],[14]. However, two features of the current study are critically different from the previous reports. First, this study examines gene-specific correlations. Most previously reported correlation coefficients for protein and transcript abundance for yeast were derived from single measurements of protein and transcript levels for many genes in a single strain, and these individual measurements for different genes were combined to derive an average correlation coefficient. The correlation coefficients we report here, in contrast, are derived from 95 measurements of protein-transcript pairs for each of 354 different genes (Figure S2). This is important because there are dramatic differences in the degree to which different genes are regulated at the transcript level versus the protein level. Second, while our previous study reported gene-specific correlation coefficients [14], we did not emphasize these results because our marginal ability to map protein regulators raised the possibility that these low correlation coefficients reflected inaccuracies in our protein measurements (see below for mapping results). (The ability to map significant numbers of regulators can be used as a metric to assess accuracy of measurements when FDRs are empirically determined through permutation testing [24].)
We found that proteins and transcripts were well correlated for only 27% of genes (94/354 at 5% significance; Table S3), with most genes showing little or no correlation. Even if we limit ourselves to proteins for which we mapped regulators (p<0.05; see below for mapping results) and thus have high confidence in our protein measurements, only 37% of genes (46/125) show significant correlations between proteins and transcripts. Furthermore, we could find no relationship between these correlation coefficients and the corresponding genes' transcript or protein half life [25],[26]. Plotting the data in terms of variance explained for transcripts and proteins similarly failed to reveal trends (unpublished data). We conclude that for most genes, inter-individual variation in protein levels does not reflect variation in underlying transcripts (Figure 3A).
To compare the genetic regulation of the two protein networks, we began by mapping loci that affect transcript and/or protein levels. (We note that heritability for proteins, like that for transcripts, was high: averages for proteins and transcripts were 0.70 and 0.71, respectively, and medians for proteins and transcripts were 0.71 and 0.74, respectively; calculation described in Materials and Methods.) All strains have been genotyped for 2,955 genetic markers, 1,969 of which exhibited unique segregation patterns among the 95 segregant strains. We looked for linkage between inheritance of these 1,969 markers and the 354 transcript and protein levels using t tests and determined FDR cutoffs for each gene on the basis of 100 permutations. At a 5% FDR, we mapped 49 and 97 loci that control the level of a total of 125 and 200 proteins and transcripts, respectively. At a 1% FDR, we mapped 30 and 74 loci that control levels of 89 and 170 proteins and transcripts, respectively (Table S4). Because proteins and transcripts can map to more than one locus, the total number of linkages at a 5% FDR was 179 for proteins and 342 for transcripts, and at a 1% FDR these numbers were 115 and 253, respectively. These results provide an objective metric for the extent to which our current algorithm (i.e., the one used in this report) has improved our accuracy: At a 5% FDR, our previous algorithm [14] allowed us to map 24 regulators that regulate levels of 18 proteins, whereas the current algorithm allowed us to map 179 regulators that regulate levels of 125 proteins. (We use the term “regulator” to denote a locus that influences transcript and/or protein levels.) With comparable measurements for proteins and transcripts, we are now able to address questions about the relationship between the two data sets that we could not address in our previous publication [14]. Consistent with our previous results, we found that (1) both proteins and transcripts show hot spots of regulation (single loci that control multiple genes), (2) these hot spots are largely but not completely overlapping, and (3) the genes regulated by a single hot spot show low overlap at the protein and transcript levels, highlighting the difference between genetic regulation of the proteome and transcriptome (Figure 3B).
The locations of the genetic regulators that control proteins and transcripts within the ribosomal network bore essentially no resemblance to each other; indeed, given the overall distribution of regulatory loci for proteins and transcripts, loci that regulated both proteins and transcripts within this network appeared much less frequently than is expected by chance (p<0.0001 based on 10,000 permutations; Figure 3C). These results call into question the widely held belief that in yeast, in contrast to vertebrates, ribosomal protein levels are controlled primarily by regulation of their transcripts [27]. (We note that genetic regulation of the ribosomal transcripts is complex, i.e. a large number of loci, each with relatively small effect.) The locations of genetic regulators of proteins and transcripts were more similar for genes in the amino acid network: Approximately a quarter of the time (20 out of 84 linkages, corresponding to 17 different regulated genes), protein linkages (5% FDR) to genes in the amino acid network showed regulation of the corresponding transcript by the same locus. Below we ask whether, at least for these 20 linkages, the mechanism by which the loci regulate the proteins is regulation of their underlying transcripts.
This subset of 20 linkages comprise the most likely examples of loci that control the levels of proteins in the protein amino acid network primarily by controlling the underlying transcripts, but even here it is possible that transcription is not the main driver of protein levels. For example, a response to alterations in cellular physiology created by polymorphisms on chromosomes 3 and 13 may include both a transcriptional response of a specific set of genes and changes in the translation of the corresponding transcripts and/or stability of the corresponding proteins. Multilevel control (i.e. transcriptional and posttranscriptional) of the same genes is a well-described phenomenon in response to environmental changes and in development that assures the magnitude and rapidity of response and that reinforces cellular decisions [28],[29]. If the translational or protein stability changes have a larger effect on protein levels than the transcriptional alterations, one would still see shared genetic regulation and high correlation coefficients, but the protein network would not be driven primarily by transcription (Figure 1). To distinguish between these possibilities, we asked whether these 20 linkages for proteins in the amino acid network maintained linkage after normalizing for transcript levels [30]. For each of the proteins and the sites to which they are linked, such as ACS2 to chromosome 12, which is shown as an example (Figure 4A), protein levels were regressed on the corresponding transcript levels (Figure 4B), the residuals were tested for linkage to the original loci (Figure 4C), and residual linkage was plotted against the original linkage (Figure 4D). In the case of ACS2, it is clear that the locus on chromosome 12 is regulating ACS2 protein levels by regulating transcript levels, because the tight linkage between the locus and protein levels (p = 6.03×10−10, Figure 4A) becomes insignificant when the effect of the locus on transcript levels is taken into account (p = 0.178, Figure 4C). Two other linkages also behaved this way; thus a total of three out of the 20 protein linkages examined appear to reflect primarily transcriptional regulation of protein levels (three points below horizontal line at −log 0.05 in Figure 4D, in which all protein linkages have been plotted). Two of these three map to the transcription factor HAP1, which is inactivated in the laboratory parent by a Ty element insertion. These two genes, ACS2 and ERG6, are both regulated by HAP1 and both have upstream HAP1 binding sites that are among the most tightly HAP1-bound sites in the genome, thus suggesting a mechanism for transcript-mediated regulation of these two proteins [31]. Extending this test for transcript-mediated control of protein levels to all 179 protein linkages (p = 0.05) shows that the three cases mentioned above are the only instances in which control of protein levels can be attributed exclusively to control of the corresponding transcript (Figure 4D).
Quantitative trait loci that affect transcript levels have been classified according to whether the regulated gene is linked (cis-regulation) or unlinked (trans-regulation) to the regulatory locus. For example a promoter mutation would be classified as cis-regulatory whereas a mutation in a transcription factor would likely appear as trans-regulatory. Consistent with numerous reports in both this collection of yeast strains and other populations [32]–[34], we find that at the transcript level, cis regulation of the 354 genes in this study is relatively common (Figure 5): at 1% FDR, we map regulators of 170 transcripts, 22 of which act in cis. In contrast, at this FDR we map regulators of 89 proteins, only three of which act in cis (Figure 5). t tests between each gene and a single cis marker to reduce the problem of multiple testing confirmed the wider prevalence of cis linkage in the transcriptome over the proteome: At a 1% significance, 50 transcripts show linkage to the nearest marker whereas only 13 proteins do. Normalizing for the 3.54 false positives expected suggests approximately 5-fold more cis linkage for transcripts than for proteins. As others have noted, we saw that cis regulators tended to have above average effect sizes and therefore should be easier to detect; for example, cis linkages at 1% significance explained on average 29% of variation in transcript levels whereas the corresponding trans linkages explained only 24%. Thus if variation in protein levels between individuals were caused by variation in the underlying transcripts, inaccurate measurements of either proteins or transcripts would lead to an overestimate of cis linkage, whereas we see the opposite. Cis linkage for transcript and protein abundance can be due to polymorphisms in the promoter region that alter transcription rates, or polymorphisms in the untranslated or coding sequences that alter transcript stability, translation rate, or protein stability. The virtual absence of cis linkage for proteins in the face of relatively common cis linkage for transcripts is consistent with our finding that variation in protein levels in this cross is largely independent of transcriptional variation. Furthermore, the rarity of cis linkage for proteins also suggests that cis-acting polymorphisms that lead to alterations in protein stability and/or translation rates are less common than those that alter transcription rates.
It is particularly noteworthy that, although ribosomal proteins and their corresponding transcripts are both tightly co-regulated, they are regulated entirely independently of one another; it is as if the loci regulating ribosomal transcripts are actively avoiding those that regulate the corresponding proteins, since permutation testing (with 10,000 permutations) demonstrates that there is less than a one in 10,000 chance that this level of “avoidance” could have happened by chance. This suggests that the loci that trigger the transcriptional response for ribosomal protein genes are acting through one pathway while the loci that trigger the corresponding translational (or protein stability) response are acting through another. The pathways that regulate transcription of ribosomal protein genes have been studied extensively. For example, nutrients activate the TOR and PKA pathways, which leads to phosphorylation of Sch9 and Sfp1. This increases the levels of Sch9 and causes Sfp1 to enter the nucleus, and this in turn triggers Fhl1- and Ifh1-dependent transcription of ribosomal protein and biogenesis genes [35],[36]. Thus the loci that regulate transcripts in the ribosomal community are likely perturbing intracellular physiology in a way that elicits a TOR and/or PKA signaling response.
So which pathway might regulate the translational (or protein stability) response that influences ribosomal protein levels? The loci that regulate the protein levels for genes in the protein ribosomal community (upper left in Figure 2) showed a striking resemblance to the loci that regulate transcripts in the transcript amino acid community (lower right in Figure 2; Figure 6), suggesting that these two networks are regulated by a common pathway. Furthermore, the major loci that regulate these genes on chromosomes 3 and 13 had diametrically opposing effects on the two groups; i.e. one allele of chromosome 3 caused essentially every amino acid transcript (65/67) to increase while also causing every ribosomal protein (36/36) to decrease, and the same was true for the locus on chromosome 13. Thus it appears that a single pathway is causing ribosomal proteins and amino acid transcripts to vary in opposition to each other. Our search for a pathway that regulates ribosomal protein translation therefore led us to consider pathways that affect transcription of genes involved in amino acid synthesis. The general amino acid control pathway (GAAC) responds to amino acid imbalances (sensed through levels of uncharged tRNAs) by activating Gcn2, which then phosphorylates the translation initiation factor eIF2A. This phosphorylation causes eIF2A to downregulate translation of a large number of genes while simultaneously promoting translation of the transcription factor GCN4, which in turn stimulates transcription of genes involved in amino acid synthesis [37]. Thus we suggest that loci on chromosomes 3 and 13 skew levels of uncharged tRNAs leading to a GAAC-dependent response that includes increasing transcription (via GCN4) of genes involved in amino acid biosynthesis and decreasing translation of ribosomal proteins. Likely candidate genes are LEU2 on chromosome 3, which is involved in leucine biosynthesis and is heterozygous in this cross and BUL2 on chromosome 13, which regulates amino acid import and is also heterozygous in this cross [38]. mRNAs encoding ribosomal proteins in higher eukaryotes have 5′ terminal oligopyrimidine tracts that are critical in regulating translation and thus protein levels [39]. The fact that yeast lack these so-called TOP mRNAs coupled with the observation that ribosomal protein transcripts in yeast are tightly regulated by the TOR signaling pathway [35],[36] has led to the widespread belief that yeast are different from other eukaryotes in that they regulate ribosomal protein levels primarily through transcription. Our observations suggest an important translational role for control of ribosomal protein levels in yeast, just like in vertebrates.
In summary, our results reveal striking differences in the effect of genetic variation on networks of proteins versus transcripts. Our results demonstrate that, in this genetically diverse population, levels of variation in transcripts are such that non-transcriptional mechanisms for controlling protein levels obscure the correlation between transcripts and proteins for most genes. This is not a consequence of selecting a subset of 354 genes whose transcripts were biased for low variance; on the contrary, every one of these transcripts was among the 522 with the highest variance, suggesting that even among highly variant transcripts, non-transcriptional variation is the predominant cause for variation in protein levels in this population. Finally, advances in proteome profiling technologies that enable high throughput profiling of low abundance proteins [40]–[42] combined with studies of other populations will help determine how general the predominance of non-transcriptional control is in controlling protein levels in genetically diverse populations, since our current analysis is limited to high abundance proteins. Given that protein abundance is more highly related to phenotype than transcript abundance, our findings underscore the fundamental role of non-transcriptional mechanisms in the creation of phenotypic diversity in genetically outbred populations.
Raw data for both protein [14] and transcript [17],[19],[21] quantitation came from previously published work. Protein quantification in this manuscript is novel, i.e. the algorithm to extract protein quantitation from published mass spectrometric data sets is novel. Protein-protein and transcript-transcript correlation coefficients were calculated for each pair of genes. Segregant identities were then “scrambled” for one of the two measurements. For example, in calculating the correlation between levels of CDC19 protein and OLA1 protein, segregant identities would be shifted by 1 such that we calculated correlations on the basis of the level of CDC19 in segregant 2 with OLA1 levels in segregant 1, CDC19 levels in segregant 3 with OLA1 levels in segregant 2, and so forth. For each pair of proteins and for each pair of transcripts, 1,000 such permutations were performed. A 1% FPR was determined as higher than the 10th highest correlation coefficient calculated with the 1,000 permutations. Each pair of proteins and each pair of transcripts was thus assigned its own 1% FPR cutoff and this FPR was arbitrarily chosen as a cutoff for whether two proteins or transcripts are connected.
Communities were identified as previously described [23]. They were all derived from the same group of 354 genes for which both protein and transcript quantitation was available. The five closely related ribosomal protein communities consisted of five 35-member communities made up exclusively of 36 genes. Enrichment for functionally related genes within communities was calculated based on the fraction of members of a community with appropriate GO term assignments within the community as compared to the fraction of the same GO term assignments in the set of 354 genes.
Heritability is the degree to which variation in measurement is due to individual-to-individual variation as opposed to variation due to unintended fluctuation in experimental conditions. For each of 95 strains, there are between zero and four measurements for each protein level, total measurements for each protein level across all segregants, n1 of which come from strain 1, n2 of which come from strain 2, and so forth. Heritability was calculated as follows:Where xi is the mean of the ith segregant, mi is the ith measurement, and GM is the “grand mean,” i.e. the mean of the entire population.
We use variance explained to quantify that proportion of the variation in phenotype (e.g. protein level) can be attributed to inheritance of a particular genetic marker. For example, imagine that the level of Aro1 protein is higher among segregants that inherited a SNP at base pair 10,000 on chromosome 3 from the RM parent than among segregants that inherited this SNP from the BY parent. meanRM is the mean level of Aro1 protein among segregants inheriting the RM SNP and meanBY is the mean level among segregants inheriting the BY SNP. GM is the “grand mean,” i.e. the mean level of Aro1 protein in all of the segregants. There are k total segregants with i segregants inheriting the RM SNP and j segregants inheriting the BY SNP. mRM_1 is the Aro1 protein level in the first segregant inheriting the RM SNP, mRM_2 is the level in the second segregant inheriting this SNP, and so forth.Data collection has been described previously [14]. The protein quantitation algorithm is described in detail below. Complete code is available from D.R. upon request. Data collection has been described previously [14]. The mapping algorithm was based on t tests with permutation testing to determine false discovery rates. Complete code is available from E.J.F. upon request. Communities were identified as described previously [23]. All protein measurements will be placed in the GEO online database upon publication (http://www.ncbi.nlm.nih.gov/geo/). Transcript measurements are available at this site.
|
10.1371/journal.ppat.1002336 | Murine Gamma Herpesvirus 68 Hijacks MAVS and IKKβ to Abrogate NFκB Activation and Antiviral Cytokine Production | Upon viral infection, mitochondrial antiviral signaling (MAVS) protein serves as a key adaptor to promote cytokine production. We report here that murine gamma herpesvirus 68 (γHV68), a model virus for oncogenic human gamma herpesviruses, subverts cytokine production via the MAVS adaptor. During early infection, γHV68 hijacks MAVS and IKKβ to induce the site-specific phosphorylation of RelA, a crucial subunit of the transcriptionally active NFκB dimer, which primes RelA for the proteasome-mediated degradation. As such, γHV68 efficiently abrogated NFκB activation and cytokine gene expression. Conversely, uncoupling RelA degradation from γHV68 infection promoted NFκB activation and elevated cytokine production. Loss of MAVS increased cytokine production and immune cell infiltration in the lungs of γHV68-infected mice. Moreover, exogenous expression of the phosphorylation- and degradation-resistant RelA variant restored γHV68-induced cytokine production. Our findings uncover an intricate strategy whereby signaling via the upstream MAVS adaptor is intercepted by a pathogen to nullify the immediate downstream effector, RelA, of the innate immune pathway.
| Innate immunity represents the first line of defense against invading pathogens chiefly through anti-viral cytokines. The mitochondrial antiviral signaling (MAVS)-dependent innate immune pathways are critical for inflammatory cytokine production. Deficiency in essential innate immune components, such as MAVS, severely impairs cytokine production and host defense that are enabled by the master transcription factor, NFκB. Here we show that murine gamma herpesvirus 68 (γHV68), a model herpesvirus for human Kaposi's sarcoma-associated herpesvirus and Epstein-Barr virus, hijacks MAVS and IKKβ to abrogate NFκB activation and cytokine production. Uncoupling RelA degradation from γHV68 infection restored NFκB-dependent cytokine gene expression and elevated cytokine production. Thus, our results demonstrate that upstream innate immune activation can be harnessed by pathogens to inactivate the downstream effector and subvert cytokine production.
| Innate immunity represents the first line of defense against invading pathogens. Eukaryotic cells express a panel of sensors, known as pattern recognition receptors (PRRs), which detect pathogen-associated molecular patterns that are either structural components or replication intermediates [1], [2]. Toll-like receptors are primarily expressed on immune cells and patrol the extracellular and endosomal compartments. The recently discovered cytosolic receptors (e.g., NOD-like receptors and RIG-I-like receptors) are more ubiquitously expressed and monitor the presence of pathogens in the cytosol. Along with C-type lectins [3], these sentinel molecules constitute the vast majority of PRRs in high eukaryotes.
The cytosolic RIG-I and MDA-5 sensors are authentic RNA helicases that contain two tandem caspase-recruitment domains (CARD) within the amino-terminus and an RNA-binding domain within the carboxyl terminus, endowing the ability to detect nucleic acids [4], [5]. Association with RNA triggers the dimerization of RIG-I and MDA-5 with the mitochondrial antiviral signaling (MAVS, also known as IPS-1, VISA, and CARDIF) adaptor via their N-terminal CARDs, which relays signal to promote antiviral cytokine production [6], [7], [8], [9]. In doing so, MAVS activates the IKKα/β/γ and TBK1/IKKε kinase complexes that, through phosphorylation, effectively promote the gene expression driven by transcription factors of the NFκB and interferon regulatory factor (IRF) family, respectively [10], [11], [12], [13]. It is believed that NFκB activation sufficiently induces the expression of inflammatory cytokines, such as IL6 and TNFα. The efficient transcriptional activation of a prototype interferon (IFN), IFN-β, requires the concerted action of multiple transcription factors including NFκB, ATF2, c-Jun, and IRFs, constituting one of the most sophisticated coordination within multiple innate immune signaling pathways to achieve optimal antiviral immune responses [14], [15]. The participation of numerous components in relaying signaling from pathogen detection to cytokine production maximizes the number of checkpoints to tune host immune responses. Conversely, the highly ordered architecture of signaling cascades also offers pathogens with opportunities to manipulate and exploit host immune responses. Key to the immune signaling cascades is the activation of NFκB transcription factors that control cytokine production, an essential determinant underlying effective host innate and adaptive immune responses.
The family of NFκB transcription factors is composed of five members, including RelA (p65), RelB, c-Rel, NFκB1 (p50 derived from its precursor p100), and NFκB2 (p52 derived from its precursor p105) [16]. All NFκB transcription factors share an N-terminal Rel homology domain that is responsible for subunit dimerization and sequence-specific DNA binding activity. Additionally, RelA, RelB, and c-Rel harbor a C-terminal transcription activation domain (TAD) that positively regulates gene transcription. Among them, RelA is the most ubiquitously and abundantly expressed subunit. By contrast, NFκB1 and NFκB2 do not contain a TAD and therefore rely on dimerization with one of the other three NFκB members to activate gene transcription. Furthermore, post-translational modifications, such as phosphorylation and acetylation, have been identified to confer specific effect on the DNA-binding, protein stability, and transcriptional activity of NFκB transcription factors [17], [18]. Although the signaling pathways that activate NFκB transcription factors have been extensively investigated, relatively little is known regarding the equally important process of NFκB termination.
Herpesviruses are large DNA viruses that establish a lifelong persistent infection. To persist within immuno-competent hosts, gamma herpesviruses in particular have evolved an arsenal of weapons to contend with host immune responses [19], [20]. Being closely-related to human oncogenic Kaposi's sarcoma-associated herpesvirus (KSHV) and Epstein-Barr virus (EBV), murine gamma herpesvirus 68 (γHV68) infects laboratory strains of mice, resulting in robust acute infection in the lung and persistent infection in the spleen. Thus, murine infection with γHV68 offers a tractable small animal model to examine the entire course of host immune responses and viral infection in vivo, which are not available for human KSHV and EBV [21].
To assess the role of immune signaling pathways downstream of cytosolic sensors in gamma herpesvirus infection, we have characterized viral infection and host innate immune responses in MAVS-deficient mice infected with γHV68. We previously reported that γHV68 activates the MAVS-IKKβ pathway to promote viral lytic infection [22]. Paradoxically, the activation of the MAVS-IKKβ pathway often instigates NFκB activation and the production of antiviral cytokines [23], [24]. Moreover, RelA was shown to inhibit γHV68 lytic replication [25]. We report here that γHV68 utilizes MAVS and IKKβ to promote RelA phosphorylation and transient degradation, thereby efficiently abrogating NFκB activation and cytokine production. Finally, loss of MAVS elevated inflammatory cytokines and immune cell infiltration in the lung of γHV68-infected mice, highlighting an essential role of MAVS in evading antiviral cytokine production. Our findings illustrate an intricate strategy whereby a pathogen usurps upstream immune signaling events of the NFκB pathway to destroy the essential downstream effector, RelA, effectively nullifying host innate cytokine production. These findings reshape our view on host innate immune responses.
We have previously shown that γHV68 loads in the lung of Mavs−/− mice were significantly lower than those in the lung of Mavs+/+ mice at 10 d.p.i. [22]. We reasoned that the reduced γHV68 acute infection may be, at least partly, due to an elevated immune response in Mavs−/− mice. To test this hypothesis, we measured inflammatory cytokines, including CCL5, CXCL1, IL6 and TNFα in Mavs+/+ and Mavs−/− littermate mice infected with a low-dose (40 plaque-forming units, PFU) γHV68 by enzyme-linked immunosorbent assay (ELISA). A low dose infection more likely resembles natural γHV68 infection. We found that, in response to γHV68 infection, levels of all four cytokines in the lung of Mavs−/− mice were approximately two-fold of those in the lung of Mavs+/+ mice at 7, 10, and 13 d.p.i. (Figure 1A). This phenomenon is in stark contrast to the observations that loss of MAVS impairs cytokine production in response to viral infection, e.g., RNA viruses [23], [24]. It is notable that γHV68 infection induced significant cytokines in Mavs−/− mice at 7 d.p.i., when cytokines were slightly reduced or unchanged in Mavs+/+ mice, indicating a faster cytokine production. Interestingly, cytokine levels in the sera of γHV68-infected Mavs+/+ and Mavs−/− mice were similar (Figure S1A). There was no statistically significant difference of IL10, an important anti-inflammatory cytokine, in either lungs or sera of γHV68-infected Mavs+/+ and Mavs−/− mice (Figures S1A, S1B). Consistent with our previous reported result [22], loss of MAVS greatly reduced γHV68 load at 10 d.p.i. in the lung, whereas had a marginal effect on viral load at 7 d.p.i. (Figure 1B). These results collectively indicate that, in response to γHV68 infection, Mavs−/− mice produce more inflammatory cytokines specifically in the lung than Mavs+/+ mice.
It was reported that γHV68 replicates to similar levels in IL6-deficient and wild-type mice, implying that IL6 is not obligate to limit γHV68 lytic replication [26]. However, our study suggests that γHV68 successfully abrogates cytokine production during early viral infection, the critical stage for cytokines to curtail viral replication. As such, we reasoned that obliterating IL6 does not enhance γHV68 lytic replication, and that the administration of exogenous cytokines likely better evaluates the effect of cytokines on γHV68 acute infection. Thus, we intranasally administered recombinant mouse IL6 or TNFα (rmIL6 or rmTNFα) after a low-dose (40 PFU/mouse) γHV68 infection. We determined that the optimal efficiency of intranasal delivery was approximately 60% (Figure S2). Treatment with either rmIL6 or rmTNFα reduced γHV68 loads in the lung to less than 5% of those in mock-treated mice, demonstrating the potent antiviral effect of rmIL6 and rmTNFα against γHV68 (Figure 1C). Importantly, we found that cytokine treatment did not affect mouse body weight (Figure S3A), spleen mass (Figure S3B), or lung cytokine levels (Figures S3C,S3D), excluding the potential side effect brought by rmIL6 and rmTNFα treatment. Furthermore, we determined whether rmIL6 and rmTNFα inhibit γHV68 lytic replication ex vivo under normal productive infection and restricted condition (in methylcellulose-containing medium). Under both conditions, treatment with rmIL6 and rmTNFα reduced γHV68 yield by 60% (Figure 1D) and plaque-forming units by 50% (Figure 1E). Collectively, these results bolster the conclusion that inflammatory cytokines, such as IL6 and TNFα, are potent antiviral effectors against γHV68 lytic replication ex vivo and in vivo.
Given that CCL5 and CXCL1 represent chemokines important for immune cell recruitment [27], [28], we surmised that the increased levels of CCL5 and CXCL1 in the lung (Figure 1A) likely translate into more robust infiltration of immune cells in Mavs−/− mice than in Mavs+/+ littermates. We examined mouse lungs from mock- or γHV68-infected (40 PFU, intranasally) mice by hematoxylin and eosin (H&E) staining. We observed similar lung architecture and no immune cell infiltration in the lung of mock-infected Mavs+/+ and Mavs−/− mice (Figures 2A,S4A; top panels). Compared to mock-infected mice, the lungs of both Mavs+/+ and Mavs−/− mice at 10 d.p.i. displayed apparent increase in cellularity (Figures 2A,S4A). In the lung of γHV68-infected Mavs+/+ mice, there was an isolated mild mixed-cell perivascular infiltration of lymphocytes, macrophages, and rare neutrophils. Strikingly, we observed an intensely increased cellularity within perivascular and peribrochial regions, caused by a massive influx of macrophages, neutrophils, and lymphocytes in the lung of γHV68-infected Mavs−/− mice. Evidently, the immune infiltrated regions extended from blood vessels and bronchioles into alveolar interstitium. It is important to note that, at 7 d.p.i., no significant immune cell infiltration was observed in the lung of either Mavs+/+ or Mavs−/− mice (data not shown), suggesting that immune cell infiltration is the consequence of the rising cytokine levels in Mavs−/− lungs at 7 d.p.i. Thus, the expansion of immune infiltrated regions in Mavs−/− lungs is likely the sequela of γHV68 infection and cytokine production thereof.
To further characterize the infiltrated immune cells, pulmonary macrophages were examined by immunohistochemistry staining using a specific antibody, anti-Iba1 (Figure S4B). We observed sporadic and evenly distributed Iba1-positive cells, likely lung-residential macrophages, in mock-infected mice (Figures 2B,S4C; top panels). Consistent with H&E staining, there were much more and larger Iba1-positive foci, in the lung of γHV68-infected Mavs−/− mice than those of Mavs+/+ littermates (Figures 2B,S4C), indicative of escalated inflammation in Mavs−/− mice. Moreover, we counted macrophages within eight randomly selected fields, and found that γHV68 infection increased lung macrophages by approximately three-fold in Mavs+/+ mice, whereas by more than five-fold in Mavs−/− mice (Figure 2C). Because of the intense staining of Iba1-positive macrophages in Mavs−/− lungs, the increase of macrophages caused by γHV68 infection is likely underestimated.
Neutrophils serve as a hallmark indicator for inflammation and CXCL1 is a major chemokine for neutrophil recruitment. We then examined neutrophils in the lungs of γHV68-infected mice by an esterase specific staining. In contrast to macrophages, neutrophils were rarely observed in the lung of mock-infected mice, or those of γHV68-infected Mavs+/+ mice (Figures 2D,S4D). However, neutrophils were easily detected within immune infiltrated regions in the lung of γHV68-infected Mavs−/− mice (Figures 2D,S4D). Counting neutrophils within eight randomly selected fields revealed that γHV68 infection increased neutrophils by approximately five-fold in the lung of Mavs−/− mice, whereas it had a negligible effect on neutrophil recruitment in Mavs+/+ mice (Figure 2E). Taken together, these results suggest that the elevated chemokine production promotes the infiltration of immune cells, including macrophages and neutrophils, into γHV68-infected lungs. Altogether, we conclude that MAVS is necessary for γHV68 to dampen cytokine production and subsequent inflammatory responses in the lung.
Professional innate immune cells, such as macrophages, are the major source to produce antiviral cytokines and MAVS is crucial for cytokine secretion from macrophages induced by intracellular pathogens [23], [24]. To test whether MAVS deficiency increases cytokine production in γHV68-infected macrophages, we isolated bone marrow-derived macrophages (BMDMs) and determined cytokines secreted by BMDMs in response to γHV68 infection. As expected, loss of MAVS greatly impaired IL6 and CXCL1 secretion in BMDMs infected by Sendai virus (SeV), a prototype RNA virus (Figure S5). Interestingly, we found that γHV68 infection induced equivalent levels of IL6 and TNFα in Mavs+/+ and Mavs−/− BMDMs, in a dose-dependent manner (Figure 3A). Moreover, loss of MAVS reduced CXCL1 secretion from γHV68-infected BMDMs, albeit with boarder-line statistical significance (Figure 3A). Thus, upon γHV68 infection, MAVS deficiency does not elevate the intrinsic ability of BMDMs to produce cytokines.
Having excluded that MAVS deficiency elevates the intrinsic cytokine production by macrophages, we surmised that the elevated cytokines in γHV68-infected Mavs−/− mice, at least during early infection (e.g., 7 d.p.i.), are likely produced by lung epithelium/fibroblasts. Next, we tested whether MEFs recapitulate the MAVS-dependent avoidance of cytokine production and examined cytokines secreted from γHV68-infected MEFs by ELISA. γHV68 infection induced significantly more IL6 and CCL5 in Mavs−/− MEFs than Mavs+/+ MEFs (Figure 3B), recapitulating the phenotypic cytokine production observed in γHV68-infected mice (Figure 1A). It is important to note that MAVS deficiency does not elevate cytokine production in mock-infected MEFs (Figure 3C). As expected, SeV infection induced much more IL6 in Mavs+/+ MEFs than in Mavs−/− MEFs, confirming the MAVS-dependent cytokine production in response to infection by a prototype RNA virus (Figure 3D). Interestingly, we found that mRNA and protein levels of cytokines were inversely correlated with the levels of γHV68 replication in the lung of BL/6 mice (Figure S6), suggesting that γHV68 inhibits cytokine production at the transcription level.
Two main signaling cascades, i.e., NFκB and IRF-IFN pathways, are known to relay MAVS-dependent innate immune activation (Figure 4A). In response to viral infection, NFκB activation downstream of MAVS and IKKβ is essential for gene expression and secretion of antiviral cytokines. Therefore, we assessed the mRNA levels of CCL5, IL6 and TNFα in γHV68-infected MEFs by quantitative real-time PCR. γHV68 infection robustly increased the mRNA abundance of all three cytokines within 6 hours post-infection (h.p.i.) in Mavs−/− MEFs, which was not observed in γHV68-infected Mavs+/+ MEFs (Figure 4B). Moreover, loss of MAVS reduced IFN-β gene expression induced by γHV68 infection (Figure S7), indicating the specificity of MAVS utilization by γHV68. Finally, γHV68 infection failed to up-regulate gene expression of these inflammatory cytokines in MEFs deficient in IKKβ and IKKγ (Figure S8), consistent with the notion that activated IKKβ is necessary for cytokine production in response to viral infection. These results suggest that MAVS is necessary for γHV68 to prevent cytokine gene expression and that up-regulated gene expression likely underpins the increased cytokine secretion in Mavs−/− MEFs. Thus, we examined NFκB activation by electrophoresis mobility shift assay (EMSA). Agreeing with the elevated cytokine gene expression in Mavs−/− MEFs, γHV68 infection imparted more robust DNA-binding activity of NFκB in nuclear extract of Mavs−/− MEFs than that of Mavs+/+ MEFs (Figure 4C). The specificity of EMSA for NFκB was confirmed by a competition assay using a cold probe and a super-shift assay using a RelA-specific antibody (Figure S9). Densitometry analysis further showed that the DNA-binding activity of NFκB, induced by γHV68 infection, in Mavs−/− MEFs was approximately two-fold of that in Mavs+/+ MEFs (Figure 4C). RelA phosphorylation at serine 536 (Ser536) was demonstrated as a marker for NFκB activation [29]. We further examined the Ser536 phosphorylated RelA by immunoblot and found that γHV68 infection also triggered a robust accumulation of the Ser536 phosphorylated RelA in Mavs−/− MEFs, but not in Mavs+/+MEFs (Figure 4D). To confirm that loss of MAVS is responsible for increased cytokine production induced by γHV68, we “reconstituted” MAVS expression by lentivirus in Mavs−/− MEFs (Figure 4E), and examined gene expression of cytokines (such as TNFα and CCL5) in response to viral infection. We found that the “reconstituted” expression of MAVS reduced cytokine gene expression triggered by γHV68 (Figure 4F), whereas up-regulated cytokine gene expression induced by VSV and SeV (Figure 4G), supporting the premise that loss of MAVS elevated γHV68-induced cytokine production. Collectively, these results indicate that MAVS is critical for γHV68 to negate NFκB activation and cytokine gene transcription in γHV68-infected cells.
To dissect the molecular mechanism by which γHV68 uncouples NFκB activation from IKKβ activation, we examined RelA protein in γHV68-infected MEFs by immunoblot analysis. RelA is the most ubiquitously and abundantly expressed subunit of the transcriptionally active NFκB dimer. We found that γHV68 infection abolished RelA protein at 4 h.p.i. Moreover, treatment by the proteasome inhibitor MG132, but not by the lysosome inhibitor chloroquine, completely restored RelA protein (Figure 5A). These results indicate that γHV68 infection induces the proteasome-dependent degradation of RelA.
We noted that γHV68 infection also induced the degradation of the inhibitor of NFκB (IκBα), which serves as an indicator of IKKβ activation (Figure 5A). To test whether IκBα degradation is necessary for γHV68-induced RelA degradation, we established Mavs+/+ MEFs stably expressing the IκBα super-suppressor, IκBαΔN, by lentivirus infection and confirmed IκBαΔN expression by immunoblot (Figure 5B). Notably, the IκBαΔN expression decreased the steady state level of RelA and we analyzed RelA protein in these two MEF cell lines separately. Surprisingly, in MEFs expressing the IκBαΔN, RelA protein gradually disappeared, albeit in a slower kinetics, with the progression of γHV68 infection (Figure 5C), indicating that IκBα degradation is dispensable for RelA degradation in γHV68-infected MEFs. As expected, IκBαΔN was not degraded in γHV68-infected MEFs, and IκBαΔN potently abrogated RelA nuclear translocation that was induced by TNFα treatment (Figure S10). RelA degradation was not observed after infection by VSV and SeV, nor after treatment with lipopolysaccharide (Figure 5D). Next, γHV68-induced RelA degradation was further examined by immunofluorescence microscopy after treatment with MG132. Consistent with the proteasome-mediated RelA degradation, MG132 treatment restored RelA protein in γHV68-infected MEFs, with significant accumulation in the cytosol (Figure 5E). Collectively, these findings support the conclusion that γHV68 induces RelA degradation in an IκBα-independent manner.
NFκB activation and cytokine gene expression induced by γHV68 appear to be transient in Mavs−/− MEFs, implying the dynamic regulation of the NFκB pathway by γHV68 infection. We assessed the kinetics of RelA protein in wild-type, Mavs−/−, Ikkβ−/−, and Ikkγ−/− MEFs infected with γHV68. Within 4 h.p.i., RelA protein gradually diminished in wild-type MEFs (Figure 6A). Remarkably, RelA protein re-appeared at 8 h.p.i. in wild-type MEFs, suggesting a transient RelA degradation triggered by γHV68 infection. However, relatively equivalent RelA protein was maintained in Mavs−/−, Ikkβ−/−, and Ikkγ−/− MEFs with the progression of γHV68 infection (Figure 6A). It is noteworthy that γHV68-induced RelA degradation in Ikkα−/− MEFs was comparable to that in Mavs+/+ MEFs (Figure S11), implying that IKKα is dispensable for RelA degradation induced by γHV68. These results imply that the integral MAVS-IKKβ signaling node is important for γHV68 to induce transient RelA degradation. When exogenous MAVS and IKKβ were re-introduced by lentivirus infection, γHV68 infection induced a transient RelA degradation (Figures 6B,6C), bolstering the specific requirement for MAVS and IKKβ in RelA degradation triggered by γHV68 infection. To determine whether elevated IKKβ is sufficient for RelA degradation induced by γHV68 infection, we expressed exogenous IKKβ with lentivirus in Mavs−/− MEFs and found that γHV68 infection failed to reduce RelA protein (Figure 6D). Thus, this result supports the conclusion that the MAVS-dependent IKKβ activation, rather than the absolute IKKβ level, is necessary for γHV68-induced RelA degradation.
To further characterize the γHV68-induced RelA degradation, we determined the half-life of RelA in γHV68-infected MEFs in the presence of cyclohexamide (CHX), an inhibitor of protein translation. For this experiment, MEFs were infected with γHV68 for 30 minutes and CHX was added to halt protein synthesis. In γHV68-infected Mavs+/+ MEFs, the half-life of RelA was reduced to approximately 2 hours (Figure 6E). CHX treatment, up to 4 hours, did not significantly alter RelA protein level in mock-infected MEFs. Furthermore, RelA protein remained relatively constant in γHV68-infected MEFs deficient in MAVS, IKKβ, or IKKγ (Figure 6E). Taken together, these results indicate that MAVS and IKKβ are necessary for γHV68 to induce rapid RelA degradation. Based on these findings, we conclude that γHV68 induces RelA degradation in a MAVS- and IKKβ-dependent, and IκBα-independent manner.
The MAVS- and IKKβ-dependent, IκBα-independent RelA degradation triggered by γHV68 infection prompted us to hypothesize that IKKα directly phosphorylates RelA to facilitate its turnover. Although IKKβ is historically known for IκBα phosphorylation and subsequent degradation, recent studies also implicated IKKβ and IKKα in phosphorylating RelA and terminating NFκB activation. Two predominant IKKβ-mediated phosphorylation sites, i.e., Serine 536 (Ser536) and Serine 468 (Ser468), have been implicated in regulating RelA turnover [29], [30], [31], [32], [33]. Because we observed significant reduction of RelA protein levels at 2 h.p.i. (Figures 6A), we reasoned that RelA phosphorylation by IKKβ precedes its degradation and focused on the first hour post γHV68 infection. To further characterize the activation of the IKKβ-NFκB ramification downstream of MAVS, we have examined TRAF6 translocation and kinase activity of IKKβ after γHV68 infection. We first determined the migration of TRAF6 into the Triton X-100-insoluable fraction that marks TRAF6 activation. Indeed, TRAF6 was detected at 30 and 60 minutes post-infection in the Triton X-100 insoluble pellet of γHV68-infected MEFs, but not in that of mock-infected MEFs (Figure 6A). We further assessed IKKβ activation by γHV68 infection with an in vitro kinase assay. Strikingly, γHV68 infection potently up-regulated IKKβ kinase activity, as demonstrated to phosphorylate the IkBa N-terminal sequence in vitro, as early as 15 min post-infection (Figure 7B). These results indicate that γHV68 infection sufficiently instigates the activation of the MAVS-IKKβ pathway.
Next, we examined the RelA phosphorylation at two IKKβ phosphorylation sites, i.e., Ser468 and Ser536, in γHV68-infected MEFs. We found an accumulation of the Ser536 phosphorylated RelA in γHV68-infected Mavs−/− MEFs, but not in Mavs+/+ MEFs (Figure 4D), which was consistent with more robust NFκB activation in Mavs−/− MEFs than those in Mavs+/+ MEFs (Figure 4C). These results suggest that Ser536-phosphorylated (Ser536p) RelA represents the activated RelA, and is selectively being targeted for degradation. We therefore infected MEFs with γHV68 and treated with MG132 to inhibit protein degradation. After treatment for 1 hour, the Ser536 phosphorylated RelA was increased by more than 8-fold, whereas total RelA was only increased approximately 2-fold (Figure 7C). This result indicates that activated RelA is being selectively degraded by γHV68 infection. Consistent with the pivotal role of the Ser468 phosphorylated form in promoting RelA degradation, we found that the Ser468p RelA gradually increased in Mavs+/+ MEFs, and conversely decreased in Mavs−/− MEFs, infected with γHV68 (Figure 7D). The Ser468p RelA, under both basal and γHV68-infected conditions, was not observed in Ikkβ−/− MEFs and severely impaired in Ikkγ−/− MEFs (Figure 7D), indicating that activated IKKβ is necessary for RelA phosphorylation at Ser468. The distinct pattern of RelA phosphorylation in γHV68-infected MEFs, dependent on MAVS expression (Figures 4D,7D), suggests the site-specificity of RelA phosphorylation instigated by γHV68. These results also agree with the requirement of IKKβ activation for RelA degradation triggered by γHV68 (Figure 6), linking RelA Ser468 phosphorylation to its degradation.
RelA phosphorylation at Ser468 represents a key step in promoting RelA ubiquitination, we then assessed the ubiquitination of RelA by immuno-precipitation with anti-RelA antibody and immunoblot with anti-ubiquitin antibody. When treated with MG132 for two hours, γHV68-infected Mavs+/+ MEFs accumulated detectable levels of high molecular species shown by the smearing of RelA, indicative of RelA ubiquitination (Figure 8A). By contrast, the smeared RelA proteins were not observed in γHV68-infected Mavs−/− MEFs (Figure 8A). In the absence of MG132, ubiquitinated RelA was not detected in both Mavs+/+ and Mavs−/− MEFs with or without γHV68 infection. The fact that RelA ubiquitination was only detected with MG132 treatment in γHV68-infected Mavs+/+ MEFs implies that RelA ubiquitination is the rate-limiting step in the process of degrading RelA.
To determine whether the Ser468 phosphorylation of RelA is necessary for its degradation, we established Mavs+/+ MEFs that stably express the RelA variant carrying the Serine 468-to-Alanine mutation (RelA.S468A) or wild-type RelA by lentivirus infection, and assessed RelA ubiquitination in γHV68-infected MEFs. We found that the exogenously expressed wild-type RelA was efficiently ubiquitinated, and that the S468A mutation abrogated the ubiquitination of the RelA.S468A variant in γHV68-infected Mavs+/+ MEFs, supporting the proposition that the Ser468 phosphorylation is necessary for efficient RelA ubiquitination (Figure 8B). When γHV68-infected Mavs+/+ MEFs were treated with MG132, the ubiquitinated RelA.S468A was observed at a lower level than the ubiquitinated wild-type RelA (Figure 8B). Next, the effect of the RelA.S468A variant on γHV68-induced RelA degradation was assessed. The exogenous RelA.S468A variant, although expressed at a lower level than wild-type RelA, was resistant to degradation induced by γHV68 infection, but wild-type RelA was not (Figure 8C). Interestingly, the exogenously expressed RelA.S468A variant also protected endogenous RelA from γHV68-induced degradation (Figure 8C), suggesting that the RelA.S468A variant functions as a dominant negative of RelA degradation.
To evaluate the contribution of transient RelA degradation to the reduced cytokine production, we determined whether the expression of the RelA.S468A variant, which inhibited RelA degradation, is sufficient to up-regulate cytokine gene expression in response to γHV68 infection. We used quantitative real-time PCR analysis to examine the mRNA levels of IL6, CCL5 and TNFα. We have noticed that the expression of the RelA.S468A variant had a marginal effect on the basal level of mRNAs of IL6 and CCL5, whereas the RelA.S468A expression elevated basal TNFα mRNA over 1,000 fold (Figure 8D). Regardless of the basal mRNA levels, the expression of the RelA.S468A variant greatly increased IL6 and CCL5 mRNA levels induced by γHV68 infection (Figure 8D). Although the basal TNFα mRNA was exceedingly high in MEFs expressing the RelA.S468A variant, γHV68 infection further increased TNFα mRNA to approximately 2,500 fold (Figure 8D). Consistent with the up-regulation of cytokine gene expression, the RelA.S468A variant also increased IL6 and CCL5 secretion in γHV68-infected MEFs (Figure 8E). Notably, despite that TNFα mRNA was highly up-regulated by RelA.S468A and γHV68 infection, TNFα secretion was under detection in MEFs (data not shown). Finally, the exogenous expression of the RelA.S468A variant reduced γHV68 lytic replication under both permissive and restricted (methylcellulose-containing) conditions (Figure S12). These results suggest that inhibiting RelA degradation is sufficient to restore cytokine gene expression and production in γHV68-infected Mavs+/+ MEFs, and demonstrate a requisite role of the Ser468 phosphorylation in efficient RelA ubiquitination and degradation induced by γHV68. Our findings uncover an essential role of MAVS in specifying the site-specific (Ser468) phosphorylation of RelA to promote RelA degradation and terminate NFκB activation, thereby effectively preventing cytokine production induced by γHV68 infection.
Recent studies have demonstrated that, in response to viral infection, the MAVS adaptor protein relays innate immune signaling from cytosolic sensors to NFκB and IRF activation that up-regulate antiviral cytokine production. Mice deficient in MAVS are severely compromised in host defense against infection of several viruses [23], [24], [34]. Moreover, viruses of the positive-stranded RNA family target MAVS for destruction to disable host innate immune responses [9], [35], [36], [37], [38]. In this study, we report that γHV68 hijacks MAVS and its immediately downstream IKKβ kinase to degrade RelA, a key subunit of the transcriptionally active NFκB dimer. In doing so, γHV68 effectively terminates NFκB activation and negates cytokine production. To our knowledge, this is the first example whereby signaling via the upstream components of the innate immune pathways, MAVS and IKKβ, is intercepted by a pathogen to degrade the critical downstream effector, RelA, a master transcription factor that governs the expression of a myriad of genes of immune function. Given the common biological properties shared by members of the gamma herpesvirus family, it is possible that human KSHV and EBV have evolved equivalent tactics to evade cytokine production.
Our observation that the MAVS-IKKβ pathway, which otherwise activates NFκB and promotes cytokine production, is directed by γHV68 to assist in degrading RelA and terminating NFκB activation is surprising. Moreover, γHV68 infection resulted in elevated cytokines in mice and fibroblasts that are deficient in MAVS, indicating that MAVS is an integral player of the active evasion scheme to abrogate host cytokine production. The elevated cytokine production in MAVS-deficient mice and MEFs, in response to γHV68 infection, is opposite to what was observed for RNA virus infection [23], [24]. Our study thus highlights an unrecognized role of MAVS and IKKβ in terminating NFκB activation. These findings also explain an early report that IL6 deficiency had no apparent effect on γHV68 infection [26], in that γHV68 effectively negates cytokine (such as IL6) production during early viral infection. Presumably, the ability of γHV68 to abolish cytokine production contributes to the minimal, if any, effect of IL6 deficiency on γHV68 lytic replication, and implies that physiological function of cytokines against γHV68 may be better defined in mouse strains (e.g., Mavs−/− mouse) that γHV68 infection induces more antiviral cytokines. Alternatively, exogenous cytokines may be delivered to γHV68-infected mice during early infection when cytokines are not produced. Indeed, we showed that treatment with IL6 and TNFα greatly reduced γHV68 replication in vivo and ex vivo, demonstrating their antiviral activity against γHV68 lytic replication.
The significance of elevated inflammatory cytokines, which were approximately two-fold of those in Mavs+/+ mice, is substantiated by escalated immune cell infiltration in the lung of Mavs−/− mice. Evidently, γHV68 infection resulted in significantly more infiltrated immune cells in the lung of Mavs−/− mice than those of Mavs+/+ mice. It is important to point out that West Nile virus was recently reported to induce an uncontrolled inflammatory response in Mavs−/− mice, including a signature of higher serum levels of multiple inflammatory cytokines [34]. The elevated cytokine production likely stems from an increased viral load, and is further compounded by the lack of negative feedback mechanisms on host immune responses. By contrast, γHV68 replicated to relatively equivalent (7 d.p.i.) or lower (10 d.p.i.) viral loads in Mavs−/− mice than in Mavs+/+ mice. This observation excludes the contribution of higher viral loads in Mavs−/− mice to the elevated cytokine production and further emphasizes the roles of MAVS in evading cytokine production by γHV68. Conceivably, loss of MAVS in lung fibroblasts/endothelial cells, which support γHV68 lytic replication, increases the cytokine production during early γHV68 infection (e.g., 7 d.p.i.). One notable common property of lung fibroblasts and embryonic fibroblasts is the abundant expression of the innate immune signaling components, providing a physiological rationale to dissect signal transduction in γHV68-infected MEFs. In fact, we recapitulated the MAVS-dependent reduction in cytokine secretion using γHV68-infected MEFs. By contrast, bone marrow-derived macrophages of Mavs−/−mice secreted either similar or lower levels of cytokines than those of Mavs+/+ mice in response to γHV68 infection. Based on these observations, we propose that γHV68-induced cytokines in the lung are produced in a biphasic process: the initial low-level production by lung fibroblasts and more robust production by infiltrated immune cells during late stages of γHV68 infection. Within this scenario, cytokines of the initial phase recruit and stimulate the proliferation of immune cells (e.g., macrophages and neutrophils) that release more cytokines during late infection, such as at 13 and 16 d.p.i., when replicating γHV68 was cleared. These results collectively indicate that loss of MAVS elevates host innate immune responses against γHV68 and support the conclusion that elevated immune responses, in turn, negatively impact the lytic replication of γHV68 in Mavs−/− mice.
Manipulating the host immune response has been a recurring theme for host-pathogen interactions [39], [40]. As intracellular pathogens, viruses are obligate to utilize host components to achieve efficient replication and dissemination. Host innate immunity is the first line of defense that plays critical roles in containing invading viruses during early viral infection. We have witnessed a growing list of strategies by which pathogens deploy to evade host innate immune responses [41], [42], [43], [44]. Many of which are evolved to thwart cytokine production or signal transduction thereof, such as IFNs. Although negating NFκB activation by pathogens has been reported previously, our study showing that MAVS and IKKβ are hijacked to promote RelA phosphorylation and degradation unravels an active evasion strategy whereby activation of upstream signaling events are exploited to nullify the immediate downstream event. Given the fundamental functions of innate immune signaling pathways in cellular physiology, the discovery that pathogens exploit these signaling events for their own benefit is not a complete surprise and emerging studies support this evolving theme [45], [46], [47].
One of the well-defined mechanisms that regulate NFκB activation is the phosphorylation and degradation of IκBs. For example, activated IKKβ phosphorylates IκBα to induce its degradation, unleashing NFκB that promotes the transcription of various genes, including IκBα. γHV68 infection appears to activate IKKβ that, in turn, triggers the degradation of both IκBα and RelA. Intriguingly, these two seemingly coupled processes are independent from each other and IκBα degradation is dispensable for RelA degradation in γHV68-infected cells. Evidently, γHV68 infection effectively uncouples NFκB activation from IKKβ activation by inducing RelA degradation in an IκBα-independent manner. Krug et al. previously reported that the expression of the IκBα super-suppressor had no apparent effect on γHV68 lytic replication [48], whereas RelA was found to inhibit γHV68 lytic replication [25]. Thus, the IκBα-independent degradation of RelA, triggered by γHV68, offers a plausible interpretation for the apparent paradoxical effect of RelA and the IκBα super-suppressor on γHV68 lytic replication. Finally, data from our studies employing the IκBαΔN super-suppressor and MG132 treatment imply that γHV68 induces RelA degradation in the cytosol, although further investigation is necessary to address this possibility. Altogether, these studies are en route to assemble an overall picture regarding the dynamic regulation of NFκB-dependent gene transcription in γHV68 lytic replication.
Besides IκBs, IKKβ differentially regulates NFκB activation via phosphorylation of two serine residues within the carboxyl terminal region of RelA, i.e., Ser468 and Ser536. Whereas the Ser536 phosphorylation of RelA potentiates NFκB-dependent gene transcription through recruiting p300 [49], the Ser468 phosphorylation instigates RelA degradation [31], [32], [33]. It is noteworthy that RelA phosphorylation at Ser536 by IKKα was reported to facilitate RelA degradation in macrophages [30]. In this study, we found that γHV68 infection induced a gradual increase of the Ser468p RelA in Mavs+/+ MEFs, whereas a gradual decrease of the Ser468p RelA was observed in γHV68-infected Mavs−/− MEFs. Importantly, the phosphorylated forms of RelA, e.g., Ser468p and Ser536p, account only a minor fraction of the total pool of RelA and that the bulky part of RelA remains constant. In fact, RelA depletion by γHV68 infection requires a high MOI (MOI = 20), although NFκB termination necessitates a relatively lower MOI (MOI = 5). The increase of the Ser468p RelA correlated with more efficient ubiquitination and rapid degradation of RelA in γHV68-infected Mavs+/+ MEFs. Conversely, the increase of the Ser536p RelA correlated with NFκB activation and the up-regulated gene expression of inflammatory cytokines in γHV68-infected Mavs−/− MEFs. The key roles of the S468 phosphorylation of RelA is further substantiated by the observations that the expression of the RelA.S468A variant sufficiently inhibited γHV68-induced RelA degradation and restored cytokine gene expression. Collectively, these results support the corollary that γHV68 usurps MAVS to facilitate RelA phosphorylation at Ser468, which primes RelA for the proteasome-mediated degradation in Mavs+/+ MEFs. We observed that MG132 treatment greatly increased the Ser536p RelA, but exhibited much less effect on total RelA protein, favoring the possibility that the Ser536p RelA may be specifically targeted for Ser468 phosphorylation in γHV68-infected Mavs+/+ MEFs. Nevertheless, our findings highlight a critical role of MAVS in the site-specific phosphorylation (Ser468) of RelA to promote RelA degradation and terminate NFκB-dependent gene transcription.
In response to γHV68 infection, the NFκB activation and cytokine gene expression in Mavs−/− MEFs is transient, suggesting that γHV68 may exploit additional unknown signaling pathways to evade NFκB-mediated cytokine production. One candidate is the MyD88-dependent pathway, which is supported by a recent study showing that KSHV utilizes TLR7 to promote viral reactivation from latency. Mouse and MEFs deficient in MAVS and/or MyD88 will enable the interrogation of these two key adaptor molecules in γHV68 infection. A closely-related and logical extension of the above question is how γHV68 is sensed to promote NFκB activation and cytokine production in MAVS-deficient mouse and MEFs. Although MAVS is dispensable, IKKβ is absolutely required for γHV68-induced NFκB activation and cytokine production, consistent with the notion that IKKβ is a major antiviral innate immune kinase responsible for NFκB activation. Additional experiments are under way to identify potential pathways that induce NFκB activation in a MAVS-independent manner.
Although IKKα and IKKβ are primarily responsible for NFκB activation downstream of a broad spectrum of physiological stimuli, IKK kinases are equally essential for NFκB termination. Through the exact same kinases, the opposing outcomes of NFκB transcription factors are likely engendered by distinct upstream signaling cascades. Therefore, biochemical studies characterizing these signaling events that differentially provoke either NFκB activation or NFκB termination will reveal insight into the mechanisms governing NFκB-mediated transcription. Such mechanisms are expected to dictate the amplitude and duration of inflammatory immune responses. It is very likely that host-virus interactions upstream of MAVS, when coupled with a downstream viral effector(s), efficiently induce RelA degradation. Several E3 ligases have been reported to promote RelA degradation under various conditions [32], [50], [51], [52], however, the cellular E3 ligases that are responsible for IKKα-dependent RelA degradation in macrophage remain unknown. Regardless of the nature of the cellular E3 ligases involved in γHV68-induced RelA degradation, it will be informative to identify the viral factor(s) that couples RelA degradation to the activation of the MAVS-IKKβ pathway during early γHV68 infection. The fact that cyclohexamide treatment, at 30 min post-infection, did not abolish γHV68-induced RelA degradation implies that a structural component(s) of the incoming virion is capable of accelerating RelA turnover. However, we can not exclude the immediately early viral gene products that are likely expressed within 30 min post-infection when cells are infected at an MOI of 20. Recently, the latent nuclear antigen LANA of γHV68 and RTA of KSHV were reported to degrade RelA and IRF3/7, respectively [53], [54]. Experiments are under way to identify viral factors (including LANA and RTA) that induce RelA degradation. Together, information obtained from these studies will provide an overview on the dynamic and intricate regulation of NFκB activation in γHV68 infection, guiding us to design better studies on human KSHV and EBV.
It is important to point out that the innate immune activation and NFκB manipulation by γHV68 occur immediately after viral infection, temporal period that we know very little. Conceivably, the effect of cytokines and the fate of incoming virions are determined within the very early phase immediately after viral infection. Indeed, γHV68 potently activated IKKβ within 15 to 30 minutes post-infection, which correlated with the Ser468 phosphorylation of RelA. Presumably, the activation kinetics of the MAVS-IKKβ pathway depends on the multiplicity of infection per cell, and higher doses of infectious γHV68 favor earlier and more robust activation of IKKβ and termination of NFκB.
We have previously demonstrated that γHV68 hijacks the MAVS-IKKβ pathway to promote viral transcriptional activation via RTA phosphorylation by IKKβ [22]. We report here that γHV68 exploits MAVS and IKKβ to promote RelA degradation and NFκB termination, thereby preventing antiviral cytokine production. Collectively, our findings argue for the corollary that γHV68 has evolved a “one stone, two birds” strategy to harness host innate immune activation through coupling viral transcription activation and cellular NFκB termination to the MAVS-IKKβ pathway, thereby enabling viral lytic replication while disabling host cytokine production (Figure 9). It is possible that RelA degradation and RTA-dependent transcriptional activation are inherently coupled to promote γHV68 lytic replication. As such, γHV68 replicates more efficiently in Mavs+/+ mice than in Mavs−/−mice. However, it should be noted that MAVS is critical for antiviral cytokine production as demonstrated by previous studies using RNA viruses and this study with isolated macrophages. Thus, the phenotypes of γHV68 infection in vivo and ex vivo likely represent “neutralized” outcome of the anti- and pro-viral activities of MAVS in γHV68 infection. Regardless, our studies uncover intricate viral exploitation mechanisms of host innate immune signal transduction.
All animal work was performed under strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Texas Southwestern Medical Center (permit number: A3472-01).
NIH 3T3 cells, 293T cells, and mouse embryonic fibroblasts (MEFs) were maintained in DMEM (Mediatech) containing 8% newborn calf serum (NCS) or fetal bovine serum (FBS), respectively. Wild-type, Mavs−/−, Ikkα−/−, Ikkβ−/−, and Ikkγ−/− MEFs were described previously [22], [23]. Bone marrow-derived macrophages (BMDMs) were obtained from Mavs+/+ and Mavs−/− littermate mice, and cultured in DMEM containing 10% FBS and 10% L929 cell-conditioned medium for 6 days before viral infection. γHV68 K3/GFP was kindly provided by Dr. Philip Stevenson (Cambridge University, UK). Wild-type γHV68 and γHV68 K3/GFP were amplified in NIH 3T3 cells, and viral titer was determined by a plaque assay using NIH 3T3 monolayer. Sendai virus (SeV) stock (Charles River Laboratories) is 4000 HA units/ml.
Wild-type (Mavs+/+) and knockout (Mavs−/−) mice were described previously [23]. Gender-matched, 6- to 8-week-old littermate mice were intranasally inoculated with 40 plaque-forming units (PFU) of wild-type γHV68. To assess cytokine production in the lung, gender- and age-matched BL/6 mice (ARC, UT Southwestern Medical Center) were intranasally infected with 1×105 PFU of γHV68. To assess the antiviral effect of IL6 and TNFα, BL/6 mice were intranasally infected with 40 PFU of γHV68, and recombinant mouse IL6 and TNFα (rmIL6 and rmTNFα, PeproTech) were intranasally administered in 30 µl of 1% BSA (Sigma) in PBS from 1 to 5 days post-infection (d.p.i.) (30 ng/mouse/day). Mouse lungs were harvested at 6 d.p.i. and homogenized in DMEM.
To determine the delivery efficiency of intranasal administration of cytokines, 1×109 relative light units (RLU) of firefly luciferase was diluted in 15 µl or 30 µl buffer (1% BSA in PBS). BL/6 mice were anaesthetized by intraperitoneally injecting a cocktail of ketamine and xylazine. Buffer alone or luciferase solution was intranasally administrated. Mouse tissues (nasal cavity, trachea, and lung) were harvested at 2 hours post administration, and homogenized by bead-beating in 300 µl passive lysis buffer (Promega). Luciferase activity was immediately quantified with the Luciferase Assay System (Promega). The delivery efficiency of intranasal administration was assessed by analyzing the relative luciferase activity in the lung.
Commercial cytokine ELISA kits used in this study include: IL6 (BD Bioscience), IL10 (BD Bioscience), TNFα (BD Bioscience), CCL5 (PeproTech), and CXCL1 (R&D Systems). Cytokine levels in mouse tissue samples or the supernatant from cultured cells were assessed according to manufacturer's instruction. Absorbance was read by FLUOstar Omega (BMG Labtech.).
Viral titer of mouse tissues or cell lysates was assessed by a plaque assay on NIH 3T3 monolayer essentially as previously described [22]. Briefly, after three rounds of freezing and thawing, 10-fold serially-diluted virus-containing supernatant was added onto NIH 3T3 cells and incubated for 2 hours at 37°C. Then, DMEM containing 2% NCS and 0.75% methylcellulose (Sigma) was added after removing the supernatant. Plaques were counted at day 6 post-infection. The detection limit for this assay is 5 PFU. To assess the antiviral effect of IL6 and TNFα, wild-type MEFs were plated at an initial cell density of 5,000 cells/cm2, and infected with γHV68 at a multiplicity-of-infection (MOI) of 0.005. DMEM containing 2 ng/ml rmIL6 or rmTNFα were added to cells at 2 hours before infection. Medium was removed at 2 hours post-infection, and cells were washed with medium and incubated in DMEM containing 2% FBS and 0.75% methylcellulose. Plaques were counted at day 5 post-infection.
Mavs+/+ and Mavs−/− littermate mice were intranasally infected with 40 PFU of γHV68 as described above. Mouse right lungs were fixed in the neutral buffered 10% formalin solution (Sigma) overnight at 4°C. Tissue specimens were dehydrated, embedded in paraffin, and cut into 3 mm sections. Lung sections were analyzed by hematoxylin and eosin (H&E), immunohistochemistry, and cytochemistry staining. Macrophages were stained with rabbit anti-Iba1 polyclonal antibody (Wako), rabbit ABC staining system (Santa Cruz), and DAB substrate kit (Vector laboratories). Neutrophils were stained with the Naphthol AS-D Chloroacetate Specific Esterase Kit (Sigma). Hematoxylin solutions for countertaining include: Gril No. 2 for macrophage staining, and Gril No. 3 for neutrophil staining.
To determine the relative levels of cytokine transcripts, RT-PCR and qRT-PCR were performed as previously reported [22]. Briefly, total RNA was extracted from MEFs or mouse tissues using TRIzol reagent (Invitrogen). To remove genomic DNA, total RNA was digested with RNase-free DNase I (New England Biolab) at 37°C for 1 hour. After heat inactivation, total RNA was re-purified with TRIzol reagent. cDNA was prepared with 1.5 µg total RNA and reverse transcriptase (Invitrogen). RNA was then removed by incubation with RNase H (Epicentre). The abundance of cytokine mRNAs and viral transcripts was assessed by qRT-PCR using 7500 Fast Real-Time PCR system (Applied Biosystems). Mouse β-actin was used as an internal control. All primers were designed by Primer Express 3.0 (Applied Biosystems) and validated individually (Table S1).
MEFs were infected with γHV68 K3/GFP (MOI = 5), and harvested at indicated time points. Cells were washed once with ice-cold PBS, scrapped into 5 ml cold PBS on ice, and centrifuged at 2,000 g, 4°C for 5 min. Cell pellets were resuspended in ice-cold hypotonic lysis buffer (10 mM Tris-HCl [pH 7.4], 150 mM NaCl, 1.5 mM MgCl2, 0.5 mM phenylmethylsulfonyl fluoride, 10 mM dithiothreitol, 0.65% Nonidet P-40). Nuclei were spun down and rinsed with ice-cold hypotonic lysis buffer without Nonidet P-40. Nuclei were resuspended in a low salt buffer (20 mM HEPES [pH 7.9], 2 mM EDTA [pH 8.0], 20 mM KCl, 1.5 mM MgCl2, 0.5 mM phenylmethylsulfonyl fluoride, 0.5 mM dithiothreitol, 25% glycerol). Then, a high salt buffer (20 mM HEPES [pH 7.9], 2 mM EDTA [pH 8.0], 800 mM KCl, 1.5 mM MgCl2, 0.5 mM phenylmethylsulfonyl fluoride, 0.5 mM dithiothreitol, 25% glycerol) was added in a dropwise fashion while stirring gently. The supernatant (nuclear extract) was collected by centrifugation at 25,000 g for 30 min at 4°C.
Nuclear extracts were analyzed for NFκB activation by EMSA. Two micrograms of nuclear extracts were incubated with a 32P-labeled oligonucleotide (Promega) containing the NFκB consensus site (5′-AGT TGA GGG GAC TTT CCC AGG C-3′) for 15 minutes at room temperature in a binding reaction containing 10 mM Tris-HCl (pH 7.5), 0.5 mM EDTA, 50 mM NaCl, 1 mM MgCl2, 0.5 mM dithiothreitol, 0.05 mg/ml poly(dI-dC) (Sigma), 4% glycerol. For the competition assay and the super-shift assay, 50-fold molar excess of cold probe or 20 µg/ml (final concentration) mouse monoclonal anti-RelA (sc-8008, Santa Cruz Biotech.) was separately pre-incubated with nuclear extracts for 10 min before adding the 32P-labeled probe. DNA-protein complexes were subjected to electrophoresis in 6% native polyacrylamide gels (0.25 × TBE) at a constant current of 9 mA. Gels were dried and analyzed by STORM 820 (Amersham Bioscience) for autoradiography.
Lentivirus production in 293T cells was carried out as previously described [22]. Briefly, 293T cells were co-transfected with the packaging plasmids (VSV-G and DR8.9) and pCDH derived plasmids expressing Flag-tagged IκBαΔN, wild-type RelA, or the S468A RelA variant carrying the Serine 468-to-Alanine mutation. At 72 hours post-transfection, supernatant was collected and passed through 0.45 mm filter. Mouse embryonic fibroblasts (MEFs) were infected with filtered lentivirus in complete DMEM containing 10 µg/ml polybrene. To maximize the infection efficiency, cells were centrifuged at 1,800 rpm, 30°C for 1 hour, and incubated at 37°C up to 6 hours. MEFs were selected and maintained in complete DMEM containing 1 µg/ml puromycin.
To assess RelA nuclear translocation, control wild-type MEFs or those stably expressing the Flag-tagged IκBαΔN were treated with 10 ng/ml TNFα for 30 minutes, or infected with γHV68 K3/GFP (MOI = 5). Thirty minutes later, γHV68-infected cells were treated with 20 µM MG132 for 2 hours or left untreated. Cells were fixed, permeabilized, stained with rabbit anti-RelA antibody and Alex 596-congugated goat anti-rabbit secondary antibody, and analyzed with confocal microscope (Leica).
Immunoprecipitation and immunoblot were as previously described [22], [55]. Briefly, cells were harvested, rinsed once with ice-cold PBS, and resuspended with RIPA buffer (50 mM Tris-HCl [pH 7.4], 150 mM NaCl, 0.5% sodium deoxycholate, 0.1% SDS, 1% NP40, 5 mM EDTA/EGTA) supplemented with protease inhibitor cocktail. Centrifuged supernatant was pre-cleared with protein A/G agarose at 4°C for one hour, and subjected to precipitation by incubating with anti-RelA antibody and protein A/G agarose, or anti-Flag M2-conjugated agarose. Precipitated proteins were extensively washed with RIPA buffer and eluted with 1 × SDS-PAGE loading buffer by boiling at 95°C for 5 - 10 min.
For immunoblot analysis, WCLs (20 µg) or precipitated proteins were resolved by SDS-PAGE, and transferred to PVDF membrane. Immunoblot detection was performed with corresponding primary antibodies as indicated by incubating at 4°C overnight and with secondary peroxidase-conjugated antibody for one hour. Proteins were visualized with SuperSignal West Pico Chemiluminescent Substrate (Thermo Scientific) and a Fujifilm LAS-3000 imaging system (FujiFilm).
Commercial antibodies used in this study include: anti-Flag (Sigma), anti-IκBα (sc-371, Santa Cruz Biotech.), rabbit anti-RelA (sc-372-G, Santa Cruz Biotech.), mouse anti-RelA (sc-8008, Santa Cruz Biotech.), anti-RelA S536p (93H1, Cell Signaling), anti-RelA S468p (Bethyl Group), anti-β-actin (Abcam.), anti-ubiquitin-conjugated protein (FK2, Affiniti Research Products), and anti-Iba1 (Wako).
The statistical significance (P-value) is calculated by unpaired two-tailed Student's t-test. *, P<0.05; **, P<0.02; ***, P<0.005. A P-value of <0.05 is considered statistically significant.
RIG-I, 230073; MDA-5, 71586; MAVS, 228607; TBK1, 56480; IKKε, 56489; IRF3, 54131; IRF7, 54123; c-Jun, 16476; ATF-2, 11909; p300, 328572; IFNβ, 15977; IKKγ, 16151; IKKα, 12675; IKKβ, 16150; IκBα, 18035; NFκB1, 18033; RelA, 19697; NFκB2, 18034; RelB, 19698; c-Rel, 19696; IL6, 16193; TNFα, 21926; CCL5, 20304; CXCL1, 14825; IL10, 16153.
|
10.1371/journal.pntd.0003064 | Dissemination of Orientia tsutsugamushi and Inflammatory Responses in a Murine Model of Scrub Typhus | Central aspects in the pathogenesis of scrub typhus, an infection caused by Orientia (O.) tsutsugamushi, have remained obscure. Its organ and cellular tropism are poorly understood.
The purpose of this study was to analyze the kinetics of bacterial dissemination and associated inflammatory responses in infected tissues in an experimental scrub typhus mouse model, following infection with the human pathogenic strain Karp. We provide a thorough analysis of O. tsutsugamushi infection in inbred Balb/c mice using footpad inoculation, which is close to the natural way of infection. By a novel, highly sensitive qPCR targeting the multi copy traD genes, we quantitatively monitored the spread of O. tsutsugamushi Karp from the skin inoculation site via the regional lymph node to the internal target organs. The highest bacterial loads were measured in the lung. Using confocal imaging, we also detected O. tsutsugamushi at the single cell level in the lung and found a predominant macrophage rather than endothelial localization. Immunohistochemical analysis of infiltrates in lung and brain revealed differently composed lesions with specific localizations: iNOS-expressing macrophages were frequent in infiltrative parenchymal noduli, but uncommon in perivascular lesions within these organs. Quantitative analysis of the macrophage response by immunohistochemistry in liver, heart, lung and brain demonstrated an early onset of macrophage activation in the liver. Serum levels of interferon (IFN)-γ were increased during the acute infection, and we showed that IFN-γ contributed to iNOS-dependent bacterial growth control.
Our data show that upon inoculation to the skin, O. tsutsugamushi spreads systemically to a large number of organs and gives rise to organ-specific inflammation patterns. The findings suggest an essential role for the lung in the pathogenesis of scrub typhus. The model will allow detailed studies on host-pathogen interaction and provide further insight into the pathogenesis of O. tsutsugamushi infection.
| Many details of the pathogenesis of scrub typhus, an infection caused by the intracellular bacterium Orientia tsutsugamushi that is endemic in Southeast Asia, have remained unclear until today. In this study, we present an experimental self-healing mouse model of scrub typhus based on footpad skin inoculation of the human pathogenic Karp strain of O. tsutsugamushi that shares many features with human infection. We established a novel quantitative PCR with increased sensitivity for the measurement of bacterial organ loads of infected mice. It was thereby shown that O. tsutsugamushi initially accumulated in the regional lymph node and subsequently spread to many organs with the highest bacterial loads found in the lung. The predominant host cells in the lung were macrophages located in the parenchymal interstitium, rather than endothelial cells. Our data also show unexpected organ-specific differences in the dynamics of macrophage activation. This mouse model will help to advance our understanding of scrub typhus pathogenesis.
| Scrub typhus, the human infection with Orientia (O.) tsutsugamushi, is a febrile, potentially lethal disease that is highly endemic in rural areas of Southeast Asia. About 1 billion people are believed to be at risk [1]. Despite detailed postmortem studies from the 1940s [2], [3] and a large number of immunological experiments conducted in the 1970s and 1980s [4], important aspects of the pathogenesis of O. tsutsugamushi infection have remained elusive.
To date, the dissemination kinetics of O. tsutsugamushi from the skin to internal organs has not been elucidated. In humans, only the eschar developing at the inoculation site is reasonably accessible for investigations of host-pathogen interactions [5]. Yet, quantitative approaches in humans assessing the involvement of internal organs are largely precluded by the invasiveness of sample retrieval, or confined to anecdotal case reports and autopsy studies. Tracking the events in the large group of subclinical or unspecific scrub typhus infections in humans, which often pass unrecognized or remain undiagnosed, will thus be challenging if not impossible. These circumstances raise the need for suitable animal models [6].
Mice belong to the wildlife host range of O. tsutsugamushi [7], [8]. A number of different genetic backgrounds and inoculation routes have been used to study O. tsutsugamushi infection in mice, including the intraperitoneal (i.p.), subcutaneous (s.c.) and intradermal (i.d.) routes [9], [10], [11], [12]. However, the mouse has so far remained an incomplete model for human infection, since no inoculation route reflects all pathogenetic details observed in humans. E.g., eschar formation is not recapitulated by any model, and i.p. infections cause a highly replicative peritonitis which primarily involves infection of peritoneal macrophages and neutrophils and may thus trigger mechanisms that are not involved in natural infection [10].
A large number of studies have analyzed O. tsutsugamushi infection after s.c. administration of large inocula (0.2 ml) [11], [13], [14], [15], which may equally not be representative for the natural i.d. infection by chiggers. We wanted to establish an inbred mouse model that is closer to the natural course of infection than previous models, and that would allow detailed immunological studies such as adoptive transfer of lymphocyte subsets. The skin of the footpad allows combined i.d. and s.c. inoculation [16], [17] of small volumes (up to 50 µl) and was thus chosen as injection site. This inoculation scheme was set up in analogy to the experimental mouse model of cutaneous leishmaniasis, where intradermal ear and combined footpad inoculations have largely comparable clinical outcomes [18].
In this study, we present a kinetic analysis of the clinical course in footpad-infected BALB/c mice, using the human pathogenic Karp strain of O. tsutsugamushi. By quantifying O. tsutsugamushi organ loads measured by a novel, highly sensitive qPCR, we provide the first evidence that O. tsutsugamushi has a distinct tropism for lung tissue in this model. Furthermore, we describe the localization and composition of histopathological changes in the lung, liver, central nervous system (CNS) and heart and demonstrate organ-specific differences in the kinetics of macrophages invasion and activation. We also provide evidence for a macrophage rather than endothelial tropism of O. tsutsugamushi in the lung and show that the infected cells preferentially reside in the parenchymal interstitium. Thus, this study provides essential and new insights into the pathogenesis of self-healing systemic infection with O. tsutsugamushi.
O. tsutsugamushi strains were obtained from Dr. J. Stenos (Australian Rickettsial Reference Laboratory, Geelong, Australia). Infected, γ-irradiated L929 cells were cultured in RPMI medium supplemented with 5% fetal calf serum (FCS), 2% glutamine and 2% HEPES buffer for 14 days and then passaged. Infectious inocula for in vivo experiments were prepared by harvesting infected cells from 14 days old cultures. Aliquots of the same stock were processed for storage in liquid nitrogen to ensure reproducibility of repeated infections. Cells were resuspended in freezing medium (45% RPMI medium, 35% FCS, 20% DMSO) and frozen in liquid nitrogen. For mock controls, non-infected L929 cells were prepared in the same way.
Methylcellulose medium contained 2/3 RPMI/5%FCS and 1/3 sterile 16.8 g methylcellulose/600 ml H2O.
Since O. tsutsugamushi organisms are delicate when isolated from host cells, infectious inocula for animal experiments consisted of L929 mouse fibroblasts infected with O. tsutsugamushi Karp. Infectivity of inocula was determined by immunofocus assay, which is based on the formation of antigen-positive foci in cell culture [19].
Briefly, 4×105 irradiated L929 fibroblasts in 24-well plates were infected with diluted inocula thawed from liquid nitrogen stocks in replicates. Cell cultures were overlayed with methylcellulose medium, continued for 14 days and fixed for 2 h with 4% paraformaldehyde. Antigen-positive foci were labeled with the 2F2 monoclonal antibody (mAb) directed against the 56 kDa surface antigen of O. tsutsugamushi (see Methods S1 for details on mAb generation) and detected with a peroxidase-labeled anti-mouse conjugate (Dianova, Hamburg, Germany). After development of substrate, spots were counted by two independent microscopists with a standard inverted microscope. The proportion of material containing 50 spot-forming units (sfu) was calculated by non-linear regression of raw data in order to determine the total number of sfu per aliquot. Comparisons between fresh and cryopreserved inocula were not performed.
In vivo experiments were carried out at the animal facility of the Bernhard Nocht Institute for Tropical Medicine in Hamburg with the permission of the Health Authorities of the State of Hamburg, Germany. Female 6–7 week-old BALB/c mice were purchased from Charles River (Sulzfeld, Germany). Mice were kept in individually ventilated cages within BSL-3 facilities. 8–10 weeks old mice received, unless otherwise stated, a 50 µl inoculum containing 5–10×103 sfu of O. tsutsugamushi Karp or uninfected irradiated L929 cells in the right hind foot pad. The course of infection was followed by assessing symptom severity by a clinical score (Fig. S1).
Organ samples were homogenized in 200 µl PBS, using Precellys ceramic beads (1.4/2.8 mm) in a Precellys homogenizer (Peqlab, Erlangen, Germany). DNA extractions from suspensions were performed using the QiaAmp DNA mini kit (Qiagen, Hilden, Germany) following the manufacturer's instructions. DNA concentrations of the extracts were determined with a Nanodrop photometer (Thermo Scientific, Wilmington, USA) and adjusted to 5 or 10 ng/µl for normalization to total tissue DNA content. Organ loads or loads of intracellular bacteria were determined by qPCR for the multi copy conjugative transfer protein D (traD) genes (traD-fw: 5′-CACAACATCCAAATGTTCAG-3′; traD-rv: 5′-GCACCATTCTTGACGAAA-3′) in a SYBR green qPCR on a Roche LightCycler 480 II instrument. In a total volume of 10 µl, the PCR reaction mix contained a final concentration of 600 nM of each primer (Tibmolbiol, Berlin, Germany), 200 µM dNTPs, 100 µg/mL bovine serum albumin (BSA), SYBR green (Invitrogen, Darmstadt, Germany) and 5 U/µL Hotstar taq DNA polymerase (Qiagen, Hilden, Germany) as well as 10 ng (blood or tissue culture samples) or 20 ng (organ samples) of template DNA. Enzyme activation at 95°C for 15 min was followed by amplification in 45 cycles of 10 s at 94°C, 15 s at 58°C and 20 s at 72°C. Each sample was measured in duplicates. Specificity of the product was confirmed by melting curve analysis.
Absolute quantification of traD copy numbers in a given sample was performed by the 2nd derivative maximum method with reference to a linearized plasmid standard. Results were depicted as log values of traD copy numbers. For calculation of genome equivalent numbers, a qPCR specific for the 56 kDa protein of O. tsutsugamushi [20] was adapted to the 2-step SYBR green format. For quantification by the 16s rRNA gene, another previously published qPCR was used [21].
A DNA extract with a known O. tsutsugamushi genome copy number was diluted serially in half-logarithmic (100.5-fold) dilutions, and 8–16 replicates of each dilution were analyzed by qPCR. The 95% detection limit (LOD95) was calculated using the PriProbit software [22].
Serum AST (aspartate aminotransferase) and ALT (alanine aminotransferase) activities were measured by using commercially available colorimetric assays (Reflotron, Roche Diagnostics, Mannheim, Germany).
For histology and immunohistochemistry (IHC) stains, standard methods were used. For immunofluorescence imaging, lung tissue was fixed with 4% paraformaldehyde overnight and frozen in cryopreservation medium (TissueTek O.C.T Compound, Sakura Finetek, Torrance, USA). Samples were sequentially reacted with goat-anti-CD31 (BD Bioscience, Heidelberg, Germany) or rabbit-anti-IBA1 (ionized calcium binding adapter molecule 1; Wako, Neuss, Germany) overnight; AlexaFluor488 donkey-anti-rabbit or AlexaFluor488 donkey-anti-goat; 2F2 mAb; AlexaFluor594 donkey-anti-mouse (Life Technologies, Darmstadt, Germany). DAPI (4′,6-diamidino-2-phenylindole; Sigma, Germany) was used for nucleus counterstains (see supplementary material for details). Sections were embedded in Fluoromount G (Southern Biotech, Birmingham, USA) and viewed with a BZ-9000 Keyence fluorescence microscope or an Olympus Confocal Microscope.
Serum samples were diluted 1∶10 in 0.1% BSA/PBS and measured by a standard sandwich ELISA for the presence of interferon (IFN)-γ (R&D Systems, Wiesbaden, Germany), according to the manufacturer's instructions.
An immortalized mouse macrophage cell line was kindly provided by Douglas Golenbock, University of Massachusetts Medical School, Worcester MA, USA [23]. O. tsutsugamushi Karp was purified from infected L929 cells 2–3 weeks post infection (p.i.) by disruption with sterile glass beads (1 mm diameter), by rocking for 5 min at 1,400 rpm on a horizontal shaker at room temperature. To remove cell debris, the suspension was centrifuged for 1 h at 1,200 rpm. The supernatant containing purified bacteria was pelleted at 4,000 rpm for 30 min and resuspended in the required volume. In 24-well plates, 2×105 macrophages were infected with purified O. tsutsugamushi. Recombinant IFN-γ (100 IU; Millipore, Billerica, USA), the inducible nitric oxide synthase (iNOS) inhibitor N-monomethylarginine (NMMA, 1 mM; Sigma, Germany) or the anti-IFN-γ mAb XMG1.2 (1 µg/ml) were added to the culture. After 3 days, DNA was extracted from detached macrophages, and the bacterial load was quantified by traD qPCR.
Data were analyzed using the Graphpad Prism 5.0 software. Descriptive statistics show mean ± SD. Hypotheses were tested by two-tailed t test, or by one-way or two-way analysis of variance (ANOVA) with Bonferroni post correction. A p value of ≤0.05 was considered significant. For quantification of stock infectivity, one-phase exponential association curves were calculated.
The animal protocol was reviewed and approved by the Animal Protection Commission and the Health Department of the State of Hamburg, Germany (approval number 74/09). The animal protocol adheres to the national guidelines as regulated by the German Animal Welfare Act.
One of the major restraints in experimental scrub typhus research has been the exactness of pathogen quantification, which has relied on light microscopy in the past [10]. For qPCR-based quantification of O. tsutsugamushi, the majority of methods are based on the quantification of single copy genes that often yield high cycle threshold (Ct) values and low bacterial copy numbers, thus compromising statistical analyses [9], [20], [24], [25].
In order to increase qPCR robustness especially for low bacterial concentrations, we designed a qPCR based on the amplification of multiple alleles of the traD (conjugative transfer protein D) gene (e.g. NCBI Gene ID 6336199) which encode for a subunit of type IV secretion systems [26]. Primer analysis by Primer-BLAST (http://www.ncbi.nlm.nih.gov/tools/primer-blast/) yielded no relevant unspecific amplification products. Negative amplification results were obtained from 27/27 non-rickettsial bacterial and fungal strains (Table S1). Occasional weak amplification was obtained from some rickettsial strains by traD, but not by 16s or 56 kDa qPCR [20], [21], suggesting that very high concentrations of rickettsial DNA might interfere with traD quantification of O. tsutsugamushi. For the purpose of experimental quantification, this specificity was satisfactory, since no cross-reactions were found with commensals, and co-infections with other rickettsiae were not part of this study.
The sensitivity of our traD qPCR was compared with two single copy qPCRs targeting the genes for 56 kDa and 16S rRNA [21]. Replicate testing of DNA extracts from O. tsutsugamushi Karp-infected cell cultures was performed in all three assays to determine the 95% limits of detection (LOD95) by probit analysis [27]. The single copy qPCRs had detection limits >10 genome copies/reaction (16s rRNA qPCR: 23.7 genome copies/reaction; 56 kDa qPCR: 11.9 genome copies/reaction; Fig. 1A). In contrast, the traD qPCR yielded a LOD95 of only 0.1 genome copies/reaction. Thus, in comparison to single copy qPCRs, our novel traD multi copy qPCR increased the sensitivity of O. tsutsugamushi Karp detection by more than 100-fold.
No prediction regarding the specificity of our traD primers for other strains of O. tsutsugamushi was possible due to lack of genomic data. Replicate testing and probit analysis was therefore performed for the other three O. tsutsugamushi strains Kato, Gilliam-like and Sido, by comparing traD and 16s qPCRs. Depending on the strain, the traD qPCR lowered the LOD95 by about 30- to 70-fold (Fig. S2). Thus, quantification by traD qPCR highly increased the sensitivity of O. tsutsugamushi detection, as shown in four genetically different strains.
It was then analyzed whether quantification by traD allows inferences about the number of bacterial copies in tissue samples. Lung samples from infected Balb/c mice collected at 7–21 days p.i. were quantified by both traD and 56 kDa qPCRs. Fig. 1B shows a high degree of correlation (R2 = 0.9463) between both quantification methods. Quantification by traD qPCR is thus a valid surrogate measure for bacterial copy numbers at both high and low bacterial concentrations. Furthermore, quantification of bacteremia in footpad-infected Balb/c mice by traD and 56 kDa qPCR at different time points showed that more blood samples were tested positive by traD compared to 56 kDa qPCR. Also, more samples yielded results above the LOD95 (Fig. 1C). This shows that the use of a multi copy qPCR is warranted to avoid false-negative results. Lower variations in repetitive testing also help to reduce the number of mice needed for statistical analysis. The traD qPCR was therefore used for all O. tsutsugamushi quantifications in this study, unless otherwise indicated.
For standardization of in vivo experiments, O. tsutsugamushi inocula have usually been quantified by mouse lethal dose 50 (MLD50) or murine infectious dose 50 measurements [9], [28]. Since repeated quantification of different lots of cryopreserved inocula by in vivo testing is ethically questionable, we developed an in vitro immunofocus assay for O. tsutsugamushi. The principle is based on immunolabeling of infected foci in methylcellulose-overlayed cell cultures (24-well format) and has proven useful for various viral pathogens [19]; one example is shown in Fig. 1D (left panel).
We wanted to know how the immunofocus assay dose related to the i.p. MLD50. Thus, Balb/c mice were infected with diluted doses of an immunofocus assay-quantified inoculum in order to determine the MLD50 by probit analysis as reported previously [29] (Fig. 1D, right panel). 1 MLD50 corresponded to 0.5 sfu. Thus, quantification of infectious units in the inoculum was similar in both methods; the immunofocus assay may thus replace MLD50 testing. A quantification of the number of single infectious O. tsutsugamushi organisms per 1 sfu was not attempted.
The human pathogenic O. tsutsugamushi Karp strain and inbred Balb/c mice were used for the establishment of our scrub typhus mouse model. When s.c. infected, Balb/c mice mount a protective immune response to otherwise lethal i.p. Karp infections [11] and thus represent an excellent model to study the development of protective immunity in scrub typhus.
The skin of the footpad allows combined i.d. and s.c. inoculation [16], [17] of volumes up to 50 µl, thus approximating the natural infection route more closely than formerly used s.c. models. An infection dose of 5,000 sfu of O. tsutsugamushi Karp was chosen in preliminary studies, since this dose produced clear symptoms of systemic infection between 2–3 weeks p.i.
To describe the course of symptoms in mice footpad-infected with O. tsutsugamushi, we created a graded scoring system that has not been described previously (Fig. S1). Uninfected L929 fibroblasts were used as mock control. As shown in Fig. 2A, footpad-infected BALB/c mice started to develop symptoms after day 14, and recovered by day 21 p.i. Mock-infected controls did not show any symptoms. Thus, an acute disease phase of about 10 days was found in our model. Skin lesions did not develop at the primary infection site. It was not investigated whether other O. tsutsugamushi strains elicit similar symptoms.
O. tsutsugamushi has been found in many organs of infected patients and experimental animals, including lung, heart, brain, liver, spleen, kidney, pancreas, appendix and skin [9], [30], [31], [32]. A clear organ preference of O. tsutsugamushi has never been revealed, and the route and kinetics of dissemination are unknown. In order to track the spread of O. tsutsugamushi, we measured the bacterial load in blood as well as secondary lymphatic and parenchymal organs every 3–4 days by traD qPCR.
O. tsutsugamushi appeared in blood at day 7 p.i., and bacteremia peaked at day 14 p.i. (Fig. 2B). Blood samples were negative on day 28 p.i. In the draining lymph node, O. tsutsugamushi was detectable at high levels at day 3 p.i., peaked at 7 days p.i. and steadily declined afterwards (Fig. 2C). Maximum spleen loads were reached at day 10 p.i.
Pneumonia, hepatitis, myocarditis and meningoencephalitis are common manifestations of solid organ involvement in scrub typhus. We therefore analyzed the dynamics of bacterial loads in lung, liver, heart and brain. The highest loads were found after day 10 in the lung, about 50-fold higher than in the brain (Fig. 2D, left panel). The organ affected with the second-highest load was the heart, where the peak was between 10 and 14 days p.i. At day 14 p.i., the heart load was still 7-fold higher compared to the liver (Fig. 2D, right panel).
Since bacteremia was maximal at day 14 p.i., the bacterial loads of ten different organs retrieved at that time point were compared (Fig. 2E). O. tsutsugamushi was detectable in all organs, including intestine and bone marrow. Interestingly, heart and lung had surpassed the bacterial load of the popliteal lymph node, while all others remained below. At 14 days p.i., lung loads were significantly higher compared to any other organ.
The results show that after footpad infection, O. tsutsugamushi initially accumulated in the regional lymph node. O. tsutsugamushi then spread to internal organs, but had a prominent tropism for lung and heart tissue, while lower bacterial loads were found in liver and brain.
Following the quantification of bacterial loads of infected organs, the histopathological changes during infection were analyzed. Liver involvement is a common feature in acute scrub typhus [32], [33], [34], [35]. Fig. 3A shows that early in infection, inflammation mainly affected the periphery of the hepatic lobule, while after 14 days p.i., the liver showed panlobular inflammation (Fig. 3A, upper row, arrowheads). At 21 days p.i., inflammation had begun to resolve, and mainly centrilobular foci of infiltrating cells were seen. Inflammation of portal fields started at day 7 and was maximal at day 14 p.i. (Fig. 3A, lower row; arrowhead).
Lesions in the centrilobular areas developing between day 14 and 21 p.i. were associated with a collapse of the reticular fiber network (Fig. S3, left panels; arrowheads), a sign for hepatocyte destruction. The periportal inflammations, which appeared earlier in infection (Fig. 3A), were also accompanied with a collapse of reticular fibers (Fig. S3, right panels; arrowhead).
Elevations of liver transaminases in serum showed kinetics similar to the cellular infiltration:
While on day 7 p.i. no changes were seen, serum AST activity at day 14 p.i. increased by 3–4 fold. AST activity was still elevated on day 21 p.i. (Fig. 3B, left panel). In contrast, serum ALT levels did not change significantly during infection (Fig. 3B, right panel).
Our findings show that O. tsutsugamushi infection in the footpad Balb/c mouse model causes mild, self-limiting hepatitis. Since transaminase levels peaked together with the maximum of cellular infiltration, and areas of hepatic inflammation coincided with hepatocyte loss, an immunopathological mechanism may be the cause of scrub typhus hepatitis.
Meningitis or meningoencephalitis are common complications of scrub typhus in humans [36], [37]. Whole brain samples from BALB/c mice footpad-infected with O. tsutsugamushi collected at 7, 14 and 21 days p.i. were processed for histopathology. Inflammatory alterations discernible in HE stains appeared at day 21 p.i. Lesions were found in the meninges and the brain parenchyma, e.g. in the thalamic region (Fig. 4A).
We analyzed the cellular composition of these lesions with IBA1, iNOS and CD3 stains (Fig. 4A). In the CNS, IBA1 was originally thought to specifically stain microglial cells, especially activated microglia [38], [39], but later studies showed that blood-derived macrophages equally stain positive for this marker under inflammatory conditions [40]. IBA1 does therefore not differentiate between both cell types and should thus be regarded as a marker for all professional phagocytic cells in the CNS.
The meninges were thickened and showed infiltrations of CD3-positive T cells and IBA1-positive cells. In the parenchyma, two different lesion patterns were seen: nodular structures with actual parenchymal infiltration of T cells (Fig. 4A, middle panels), and tight vascular cuffs associated with larger gauge blood vessels (Fig. 4A, right panels). Interestingly, the highest content of iNOS-positive cells was seen in the parenchymal noduli. Only isolated B cells, and no neutrophils or astrocytes contributed to the formation of these lesions (data not shown). Increased abundance of glial cells was observed for both microglia and GFAP (glial fibrillary acidic protein)-positive astroglia at day 21 p.i. (Fig. 4B).
Thus, activation of glial cells in the brain parenchyma, with contribution of both astrocytes and microglia/macrophages, as well as infiltration of T cells with a perivascular predominance were the hallmarks of CNS inflammation during O. tsutsugamushi infection. The nodular infiltrative lesions had a higher content of activated iNOS-expressing cells compared to meninges and vascular cuffs.
Since O. tsutsugamushi preferentially infected the lung, we also investigated the development of pulmonary pathology. Whole lung samples from BALB/c mice footpad-infected with O. tsutsugamushi were taken for histopathological examination 7, 14 and 21 days p.i.
During the first two weeks after infection, hematoxylin/eosin (H&E)-stained sections did not show conspicuous alterations. At day 21 p.i. the infected animals developed inflammatory infiltrates, while L929-mock-treated animals showed no pathological changes (Fig. 5A, details 1–3). The cellular infiltrates were typically found in peribronchial areas, in the parenchyma and in the visceral pleura. Peribronchial lesions showed orientation towards the adjacent arterial blood vessels (Fig. 5A, detail 4), the characteristic localization of inducible bronchus-associated lymphatic tissue (BALT). Lesions in the alveoli had a nodular appearance (Fig. 5A, detail 5). The visceral pleura was focally invaded by inflammatory infiltrates (Fig. 5A, detail 6).
For a detailed analysis of the cellular composition, pulmonary serial sections were stained for macrophages, iNOS, neutrophils, B cells and T cells. As shown in Fig. 5B (left panels), macrophages formed a unicellular layer around bronchi and, together with T cells, were the predominant cell type in BALT lesions. Some neutrophils were found in these areas, and B cells appeared focally but in small numbers. Although these structures had the typical localization of BALT, they contained only very small B cell areas, unlike the larger germinal centers found in BALT caused by other pulmonary infections [41], [42]. The nodular alveolar lesions contained macrophages, but also neutrophils and T cells (Fig. 5B, middle panels). Notably, Ly6G-stained cells appeared relatively large and fuzzy in these regions. B cells were barely detected. According to their localization, these lesions correspond to nodular inflammatory foci (NIFs), similar to lesions described in cytomegalovirus-infected lungs [43]. The lesions in the pleura had a similar cellular composition (Fig. 5B, right panels).
Quantification of IBA1-positive cells showed that all three lesion sites had a similar macrophage content (Fig. 5C). However, significantly more iNOS-positive cells and a higher iNOS-/IBA1 ratio were found in NIF and pleura lesions, compared to the perivascular BALT (Fig. 5C). This observation parallels our findings in the brain, where the infiltrative parenchymal lesions also had a higher content of iNOS-expressing cells (Fig. 4A). This suggests that activation of macrophages may be differently regulated in perivascular and infiltrative inflammatory sites, resulting in stronger iNOS induction in the latter. Possibly, even cells other than macrophages may become activated and express iNOS in these regions (Fig. 5C, right panel).
In summary, it was found that the peak of pulmonary invasion by O. tsutsugamushi was followed by pleural, parenchymal and peribronchial infiltrations. Perivascular orientation was a typical finding. The peribronchial/perivascular BALT lesions consisted of T cells and iNOS-negative macrophages, while NIFs in the lung parenchyma and pleural lesions harbored significantly more iNOS-positive macrophages, as well as neutrophils and few B or T cells.
We assumed that at the time of maximal O. tsutsugamushi DNA load, intact bacteria were still present within lung tissue. To demonstrate their localization, fixed cryosections of lung samples from day 14 p.i. were examined by immunohistochemistry, using the 2F2 mAb and co-staining for the endothelial marker CD31 or IBA1 in order to differentiate between an endothelial or macrophage tropism.
As shown in Fig. 6A, O. tsutsugamushi antigen was found scattered in spatially separated compartments of the lung (Fig. 6A, details 1–3). Three typical localizations were identified. Intact bacteria with coccoid shape, in some cases with a clear distinction of the bacterial cell wall, were demonstrable in lung parenchyma (Fig. 6A, detail 1). The largest accumulation of antigen was found in the pleura and perivascular BALT (Fig. 6A details 2 and 3). These structures consisted mostly of fuzzy aggregates of bacterial antigen, suggesting the presence of fragmented or degraded bacterial remnants.
Despite a close spatial relationship between bacteria and CD31-positive blood vessels, confocal imaging revealed that bacteria were merely in very close contact to endothelial cells. They protruded from infected cells crossing the endothelium, but were not located intracellularly (Fig. 6B). True intraendothelial infection was not found at any instance.
In perivascular BALT (Fig. 6C) and pleura, the majority of structures stained by the 2F2 mAb were extracellular. In BALT areas, only a small number of IBA1-positive cells harbored singular intracellular, round-shaped bacteria (Fig. 6D, upper panels, arrow). While these cells expressed IBA1 on their surface, intracellular IBA1 co-localized to the bacteria. The concentration of extracellular antigen in these regions (Fig. 6D, upper panels, arrowhead) may have been caused by successful degradation and externalization of degraded bacteria.
Contrarily, in the lung parenchyma, IBA1-positive macrophages harboring much larger numbers of intracellular bacteria were identified. IBA1 expression in these cells was mainly intracellular and co-localized with the majority of intracellular bacteria (Fig. 6D, lower panels).
Our data show that in pleura and BALT, only solitary bacteria inside macrophages were present. O. tsutsugamushi antigen accumulated in the extracellular space in these areas, possibly after exocytosis of degraded bacterial remnants. This bacterial degradation was likely mediated by macrophages. Infected macrophages harboring much larger numbers of intact intracellular bacteria, in contrast, were found in the lung parenchyma. In these cells, IBA1 expression was focused around the intracellular bacteria rather than on the cell membrane. Importantly, no bacteria were found in CD31-positive endothelial cells despite a close spatial relationship.
Lung, heart, liver and CNS are commonly involved in the clinical symptomatology of scrub typhus. In our model, they showed differences in O. tsutsugamushi organ loads and clearance dynamics (Fig. 2D). We wanted to know whether these differences relate to different kinetics of macrophage appearance and activation. IHC for IBA1 and iNOS was thus performed on tissue samples collected at day 7, 14 and 21 p.i. from footpad-infected Balb/c mice or mock-infected controls (Fig. 7A).
In serial sections of liver, heart and lung, iNOS staining co-localized with focal aggregates of IBA1-positive macrophages (Fig. 7A). In the brain, clear morphological alterations of IBA1-positive cells were present on day 14 p.i., when some cells had assumed a plumper shape and showed a lower degree of ramification (Fig. 7A). This morphology resembled amoeboid microglia, consistent with an activated state. Few iNOS-positive cells were present in the brain, but they also co-localized with IBA1 staining (Fig. 4A).
In order to compare the dynamics of macrophage activation, the relative tissue contents of iNOS- and IBA1-positive cells were quantified (Fig. 7B). Of the four organs analyzed, the macrophage response was fastest and strongest in the liver. Here, a 3-fold increase of macrophage content was present at day 7 p.i. Hepatic macrophage infiltration reached a peak of 25% at day 14 p.i. and decreased thereafter. In comparison, a significant increase of cardiac macrophages set in between day 7 and 14 p.i. but remained low compared to other organs.
The macrophage response in lung and brain was protracted, with highest values at day 21 p.i.
Induction of iNOS was strongest in the liver with an iNOS/IBA1 ratio of about 6% at day 14 p.i., In heart and brain, the iNOS/IBA1 ratio remained below 0.5%. In the lung, iNOS-positive macrophages were present from day 14 p.i., with a predominance in the alveolar interstitium. By day 21 p.i., iNOS-positive macrophages in the lung were mainly found in NIFs and pleura, and to a significantly lower extent in BALT areas (Fig. 5B,C). At that time, more iNOS-positive cells per area were present, but the overall iNOS/IBA1 ratio in the lung had not further increased (Fig. 7B). Overall, brain and heart were infiltrated by a very low percentage of iNOS-positive macrophages.
These findings suggest that the strongest and most rapid macrophage response occurred in the liver, while the response was reduced or delayed in heart, lung and brain. This strong response was associated with a more efficient bacterial growth control in the liver, compared to significantly higher bacterial loads in heart and lung (Fig. 2D,E). This pattern was not observed in the brain, where a delayed infiltration of IBA1-positive cells was observed despite an efficient bacterial clearance, suggesting that other antibacterial mechanisms may become involved.
Bacterial products are able to induce the expression of iNOS in macrophages [44], [45]. As seen in our experiments, the mere presence of O. tsutsugamushi organisms was not sufficient for iNOS expression in all organs, e.g. in the lung. IFN-γ is the most important synergistic activator of nitric oxide (NO) production by macrophages [46] and was shown to be produced during O. tsutsugamushi infection in other models [47], [48]. The systemic availability of significant amounts of IFN-γ in our infection model was demonstrated by ELISA from serum samples at day 14 p.i. (Fig. 7C). To analyze whether IFN-γ is able to contribute at all to bacterial elimination via the induction of iNOS in O. tsutsugamushi infection, an in vitro approach was chosen. Infected macrophages were treated with recombinant IFN-γ in the presence or absence of the iNOS inhibitor NMMA. As shown in Fig. 7C, NMMA reversed the anti-bacterial effects of IFN-γ.
IFN-γ-dependent iNOS induction is thus an important antibacterial effector mechanism in O. tsutsugamushi infection. The expression of iNOS in infected tissues does not correlate with the systemic availability of IFN-γ. It is possible that IFN-γ is made available locally by contact-dependent secretion to infected host cells e.g. by T lymphocytes, thereby inducing iNOS expression.
In summary, these data highlight profound differences between macrophage infiltration and activation in different target organs during acute infection with O. tsutsugamushi in the mouse model. The induction of iNOS seems to be subjected to a tight temporal and spatial regulation. Although IFN-γ-induced iNOS production contributes to bacterial degradation, the serum levels of IFN-γ were not sufficient to explain the iNOS expression pattern in different target organs.
In this study we present the kinetic analysis of a scrub typhus mouse model in inbred BALB/c mice, using the hind footpad as inoculation site. By a novel qPCR using conserved traD sequences as primer targets, we provide a detailed analysis of bacterial dissemination and show that O. tsutsugamushi infects internal organs to variable degrees, of which the lung becomes the major target organ during systemic spreading. By immunofluorescence, we show that O. tsutsugamushi has a predominant macrophage rather than endothelial tropism in the lung. Moreover, we identified specific lesions in lung, liver, heart and CNS and demonstrate differences in the kinetics of macrophage responses between these important target organs.
By the novel traD qPCR, an about 100-fold increased sensitivity in detection of the Karp strain was obtained compared to single copy gene qPCRs. Importantly, the traD qPCR also detected three other strains with increased sensitivity. We furthermore showed that results from traD quantifications correlated with results from single copy gene assays. In our model, the new qPCR allowed the comparison of bacterial loads in small samples containing only low bacterial copy numbers. This strategy proved valuable for our defined infection model. Its use in clinical diagnostics, however, is limited at this stage, due to the unknown extent of genomic variations between wild type isolates and a low degree of cross reactions with other rickettsial pathogens.
We furthermore established and analyzed the inbred Balb/c mouse model of scrub typhus. To more closely approximate the natural transmission route, a footpad infection model was chosen. Footpad inoculation allows a combined i.d./s.c. administration [16], [17] of small volumes and thus approximates the natural inoculation route by dermal chigger bites to a higher degree than earlier studies that relied on the s.c. or i.p. infection routes [11], [13], [14], [15].
The novel traD qPCR was used to measure bacterial organ loads after Karp strain infection over a course of four weeks. While a recent study analyzed the dissemination of three O. tsutsugamushi strains by qPCR in outbred mice during the first week of infection [9], the present study is, to our knowledge, the first analysis on quantification of O. tsutsugamushi in an experimental infection that encompasses the entire acute phase including pathogen dissemination and clearance. As shown here, O. tsutsugamushi first accumulated in the regional lymph node before it spread in higher numbers to internal organs, suggesting two distinct phases of dissemination. The highest loads of O. tsutsugamushi were found in lung and heart tissue. With regard to the similar curves of blood and organ loads, dissemination may have occurred hematogenically, but especially the lung remained positive after completion of bacteremia. Whether this protracted course in the lung really reflects delayed bacterial clearance or ongoing replication is not certain, since we do not know to what degree dead bacteria influence qPCR results. Possibly, quantification of bacterial mRNA rather than DNA may more accurately reflect bacterial replication. It will also be interesting to know whether other infection routes such as i.p. or i.v. infection predispose for different dissemination kinetics.
In the lung, two types of IBA1-positive macrophages with distinct localizations were identified as the main host cells. O. tsutsugamushi was found as single coccoid bacteria in macrophages within BALT; here, IBA1 was expressed on the membrane and also co-localized to intracellular bacteria. It was shown that IBA1 is translocated from the cytosol to the membrane early during phagocytosis, where it is cross-linked to filamentous actin and becomes a significant component of the phagocytic cup [49], [50]. Our finding therefore suggests that BALT macrophages contribute to bacterial phagocytosis. Contrarily, highly infected macrophages were found in the parenchyma. IBA1 expression in these cells was mainly intracellular, but the reason for the absence of IBA1 from the cell membrane in these cells remains unclear. Our findings parallel previous reports on macrophages as important host cells in humans [5], [30].
While we found bacteria that were spatially associated with lung endothelia, no true intra-endothelial infection was recognized in our model. Endothelial infection was reported to be a hallmark of lethal human infection [30], but it is not known to what degree endothelial infection occurs during self-limiting courses of scrub typhus [5].
Since pneumonia, myocarditis, hepatitis and meningoencephalitis are important determinants of morbidity in scrub typhus, we characterized the histopathological changes in lung, heart, liver and CNS. In the lung, three different types of infiltrates were defined: perivascular BALT, parenchymal noduli and pleuritic lesions. By immunofluorescence analysis we found that mainly degraded, extracellular bacterial antigen accumulated in developing BALTs and pleural lesions already at 14 days p.i., possibly as a consequence of exocytosis of bacterial remnants. In contrast, infected cells with large numbers of intracellular bacteria were mainly identified in the parenchyma. Thus, de novo formed BALT and pleuritic infiltrates could contribute to early bacterial degradation, while solitary infected cells in the parenchyma may have escaped immunosurveillance during the first two weeks of infection. However, inflammatory noduli appeared in the parenchyma by day 21 p.i. and showed a high content of iNOS-positive macrophages. Similar parenchymal structures were shown to actively contribute to pathogen clearance in a mouse model of cytomegalovirus infection and were recently termed nodular inflammatory foci (NIFs) [43]. It is possible that NIFs form de novo as specialized compartments to fight infected cells that have escaped from immune surveillance.
The pulmonary lesions recapitulate important features of pulmonary lesions in human scrub typhus [51], [52], [53]. Subpleural, interlobular septal thickening and peribronchial infiltrates demonstrated by computer tomography [52], [54] closely resemble the morphology of our findings in murine lungs. Similar to the murine infection, pulmonary inflammation occurs rather late, being usually progressive until one week after the onset of symptoms [55].
Myocardial lesions consisted mainly of macrophages and were mainly located interstitially. Further studies on scrub typhus myocarditis will have to address which cell types or subsets are either required for pathogen clearance or are responsible for myocardial damage that may lead to myocardial infarction or myocarditis [56], [57].
In the liver, transient periportal inflammation with hepatocyte loss was found as sign of acute infection, similar to human scrub typhus hepatitis [32], [58], [59]. Interestingly, transaminase elevations in infected mice involved AST rather than ALT, thus reflecting the findings of clinical studies [34], [60]. In comparison to the lung or the heart, proliferation of O. tsutsugamushi was low in the liver, paralleled by early macrophage activation. While the degree of liver tissue destruction was mild in our self-limiting model of scrub typhus in BALB/c mice, the strong and early onset of the hepatic immune response suggests a protective role for the liver during scrub typhus. In humans, the strong association between underlying liver cirrhosis and fatal outcomes of scrub typhus [61] underlines that a functional hepatic immune response might be a critical determinant of favorable disease courses which deserves further study.
Central nervous symptoms in scrub typhus are common [36], [37] and have early been linked to inflammation in brain and meninges [62], [63]. In our murine model, CNS lesions appeared after the bacterial load had already declined. Leptomeningeal lesions were mainly composed of macrophages and T cells. In brain parenchyma, lesions equally consisted of macrophage/microglia and T cells. Both tight vascular cuffs and infiltrative lesions were seen, suggesting a breakdown of the blood-brain barrier. Leukocyte infiltration into the parenchyma has been shown to involve temporary residency in perivascular cuffs before the outer parenchymal basement membrane is crossed [64], [65]. Probably, the vascular cuffs and the infiltrative nodule represent two subsequent steps of leukocyte invasion into the CNS, the latter giving rise to stronger induction of iNOS.
This study also shows for the first time the kinetics of phagocyte infiltration and activation in important target organs of O. tsutsugamushi by histopathology. Interestingly, the maximum of IBA1-positive phagocyte invasion was observed earlier in the liver compared to heart, brain and lungs. In general, invasion of leukocytes has to be preceded by transmigration through the endothelia. The different dynamics may be linked to the differential permissiveness of fenestrated endothelia in the liver on the one hand, allowing a rapid reaction of Kupffer cells, and the tight endothelia e.g. in the brain on the other [66], [67]. It remains unexplained, however, why the macrophage reaction in the lung is delayed. Moreover, the IBA1 stain could not differentiate between tissue-resident macrophages (e.g. Kupffer cells, microglia) and monocyte-derived macrophages. Sophisticated models will be needed to differentiate the origin of macrophages and mutual contributions to organ pathology and pathogen clearance [68]. The relative content of iNOS-positive macrophages was remarkably low in heart and brain. Other effector mechanisms may be preferentially induced in order to avoid iNOS-mediated tissue damage, but further studies have to show this. We furthermore showed that IFN-γ induces an iNOS-dependent reduction of bacterial growth in macrophages in vitro. IFN-γ serum levels were elevated during the acute infection, but correlated poorly with iNOS expression in target organs. The measurement of IFN-γ production in the respective tissues, e.g. by mRNA expression analysis, may more accurately correlate with local iNOS expression.
In conclusion, we have analyzed an experimental mouse infection that closely approximates the natural transmission and shows a strong tropism of O. tsutsugamushi for lung and heart tissue. While only the immune response towards the Karp strain of O. tsutsugamushi was analyzed in the present study, it will be interesting to assess the effect of other, especially less pathogenic, strains. This model will be useful to understand better the immunology and pathogenesis of scrub typhus.
|
10.1371/journal.pcbi.1000236 | Spontaneous Reaction Silencing in Metabolic Optimization | Metabolic reactions of single-cell organisms are routinely observed to become dispensable or even incapable of carrying activity under certain circumstances. Yet, the mechanisms as well as the range of conditions and phenotypes associated with this behavior remain very poorly understood. Here we predict computationally and analytically that any organism evolving to maximize growth rate, ATP production, or any other linear function of metabolic fluxes tends to significantly reduce the number of active metabolic reactions compared to typical nonoptimal states. The reduced number appears to be constant across the microbial species studied and just slightly larger than the minimum number required for the organism to grow at all. We show that this massive spontaneous reaction silencing is triggered by the irreversibility of a large fraction of the metabolic reactions and propagates through the network as a cascade of inactivity. Our results help explain existing experimental data on intracellular flux measurements and the usage of latent pathways, shedding new light on microbial evolution, robustness, and versatility for the execution of specific biochemical tasks. In particular, the identification of optimal reaction activity provides rigorous ground for an intriguing knockout-based method recently proposed for the synthetic recovery of metabolic function.
| Cellular growth and other integrated metabolic functions are manifestations of the coordinated interconversion of a large number of chemical compounds. But what is the relation between such whole-cell behaviors and the activity pattern of the individual biochemical reactions? In this study, we have used flux balance-based methods and reconstructed networks of Helicobacter pylori, Staphylococcus aureus, Escherichia coli, and Saccharomyces cerevisiae to show that a cell seeking to optimize a metabolic objective, such as growth, has a tendency to spontaneously inactivate a significant number of its metabolic reactions, while all such reactions are recruited for use in typical suboptimal states. The mechanisms governing this behavior not only provide insights into why numerous genes can often be disabled without affecting optimal growth but also lay a foundation for the recently proposed synthetic rescue of metabolic function in which the performance of suboptimally operating cells can be enhanced by disabling specific metabolic reactions. Our findings also offer explanation for another experimentally observed behavior, in which some inactive reactions are temporarily activated following a genetic or environmental perturbation. The latter is of utmost importance given that nonoptimal and transient metabolic behaviors are arguably common in natural environments.
| A fundamental problem in systems biology is to understand how living cells adjust the usage pattern of their components to respond and adapt to specific genetic, epigenetic, and environmental conditions. In complex metabolic networks of single-cell organisms, there is mounting evidence in the experimental [1]–[6] and modeling [7]–[14] literature that a surprisingly small part of the network can carry all metabolic functions required for growth in a given environment, whereas the remaining part is potentially necessary only under alternative conditions [15]. The mechanisms governing this behavior are clearly important for understanding systemic properties of cellular metabolism, such as mutational robustness, but have not received full attention. This is partly because current modeling approaches are mainly focused on predicting whole-cell phenotypic characteristics without resolving the underlying biochemical activity. These approaches are typically based on optimization principles, and hence, by their nature, do not capture processes involving non-optimal states, such as the temporary activation of latent pathways during adaptive evolution towards an optimal state [16],[17].
To provide mechanistic insight into such behaviors, here we study the metabolic system of single-cell organisms under optimal and non-optimal conditions in terms of the number of active reactions (those that are actually used). We implement our study within a flux balance-based framework [18]–[23]. Figure 1 illustrates key aspects of our analysis using the example of Escherichia coli. For any typical non-optimal state (Figure 1B), all the reactions in the metabolic network are active, except for those that are necessarily inactive due either to mass balance constraints or environmental conditions (e.g., nutrient limitation). In contrast, a large number of additional reactions are predicted to become inactive for any metabolic flux distribution maximizing the growth rate (Figure 1A). This spontaneous reaction silencing effect, in which optimization causes massive reaction inactivation, is observed in all four organisms analyzed in this study, H. pylori, S. aureus, E. coli, and S. cerevisiae, which have genomes and metabolic networks of increasing size and complexity (Materials and Methods). Our analysis reveals two mechanisms responsible for this effect: (1) irreversibility of a large number of reactions, which under intracellular physiological conditions [14] is shared by more than 62% of all metabolic reactions in the organisms we analyze (Table 1 and Note 1); and (2) cascade of inactivity triggered by the irreversibility, which propagates through the metabolic network due to stoichiometric and flux balance constraints. We provide experimental evidence of this phenomenon and explore applications to data interpretation by analyzing intracellular flux and gene activity data available in the literature.
The drastic difference between optimal and non-optimal behavior is a general phenomenon that we predict not only for the maximization of growth, but also for the optimization of any typical objective function that is linear in metabolic fluxes, such as the production rate of a metabolic compound. Interestingly, we find that the resulting number of active reactions in optimal states is fairly constant across the four organisms analyzed, despite the significant differences in their biochemistry and in the number of available reactions. In glucose media, this number is ∼300 and approaches the minimum required for growth, indicating that optimization tends to drive the metabolism surprisingly close to the onset of cellular growth. This reduced number of active reactions is approximately the same for any typical objective function under the same growth conditions.
We suggest that these findings will have implications for the targeted improvement of cellular properties [24]. Recent work predicts that the knockout of specific enzyme-coding genes can enhance metabolic performance and even rescue otherwise nonviable strains [25]. The possibility of such knockouts bears on the issue of whether the inactivation of the corresponding enzyme-catalyzed reactions would bring the whole-cell metabolic state close to the target objective. Thus, our identification of a cascading mechanism for inducing optimal reaction activity for arbitrary objective functions provides a natural set of candidate genetic interventions for the knockout-based enhancement of metabolic function [25].
We model cellular metabolism as a network of metabolites connected through reaction and transport fluxes. The state of the system is represented by the vector v = (v1,…,vN)T of these fluxes, including the fluxes of n internal and transport reactions, as well as nex exchange fluxes for modeling media conditions. Under the constraints imposed by stoichiometry, reaction irreversibility, substrate availability, and the assumption of steady-state conditions, the state of the system is restricted to a feasible solution space (Materials and Methods). Within this framework, we first consider the number of active reactions in a typical non-optimal state v∈M.
We can prove that, with the exception of the reactions that are inactive for all v∈M, all the metabolic reactions are active for almost all v∈M, i.e., for any typical state chosen randomly from M (Text S1, Section 1). Accordingly, the number n+(v) of active reactions in a typical non-optimal state is constant, i.e.,(1)The reactions that are inactive for all states are so either due to mass balance or environmental conditions, and can be identified numerically using flux coupling [26] and flux variability analysis [9].
We now turn to the maximization of growth rate, which is often hypothesized in flux balance-based approaches and found to be consistent with observation in adaptive evolution experiments [31]–[34]. Performing numerical optimization in glucose minimal media (Materials and Methods), we find that the number of active reactions in such optimal states is reduced by 21%–50% compared to typical non-optimal states, as indicated in the middle bars of Figure 2. Interestingly, the absolute number of active reactions under maximum growth is ∼300 and appears to be fairly independent of the organism and network size for the cases analyzed. We observe that the minimum number of reactions required merely to sustain positive growth [7],[8] is only a few reactions smaller than the number of reactions used in growth-maximizing states (bottom bars, Figure 2). This implies that surprisingly small metabolic adjustment or genetic modification is sufficient for an optimally growing organism to stop growing completely, which reveals a robust-yet-subtle tendency in cellular metabolism: while the large number of inactive reactions offers tremendous mutational and environmental robustness[52], the system is very sensitive if limited only to the set of reactions optimally active. The flip side of this prediction is that significant increase in growth can result from very few mutations, as observed recently in adaptive evolution experiments [35].
We now turn to mechanisms underlying the observed reaction silencing, which is spread over a wide range of metabolic subsystems (see Figure 1 for E. coli). The phenomenon is caused by growth maximization via reaction irreversibility and cascading of inactivity.
Although we have focused so far on maximizing the biomass production rate, the true nature of the fitness function driving evolution is far from clear [44]–[47]. Organisms under different conditions may optimize different objective functions, such as ATP production or nutrient uptake, or not optimize a simple function at all. In particular, some metabolic behaviors, such as the selection between respiration and fermentation in yeast, cannot be explained by growth maximization [48]. Other behaviors may be systematically different in situations mimicking natural environments [49]. Moreover, various alternative target objectives can be conceived and used in metabolic engineering applications. We emphasize, however, that while specific numbers may differ in each case, all the arguments leading to Eqs. (2)–(4) are general and apply to any objective function that is linear in metabolic fluxes. To obtain further insights, we now study the number of active reactions in organisms optimizing a typical linear objective function by means of random uniform sampling.
We first note the fact (proved in Text S1, Section 4) that with probability one under uniform sampling, the optimal solution set Mopt consists of a single point, which must be a “corner” of M, termed an extreme point in the linear programming literature. In this case, dopt = 0, and Eq. (2) becomes(5)With the additional requirement that the organism show positive growth, we uniformly sample these extreme points, which represent all distinct optimal states for typical linear objective functions.
We find that the number of active reactions in typical optimal states is narrowly distributed around that in growth-maximizing states, as shown in Figure 4. This implies that various results on growth maximization extend to the optimization of typical objective functions. In particular, we see that a typical optimal state is surprisingly close to the onset of cellular growth (estimated and shown as dashed vertical lines in Figure 4). Despite the closeness, however, the organism maintains high versatility, which we define as the number of distinct functions that can be optimized under growth conditions. To demonstrate this, consider the H. pylori model, which has 392 reactions that can be active, among which at least 302 must be active to sustain growth (Table 3). While only a few more than 302 active reactions are sufficient to optimize any objective function, the number of combinations for choosing them can be quite large. For example, there are combinations for choosing, say, 5 extra reactions to be active. Moreover, this number increases quickly with the network size: S. cerevisiae, for example, has less than 2.5 times more reactions than H. pylori, but about 500 times more combinations ().
Our results help explain previous experimental observations. Analyzing the 22 intracellular fluxes determined by Schmidt et al. [50] for the central metabolism of E. coli in both aerobic and anaerobic conditions, we find that about 45% of the fluxes are smaller than 10% of the glucose uptake rate (Table 4). However, less than 19% of the reversible fluxes and more than 60% of the irreversible fluxes are found to be in this group (Fisher exact test, one-sided p<0.008). For the 44 fluxes in the S. cerevisiae metabolism experimentally measured by Daran-Lapujade et al. [51], less than 8% of the reversible fluxes and more than 42% of the irreversible fluxes are zero (Table 5; Fisher exact test, one-sided p<10−7). This higher probability for reduced fluxes in irreversible reactions is consistent with our theory and simulation results (Table 6) combined with the assumption that the system operates close to an optimal state. For the E. coli data, this assumption is justified by the work of Burgard & Maranas [44], where a framework for inferring metabolic objective functions was used to show that the organisms are mainly (but not completely) driven by the maximization of biomass production. The S. cerevisiae data was also found to be consistent with the fluxes computed under the assumption of maximum growth [52].
Additional evidence for our results is derived from the inspection of 18 intracellular fluxes experimentally determined by Emmerling et al. [53] for both wild-type E. coli and a gene-deficient strain not exposed to adaptive evolution. It has been shown [21] that while the wild-type fluxes can be approximately described by the optimization of biomass production, the genetically perturbed strain operates sub-optimally. We consider the fluxes that are more than 10% (of the glucose uptake rate) larger in the gene-deficient mutant than in the wild-type strain. This group comprises less than 27% of the reversible fluxes but more than 45% of the irreversible fluxes (Table 7; Fisher exact test, one-sided p<0.12). This correlation indicates that irreversible fluxes tend to be larger in non-optimal metabolic states, consistently with our predictions.
Altogether, our results offer an explanation for the temporary activation of latent pathways observed in adaptive evolution experiments after environmental [16] or genetic perturbations [17]. These initially inactive pathways are transiently activated after a perturbation, but subsequently inactivated again after adaptive evolution. We hypothesize that transient suboptimal states are the leading cause of latent pathway activation. Since we predict that a large number of reactions are inactive in optimal states but active in typical non-optimal states, many reactions are expected to show temporary activation if we assume that the state following the perturbation is suboptimal and both the pre-perturbation and post-adaptation states are near optimality. To demonstrate this computationally for the E. coli model, we consider the idealized scenario where the perturbation to the growth-maximizing wild type is caused by a reaction knockout and the initial response of the metabolic network—before the perturbed strain evolves to a new growth-maximizing state—is well approximated by the method of minimization of metabolic adjustment (MOMA) [21]. MOMA assumes that the perturbed organisms minimize the square-sum deviation of its flux distribution from the wild-type distribution (under the constraints imposed by the perturbation).
Figure 5A shows the distribution of the number of active reactions for single-reaction knockouts that alter the flux distribution but allow positive MOMA-predicted growth. While the distribution is spread around 400–500 for the suboptimal states in the initial response, it is sharply peaked around 300 for the optimal states at the endpoint of the evolution, which is consistent with our results on random sampling of the extreme points (Figure 4). We thus predict that the initial number of active reactions for the unperturbed wild-type strain (which is 297, as shown by a dashed vertical line in Figure 5A) typically increases to more than 400 following the perturbation, and then decays back to a number close to 300 after adaptive evolution in the given environment, as illustrated schematically in Figure 5B. A neat implication of our analysis is that the active reactions in the early post-perturbation state includes much more than just a superposition of the reactions that are active in the pre- and post-perturbation optimal states, thus explaining the pronounced burst in gene expression changes observed to accompany media changes and gene knockouts [16],[17]. For example, for E. coli in glucose minimal medium, temporary activation is predicted for the Entner-Doudoroff pathway after pgi knockout and for the glyoxylate bypass after tpi knockout, in agreement with recent flux measurements in adaptive evolution experiments [17].
Another potential application of our results is to explain previous experimental evidence that antagonistic pleiotropy is important in adaptive evolution [54], as they indicate that increasing fitness in a single environment requires inactivation of many reactions through regulation and mutation of associated genes, which is likely to cause a decrease of fitness in some other environments [15].
Combining computational and analytical means, we have uncovered the microscopic mechanisms giving rise to the phenomenon of spontaneous reaction silencing in single-cell organisms, in which optimization of a single metabolic objective, whether growth or any other, significantly reduces the number of active reactions to a number that appears to be quite insensitive to the size of the entire network. Two mechanisms have been identified for the large-scale metabolic inactivation: reaction irreversibility and cascade of inactivity. In particular, the reaction irreversibility inactivates a pathway when the objective function could be enhanced by hypothetically reversing the metabolic flow through that pathway. We have demonstrated that such pathways can be found among non-equivalent parallel pathways, transverse pathways connecting structures that lead to the synthesis of different biomass components, and pathways leading to metabolite excretion. Although the irreversibility and cascading mechanisms do not require explicit maximization of efficiency, massive reaction silencing is also expected for organisms optimizing biomass yield or other linear functions (of metabolic fluxes) normalized by uptake rates [18]. Furthermore, while we have focused on minimal media, we expect the effect to be even more pronounced in richer media. On one hand, a richer medium has fewer absent substrates, which increases the number of active reactions in non-optimal states. On the other hand, a richer medium allows the organism to utilize more efficient pathways that would not be available in a minimal medium, possibly further reducing the number of active reactions in optimal states.
Our study carries implications for both natural and engineered processes. In the rational design of microbial enhancement, for example, one seeks genetic modifications that can optimize the production of specific metabolic compounds, which is a special case of the optimization problem we consider here and akin to the problem of identifying optimal reaction activity [24],[25]. The identification of a reduced set of active reactions also provides a new approach for testing the existence of global metabolic objectives and their consistency with hypothesized objective functions [46]. Such an approach is complementary to current approaches based on coefficients of importance [44],[45] or putative objective reactions [47] and is expected to provide novel insights into goal-seeking dynamic concepts such as cybernetic modeling [55]. Our study may also help model compromises between competing goals, such as growth and robustness against environmental or genetic changes [56].
In particular, our results open a new avenue for addressing the origin of mutational robustness [57]–[62]. Single-gene deletion experiments on E. coli and S. cerevisiae have shown that only a small fraction of their genes are essential for growth under standard laboratory conditions [1],[5],[6]. The number of essential genes can be even smaller given that growth defect caused by a gene deletion may be synthetically rescued by compensatory gene deletions [25], an effect not accounted for in single-gene deletion experiments. Under fixed environmental conditions, large part of this mutational robustness arises from the reactions that are inactive under maximum growth, whose deletion is predicted to have no effect on the growth rate [52]. For example, for E. coli in glucose medium, we predict that 638 out of the 931 reactions can be removed simultaneously while retaining the maximum growth rate (Note 4), which is comparable to 686 computed in a minimal genome study in rich media [11]. But what is the origin of all these non-essential genes?
A recent study on S. cerevisiae has shown that the single deletion of almost any non-essential gene leads to a growth defect in at least one stress condition [15], providing substantive support for the long-standing hypothesis that mutational robustness is a byproduct of environmental robustness [61] (at least if we assume that the tested conditions are representative of the natural conditions under which the organisms evolved). An alternative explanation would be that in variable environments, which is a natural selective pressure likely to be more important than considered in standard laboratory experiments, the apparently dispensable pathways may play a significant role in suboptimal states induced by environmental changes. This alternative is based on the hypothesis that latent pathways provide intermediate states necessary to facilitate adaptation, therefore conferring competitive advantage even if the pathways are not active in any single fixed environmental condition. In light of our results, this hypothesis can be tested experimentally in medium-perturbation assays by measuring the change in growth or other phenotype caused by deleting the predicted latent pathways in advance to a medium change.
We conclude by calling attention to a limitation and strength of our results, which have been obtained using steady-state analysis. Such analysis avoids complications introduced by unknown regulatory and kinetic parameters, but admittedly does not account for constraints that could be introduced by the latter. Nevertheless, we have been able to draw robust conclusions about dynamical behaviors, such as the impact of perturbation and adaptive evolution on reaction activity. Our methodology scales well for genome-wide studies and may prove instrumental for the identification of specific extreme pathways [63],[64] or elementary modes [65],[66] governing the optimization of metabolic objectives. Combined with recent studies on complex networks [67]–[73] and the concept of functional modularity [74], our results are likely to lead to new understanding of the interplay between network activity and biological function.
All the stoichiometric data for the in silico metabolic systems used in our study are available at http://gcrg.ucsd.edu/In_Silico_Organisms. For H. pylori 26695 [77], we used a medium with unlimited amount of water and protons, and limited amount of oxygen (5 mmol/g DW-h), L-alanine, D-alanine, L-arginine, L-histidine, L-isoleucine, L-leucine, L-methionine, L-valine, glucose, Iron (II and III), phosphate, sulfate, pimelate, and thiamine (20 mmol/g DW-h). For S. aureus N315 [78], we used a medium with limited amount of glucose, L-arginine, cytosine, and nicotinate (100 mmol/g DW-h), and unlimited amount of iron (II), protons, water, oxygen, phospate, sulfate, and thiamin. For E. coli K12 MG1655 [75], we used a medium with limited amount of glucose (10 mmol/g DW-h) and oxygen (20 mmol/g DW-h), and unlimited amount of carbon dioxide, iron (II), protons, water, potassium, sodium, ammonia, phospate, and sulfate. For S. cerevisiae S288C [76], we used a medium with limited amount of glucose (10 mmol/g DW-h), oxygen (20 mmol/g DW-h), and ammonia (100 mmol/g DW-h), and unlimited amount of water, protons, phosphate, carbon dioxide, potassium, and sulfate. The flux through the ATP maintenance reaction was set to 7.6 mmol/g DW-h for E. coli and 1 mmol/g DW-h for S. aureus and S. cerevisiae.
Under steady-state conditions, a cellular metabolic state is represented by a solution of a homogeneous linear equation describing the mass balance condition,(6)where S is the m×N stoichiometric matrix and is the vector of metabolic fluxes. The components of v = (v1,…,vN)T include the fluxes of n internal and transport reactions as well as nex exchange fluxes, which model the transport of metabolites across the system boundary. Constraints of the form vi≤βi imposed on the exchange fluxes are used to define the maximum uptake rates of substrates in the medium. Additional constraints of the form vi≥0 arise for the reactions that are irreversible. Assuming that the cell's operation is mainly limited by the availability of substrates in the medium, we impose no other constraints on the internal reaction fluxes, except for the ATP maintenance flux for S. aureus, E. coli, and S. cerevisiae (see Strains and media section above). The set of all flux vectors v satisfying the above constraints defines the feasible solution space , representing the capability of the metabolic network as a system.
The flux balance analysis (FBA) [18]–[20],[22],[23] used in this study is based on the maximization of a metabolic objective function cTv within the feasible solution space M, which is formulated as a linear programming problem:(7)We set αi = −∞ if vi is unbounded below and βi = ∞ if vi is unbounded above. For a given objective function, we numerically determine an optimal flux distribution for this problem using an implementation of the simplex method [43]. In the particular case of growth maximization, the objective vector c is taken to be parallel to the biomass flux, which is modeled as an effective reaction that converts metabolites into biomass.
To find a set of reactions from which none can be removed without forcing zero growth, we start with the set of all reactions and recursively reduce it until no further reduction is possible. At each recursive step, we first compute how much the maximum growth rate would be reduced if each reaction were removed from the set individually. Then, we choose a reaction that causes the least change in the maximum growth rate, and remove it from the set. We repeat this step until the maximum growth rate becomes zero. The set of reactions we have just before we remove the last reaction is a desired minimal reaction set. Note that, since the algorithm is not exhaustive, the number of reactions in this set is an upper bound and approximation for the minimum number of reactions required to sustain positive growth.
|
10.1371/journal.pntd.0001929 | Clinical Manifestations of Human Brucellosis: A Systematic Review and Meta-Analysis | The objectives of this systematic review, commissioned by WHO, were to assess the frequency and severity of clinical manifestations of human brucellosis, in view of specifying a disability weight for a DALY calculation.
Thirty three databases were searched, with 2,385 articles published between January 1990–June 2010 identified as relating to human brucellosis. Fifty-seven studies were of sufficient quality for data extraction. Pooled proportions of cases with specific clinical manifestations were stratified by age category and sex and analysed using generalized linear mixed models. Data relating to duration of illness and risk factors were also extracted. Severe complications of brucellosis infection were not rare, with 1 case of endocarditis and 4 neurological cases per 100 patients. One in 10 men suffered from epididymo-orchitis. Debilitating conditions such as arthralgia, myalgia and back pain affected around half of the patients (65%, 47% and 45%, respectively). Given that 78% patients had fever, brucellosis poses a diagnostic challenge in malaria-endemic areas. Significant delays in appropriate diagnosis and treatment were the result of health service inadequacies and socioeconomic factors. Based on disability weights from the 2004 Global Burden of Disease Study, a disability weight of 0.150 is proposed as the first informed estimate for chronic, localised brucellosis and 0.190 for acute brucellosis.
This systematic review adds to the understanding of the global burden of brucellosis, one of the most common zoonoses worldwide. The severe, debilitating, and chronic impact of brucellosis is highlighted. Well designed epidemiological studies from regions lacking in data would allow a more complete understanding of the clinical manifestations of disease and exposure risks, and provide further evidence for policy-makers. As this is the first informed estimate of a disability weight for brucellosis, there is a need for further debate amongst brucellosis experts and a consensus to be reached.
| Brucellosis is a bacterial disease transmitted to humans by consumption of infected, unpasteurised animal milk or through direct contact with infected animals, particularly aborted foetuses. The livestock production losses resulting from these abortions have a major economic impact on individuals and communities. Infected people often suffer from a chronic, debilitating illness. This systematic review on the symptoms of human brucellosis is the first ever conducted. Using strict exclusion criteria, 57 scientific articles published between January 1990–June 2010 which included high quality data were identified. Severe complications of brucellosis infection were not rare, with 1 case of endocarditis and 4 neurological cases per 100 patients. One in 10 men suffered from testicular infection, which can case sterility. Debilitating conditions such as joint, muscle, and back pain affected around half of the patients. Given that most patients had fever, brucellosis poses a diagnostic challenge in malaria-endemic areas where fever is often assumed to be malaria. More high quality data is needed for a more complete understanding of the clinical manifestations of disease and exposure risks, and to provide further evidence for policy-makers.
| Brucellosis is one of the most common zoonotic infections globally [1]. This bacterial disease causes not only a severely debilitating and disabling illness, but it also has major economic ramifications due to time lost by patients from normal daily activities [2] and losses in animal production [3]. In a review of 76 diseases and syndromes of animals, brucellosis lies within the top ten in terms of impact on impoverished people [4]. A brucellosis disability weighting of 0.2 has been previously proposed for Disability-Adjusted Life Years (DALY) calculation, based on the pain and impaired productivity known to result from infection [3]. However, a more informed estimate is needed for an accurate assessment of disease burden.
In 1992, the World Bank commissioned the original Global Burden of Disease (GBD) study, providing a comprehensive assessment of 107 diseases and injuries and 10 risk factors in eight major regions [5]. This review did not include any neglected tropical zoonoses. Such diseases often do not attract the interest of health researchers or sufficient resources for adequate control, yet they continue to impact significantly on human health and wellbeing, livestock productivity, and local and national economies [6]. There is a need for more accurate data relating to the burden of neglected zoonoses to facilitate more effective implementation of disease control interventions. In 2009, the Foodborne Disease Burden Epidemiology Reference Group (FERG) of the World Health Organization (WHO) commissioned a series of systematic reviews on the burden of neglected zoonotic diseases, with the aim of incorporating the findings into the overall global burden of disease assessments.
This report presents a systematic review of scientific literature published between 1990–June 2010 relating to morbidity from human brucellosis infection. The objectives of this review were to assess the frequency and severity of the clinical manifestations of brucellosis, the duration of disease, the associated disabilities and important risk factors, with a view to estimating an appropriate disability weight for calculation of the brucellosis DALY. A systematic review of scientific literature investigating the incidence and prevalence of brucellosis globally is the subject of a companion paper [7].
Thirty three databases were searched for relevant articles using the search terms of (brucellosis OR malta fever OR brucella melitensis OR brucella abortus) AND (symptom* OR sequelae* OR morbidity OR mortality OR transmission mode OR foodborne), with a publication limitation of 1990–30 June, 2010. The search term was adapted to the predominate language of the database. If a database did not allow the combining of Boolean operators, (18 of 33 databases), ‘brucellosis’ was used as the sole term.
Reference Manager bibliographic software was used to manage citations. Duplicate entries were identified by considering the author, the year of publication, the title of the article, and the volume, issue and page numbers of the source. In questionable cases, the abstract texts were compared.
The articles were sorted by a team of four reviewers with a combined fluency in English, German, French, and Spanish. Articles in other languages were noted for future translation, pending resources.
All reports were classified into one of two categories, based on their abstracts:
Category 1: Relevant – articles related to human brucellosis related to brucellosis infection in populations (i.e. disease frequency) or cases of human brucellosis (i.e. disease morbidity);
Category 2: Irrelevant - articles related to non-human brucellosis; articles addressing topics not related to the current review, such as genetics, laboratory diagnostic tests, experimental laboratory animal studies.
The abstracts of studies belonging to Category 1 and meeting the following criteria for disease morbidity studies were retained: published between 1990 and 30 June 2010, at least 10 study subjects, clinical symptoms/syndromes described, and some information relating to diagnostic tests provided. Articles relating to disease frequency and meeting the following criteria were also retained: published between 1990 and 30 June 2010, at least 100 study subjects drawn from the general population, prevalence or incidence data included, and some information relating to diagnostic tests provided. The assessment and classification of frequency articles will be the subject of a companion paper and will not be considered further here.
Articles for which the necessary data for classification could not be obtained were identified for possible future assessment, according to availability of resources. In general, non peer-reviewed or review articles, conference proceedings and book chapters were excluded.
After applying the aforementioned screening steps, the full text of each selected article was retrieved for detailed analysis. Each article was reviewed by two or three reviewers, and classification discrepancies were resolved by discussion.
Using a pre-designed Access database, articles were coded according to the following parameters:
1) Study type
Studies were classified as a prospective case series, a retrospective case series, a case-control study, or of another type.
2) Study population
The populations studied were grouped according to age category – children only (<15 years), adults only (≥15 years), or including both children and adults. Additionally, they were coded according to whether the study population represented the general population of brucellosis cases in the age category, or only a specific sub-group.
3) Diagnostic methods
Studies were classified according to their use of microbial culture to diagnose brucellosis patients. In order for studies to be included in the review, they had to not only mention culture in their methods but to also present laboratory results.
4) Overall study quality
Studies were given an overall quality grade of 1, 2, or 3. Quality 1 studies provided data drawn from general brucellosis cases, of which 75% or more were diagnosed by culture, and had well described study design and methods. Quality 2 studies also presented data from general brucellosis cases, utilised culture as a method and presented relevant laboratory results. However, unlike for Quality 1 studies, the majority of cases did not have to be diagnosed by positive culture in order to be included as Quality 2. Quality 3 studies were either drawn from only a specific sub-group of brucellosis cases such that general conclusions could not be drawn, did not use culture as a diagnostic method or failed to present culture results, or had poorly described study design and methods such that the quality of the data could not be assured.
Based on brucellosis literature [8] a comprehensive list of clinical manifestations associated with brucellosis cases was developed:
Numbers of subjects with each symptom/syndrome were recorded for each study, as well as the number of male and female patients. For the sex-related outcomes of epididymo-orchitis and abortion, the study population was considered to be only the male and pregnant female sub-groups of the study population respectively. Information relating to duration of disease prior to treatment and exposure to potential risk factors were also recorded wherever provided.
To calculate the proportion of patients by sex, numbers of male and female patients were aggregated across all studies as well as within each age category. 95% confidence intervals were calculated using the normal approximation to the binomial.
Where appropriate data were available from two or more studies, pooled proportions of patients with each clinical manifestation were estimated using generalized linear mixed models. Pooled estimates with 95% confidence intervals were calculated both within age categories and overall across all studies, using a Freeman-Tukey double arscine transformation. Homogeneity across studies was assessed using a Cochrane's Q test and total variability due to between-study variation was reflected in the I2 index. The meta-analysis was performed with R statistical software [9] using the meta package [10]. Additionally, in order to assess the impact of study design, the same analysis was conducted according to study type category.
The pooled estimates for proportions of patients with each clinical manifestation were compared with the disability weights used in the GBD 2004 study [11]. A disability weight for brucellosis was then proposed.
Median proportions of patients with exposure to particular risk factors were calculated. Data relating to duration of illness and diagnostic delay were recorded. In order to assess the duration of untreated illness, an additional, non-systematic search for data prior to the availability of appropriate antibiotics was undertaken by manually searching library records.
Table 1 lists the databases searched and the number of hits obtained for each. A total of 28,824 studies were identified, of which 59% were duplicates, leaving 11,000 original reports.
Figure 1 shows a flow diagram of the process for the selection of articles included in the review. In total, 289 frequency and morbidity studies were selected, for which full text was available for 153. However, 14 of these were in languages in which the team was not competent (Croatian (6), Turkish (4), Korean (2), Persian (1), Mandarin (1)), leaving 96 morbidity studies for quality assessment. Some articles contained both frequency and morbidity data and were thus counted in both categories.
Of the 96 morbidity studies for quality assessment, five were classified as Quality 1 and 52 as Quality 2. Thirty-nine were excluded from further analysis as Quality 3, one of which was due to duplication of data from another larger study. Two pairs of Quality 2 studies were based on the same data [12]–[15]. These studies were included because each provided some unique information; however, the duplicated data were only included once in the meta-analysis. Except for two articles in Spanish and one in French, all Quality 1 and 2 studies were in English.
The median number of study subjects was 143 (IQR: 85-283), ranging from 20-1028. Studies from high income countries such as Germany, France, and USA were generally situated at the lower end of the range (less than 60 subjects), although larger studies were reported from Spain, including one study of over 900 subjects. Of the 57 studies selected, 24 were from Turkey. The next most represented country was Saudi Arabia, with 8 studies, followed by Spain with 4 and Greece with 4. One or two studies each came from Cuba, France, Germany, Israel, India, Iran, Jordan, Kuwait, Tunisia, USA, Uzbekistan and Yemen. The geographic distribution of the selected studies is shown in Figure 2.
In terms of study type, 37 were classified as retrospective case series with data retrieved from medical records, and 19 as prospective case series. One study was a case-control. Seventeen studies provided detailed information about cases with specific syndromes, e.g. neurological brucellosis [16]–[19], epididymoorchitis [20]–[23], osteoarticular complications [13], [14], [24], [25], spondylitis [26], [27], pulmonary brucellosis [28], pancytopaenia [29], and pregnant women [30]. As these studies also provided some information about proportions of general brucellosis cases with specific symptoms/syndromes, they were included in the review.
Twenty-three studies included both children and adult participants [12]–[15], [18], [20], [24], [30]–[44]. Twelve studies investigated only children [29], [45]–[55], with an upper age limit ranging from 13 years to 18 years. Of the 19 studies with an adult population of 15 years or older [16], [17], [21]–[23], [25]–[27], [56]–[67], five consisted of only male participants [21]–[23], [64], [65]. Three studies did not clearly state the age category [19], [28], [68] and were analysed as if containing data for both adults and children.
In studies consisting of only children, 64% patients (95% CI: 60–68%) were male. The proportion of male patients in adult studies was significantly lower, at 56% (95% CI: 55–58%). In studies including both children and adult patients, 48% were male (95% CI: 46–51%). Overall, 55% patients (95% CI: 54–56%) across all studies were male.
Table 2 shows the pooled proportions of patients estimated by the random-effects model, according to clinical manifestations by age category. Forest plots are provided as Supplementary Information. An analysis by study type did not show any significant changes or trends.
Documented fever was common, with an estimated 78% of patients affected across the three age categories. Estimates of the proportions of patients with self-reported symptoms of sweats, chills, fatigue, headache, and malaise, were significantly lower in children, ranging from 9–24% depending on symptom, compared to 33–81% for adults. Weight loss in children, at 13%, was also lower than the 31% reported in adults.
Abdominal-related manifestations of pain, splenomegaly and hepatomegaly were fairly uniformly distributed across age categories, with overall estimated proportions of 19%, 26% and 23%, respectively. The number of studies reporting the presence of hepatitis was small, totalling only seven, with an estimated 4% patients affected overall.
Arthralgia was common, affecting 65% patients overall, whereas arthritis affected only 26% patients. In adult patients, 56% and 49% suffered from myalgia and back pain, respectively. Only two studies reported myalgia and back pain in children. Overall, spondylitis and sacroiliitis were detected in 12–36% adults.
In relation to reproductive problems, only one study reported abortion rates as a proportion of pregnant female participants, which was 46% [30]. Overall, 10% male patients had epididymo-orchitis.
For more severe outcomes, endocarditis was reported in an overall 1% patients, and neurological manifestations in 4%. Neurological outcomes reported included motor deficits, cranial nerve deficits, sciatica, confusion and/or psychological disturbances, meningitis and seizures. 6% of patients suffered from respiratory manifestations, including cough, bronchopneumonia, pleural adhesion and pleural adhesion. Cutaneous changes were reported in 6% patients.
As most studies were case series without a control group, an evaluation of the importance of risk factors was not possible. However, median proportions were calculated from 27 studies which provided some exposure history. Median proportions of brucellosis cases with exposure to a potential risk factor were 64% (IQR: 34–78%) for consumption of unpasteurised dairy products, 42% (IQR: 23–59%) for contact with livestock, and 6% (IQR: 3–19%) for occupational exposure, including veterinarians, butchers, and abattoir workers. From fifteen studies, the median proportion of cases with a history of brucellosis in a family member was 20% (IQR: 17–46%).
Only six studies included in the systematic review provided data regarding duration of illness prior to diagnosis and treatment [32], [41], [52], [55], [57], [62]. The age of the patient and the nature of the illness were influential factors. One study reported a longer duration of illness in adults compared to children under 15 years, averaging 8 weeks versus 4 weeks, respectively [41]. In another study, the average duration of illness prior to diagnosis and treatment was 40 days, but cases with osteoarticular disease generally experienced longer periods of illness, extending to 6 months [62].
The GBD 2004 study estimated the disability weights for low back pain due to chronic intervertebral disc disease and osteoarthritis of the knee to be 0.121 (range 0.103–0.125) and 0.129 (range 0.118–0.147), respectively [11]. Given the high proportion of patients in our systematic review with joint, back, or muscular pain, a disability weight of at least 0.150 is proposed as a minimum estimate for localised, chronic brucellosis. Generalised, non-specific clinical manifestations were also common. Acute, non-localised brucellosis could be approximated by an episode of malaria, estimated to be 0.191 (range 0.172–0.211) by the GBD 2004 study [11].
The clinical picture of brucellosis presented in this systematic review is consistent with other literature [69]. Although a large amount of data are available regarding clinical manifestations of brucellosis, its geographical distribution is limited. No high quality studies were identified from Sub-Saharan Africa, Central and South America or South-East Asia. This could potentially reflect either a lower disease burden or a poorer brucellosis surveillance system.
The proportion of male patients was greater than female patients amongst both children and adults. Although this difference was only small in adults, it was more pronounced in children. Possible explanations could be a greater risk of exposure amongst boys, with household responsibilities such as shepherding of livestock being preferentially delegated to boys, or gender-related differences in accessing to health care.
Given the high proportion of brucellosis cases with fever, brucellosis should be considered as a differential diagnosis for fevers of unknown origin. In malaria-endemic countries, fever patients are often diagnosed and treated for malaria based solely on clinical findings [70]. Improved diagnostic capacity would reduce the diagnostic delay and facilitate prompt and appropriate treatment. These health service inadequacies are compounded by socioeconomic factors, with brucellosis affecting poor, marginalised communities who often do not have the means to seek treatment. Although studies included in this systematic review did not investigate health-seeking behaviour, a study from rural Tanzania revealed that 1 in 5 patients did not present to a health centre for assessment until more than one year after the onset of illness. Once at the health centre, nearly half (45%) were not diagnosed with brucellosis at their first visit [71]. In children, particularly, under-diagnosis of brucellosis is likely. The lower proportions of reported general symptoms such as sweats, chills, fatigue, and headache in study populations consisting only of children in this systematic review could reflect difficulty in obtaining accurate case histories from this group.
One in 10 men experienced epididymo-orchitis, the most common genitourinary complication of brucellosis infection. This can have serious repercussions such as abscessation and infertility. Although other severe outcomes were less common, 4 neurological cases and 1 endocarditis case per 100 brucellosis patients were reported, which is substantial.
Arthralgia, myalgia, and back pain were common manifestations. The relative lower proportions of patients with sacroiliitis and spondylitis compared to those reporting back pain might reflect limitations in diagnostic capacity. Chronic pain has been shown to severely affect the quality of sufferers' social and working lives [72]. As the majority of the brucellosis disease burden is in less developed countries, where livelihoods are often reliant on physical activities, the impact of musculoskeletal pain and impaired function in these settings may be even more serious.
One study reported that patients with osteoarticular disease experienced a greater diagnostic delay than other cases [62], reflecting the chronic debilitation that can result from brucellosis infection. Indeed, in an endemic area of Russia prior to the availability of effective antibiotic therapies approximately 40% of 1,000 brucellosis cases followed over a 20 year period continued to suffer from clinical manifestations two years after disease onset. In this study, cited by Wund in 1966, approximately 90% of cases had self-cured after 6 years. [73].
Given the complexity of the clinical manifestations of brucellosis, summarising its impact into a single disability weight risks being too reductionist. However, a disability weight is required for an assessment of the global burden of disease which is, in turn, essential for engagement of policy-makers and funding bodies. Using the disability classes formerly used by the GBD 2004 study [74], a disability weight of 0.2 has been previously proposed based on Mongolian patient data [3]. This estimate fell between Class 1 (0.096), which referred to a limited ability to perform at least one activity in the one of the following areas: recreation, education, procreation or occupation; and Class 2 (0.22), referring to a limited ability to perform most activities in one of the aforementioned areas.
Based on this systematic review and meta-analysis, better informed estimates of disability weights are proposed: at least 0.150 for chronic, localised brucellosis and 0.190 for acute brucellosis. However, as this is the first informed estimate of a brucellosis disability weight, there is a need for further debate amongst brucellosis experts and a consensus to be reached.
Morbidity could vary geographically according to epidemiological setting. Well designed epidemiological studies from regions under-represented in this review would greatly contribute to an overall assessment of the global disease burden. A surveillance system amongst fever patients in malaria-endemic countries could be particularly informative. Additionally, risk factors for disease should be investigated through case-control studies. This would provide invaluable information to guide disease control interventions and policy.
Studies for which a title or abstract was not published in a language using the Latin alphabet, such as those published only in Chinese characters or Arabic script, may not have been identified during the original database search. Of the foreign language studies that were identified, those published in languages in which the team was not competent were excluded from the analysis. It is possible that some of these studies contained data that could have contributed to this global assessment of brucellosis morbidity. Additionally, although studies in English were independently reviewed by three team members, this was not always possible for studies reviewed in other languages (German, French, Spanish).
There were likely some differences between the case definitions and diagnostic capacity of different studies. For neurological and respiratory syndromes, many studies provided only an overall aggregated estimate without details of the different disease forms. A respiratory case could potentially vary from a patient with only a cough to severe bronchopneumonia, or a neurological case from altered behaviour and confusion to nerve deficits, meningitis or seizures. All patients were positive by culture in only 3 studies. Given the complexity of brucellosis serology interpretation, it is possible that some patients in other studies were misdiagnosed as cases of active brucellosis.
The studies provide data from brucellosis patients presenting to health centres. It is possible that cases that do not present to health centres are less severe. The results of this review may, therefore, be biased towards more severe cases. As with the estimation of other disability weights, the proposed brucellosis disability weight estimate assumes that a given clinical manifestation will result in the same disability in all settings, which is unlikely [75].
This systematic review adds to the understanding of the global burden of brucellosis, one of the most common and important zoonotic diseases worldwide. Brucellosis is shown to have a severe, debilitating, and often chronic impact on its sufferers. Significant delays in appropriate diagnosis and treatment are the result of both health system inadequacies and socioeconomic factors. Well designed epidemiological studies from those regions identified to be lacking in data would allow a better understanding of the clinical manifestations of disease and exposure risks and provide further evidence for policy-makers. Based on the findings of this systematic review and the disability weights from the 2004 Global Burden of Disease Study, a disability weight of 0.150 is proposed as the first informed estimate for chronic, localised brucellosis and 0.190 for acute brucellosis. As this is the first informed estimate of a disability weight for brucellosis, there is a need for further debate amongst brucellosis experts and a consensus to be reached.
|
10.1371/journal.pcbi.1003927 | Estimating Location without External Cues | The ability to determine one's location is fundamental to spatial navigation. Here, it is shown that localization is theoretically possible without the use of external cues, and without knowledge of initial position or orientation. With only error-prone self-motion estimates as input, a fully disoriented agent can, in principle, determine its location in familiar spaces with 1-fold rotational symmetry. Surprisingly, localization does not require the sensing of any external cue, including the boundary. The combination of self-motion estimates and an internal map of the arena provide enough information for localization. This stands in conflict with the supposition that 2D arenas are analogous to open fields. Using a rodent error model, it is shown that the localization performance which can be achieved is enough to initiate and maintain stable firing patterns like those of grid cells, starting from full disorientation. Successful localization was achieved when the rotational asymmetry was due to the external boundary, an interior barrier or a void space within an arena. Optimal localization performance was found to depend on arena shape, arena size, local and global rotational asymmetry, and the structure of the path taken during localization. Since allothetic cues including visual and boundary contact cues were not present, localization necessarily relied on the fusion of idiothetic self-motion cues and memory of the boundary. Implications for spatial navigation mechanisms are discussed, including possible relationships with place field overdispersion and hippocampal reverse replay. Based on these results, experiments are suggested to identify if and where information fusion occurs in the mammalian spatial memory system.
| Spatial navigation is one of the most important functions of animal brains. Multiple regions and cell types encode the current location in mammalian brains, but the underlying interactions between sensory and memory information remain unclear. Recent experimental and theoretical evidence have been found to suggest that the presence of a boundary fundamentally alters the task of navigation. In this paper, evidence is provided that it is possible to determine the location inside any familiar arena with 1-fold rotational symmetry, while completely ignoring sensory cues from the outside world. Surprisingly, the results show that the mere knowledge of the boundary's existence is enough, without requiring direct physical contact. Localization is robust despite the presence of noise modelled from the rodent head direction system, and even inaccuracies in the navigation system's memory of the boundary or internal models of noise. In circular arenas, rotational asymmetry can arise from interior structures such as barriers or voids, also without contact information. This theoretical evidence highlights the need to distinguish arena-based navigation common to most experimental studies, from open field navigation. These findings also point to novel ways to study information fusion in mammalian brains.
| Accurate spatial navigation is crucial to animal survival. Localization is the process of determining current location, critical for many navigation behaviours. Starting from an unknown location and direction (jointly called “pose”), the ability to localize is thought to depend on the detection of world-based (‘allothetic’) cues such as visual landmarks. In contrast, it is thought that animal-based (‘idiothetic’) cues which provide self-motion estimates, e.g., from vestibular, proprioceptive or motor command signals, can only serve to maintain localization briefly, requiring allothetic cues for error correction [1]–[5]. This is because cumulative errors will degrade self-motion estimates of position over time, not improve it.
In the rodent brain, the firing of both place and grid cells strongly correlate with the animal's physical location in a familiar space [6]–[10], with firing patterns being stable over days to weeks in the same environment [11]. Such neural correlates demonstrate that an animal can robustly localize itself within a familiar arena. A consistent feature of these neural correlates of localization is their persistence without vision, for upwards of 30 minutes [12], [13]. One possible strategy is to use idiothetic path integration (iPI) whereby an animal keeps track of its current position by summing idiothetic estimates of displacement [8], [14]–[16], but this process is known to suffer from cumulative errors. In an open field, error accumulation due to iPI will lead to a rapid increase in discrepancy between true position and estimated position [14]–[16]. It follows that, if using only iPI, a navigation system cannot accurately estimate its location in the long term, and certainly cannot localize itself starting from an unknown pose. An important biological implication is that stable, spatially-selective firing patterns such as those of rodent hippocampal place cells or medial entorhinal grid cells cannot depend purely on iPI.
Most experiments investigating neural correlates of spatial behaviour have been performed in either linear tracks or 2D arenas, the latter termed ‘open fields’ [9], [10]. On their own, featureless arena boundaries do not provide sufficient spatial information for localization without vision. This is due to a combination of geometric properties [17], and infinite poses which equally account for the detection of a point along a featureless boundary during boundary contact. The problem is compounded further if boundary contact (an allothetic cue) is not available or used. The mere knowledge of a boundary's geometry is therefore insufficient for localization, and might be interpreted as support that arena boundaries do not significantly aid localization compared to boundary-less open fields.
Contrary to the above supposition, it is demonstrated here that long-term accurate localization is possible if idiothetic self-motion cues are combined with boundary information already in memory. In particular, the previously acquired boundary map limits the growth of uncertainty due to noisy self-motion information. Surprisingly, the act of sensing the boundary or any other landmark is not necessary. Metrics based on information theoretic principles are used to quantify localization performance. Grid cell simulations are used to provide both a visual display of time-averaged localization performance, and to predict the optimal spatial selectivity that may be expected of neural firing patterns, based on published rodent neural data.
As a baseline, both iPI and aPI (allothetic PI implies PI using a compass) performance were quantified without arena memory, in a kite-shaped arena. Even for the simplified task of PI initially orientated, localization failed (Fig. 1A, 1B). The median place stability index, , fell below 0.5 (chance level) in 3 minutes using iPI alone, and under 6 minutes using aPI alone (Fig. 1C), consistent with the generally-accepted idea that cumulative PI errors degrade location estimates over time. Using iPI, the circular variance which measures the particle filter's directional performance across trials, increased from 0 (no error) to close to 1 (uniformly random orientation). In contrast, remained close to 0 using aPI since the compass continually reset orientation errors. Using either type of PI, spatially-selective firing patterns could not be maintained beyond 1–2 minutes (Fig. 1D). The more general task of localization initially disoriented (Fig. 1E) further increased localization difficulty, with throughout the simulation period, and no grid-like firing pattern was observed. Taken together, it was clear that neither aPI or iPI alone could enable localization. Next, idiothetic self-motion cues were combined with arena memory information.
Figure 2A shows snapshots of the positional uncertainty distribution along random trajectories, combining idiothetic self-motion cues with arena boundary memory. Starting either oriented or disoriented, the true pose remained close to the estimated pose. remained above 0.5 using idiothetic cues (Fig. 2B) demonstrating localization success. Similarly, the directional component of the pose estimate, θ, was centred on the true direction. Not surprisingly, initial orientation improved position estimation but its effect on Ip was no longer detectable at 96 minutes (Wilcoxon test, p = 0.49). Likewise, V(θ) remained consistently higher when initially disoriented (Fig. 1B, dotted lines), but the effect persisted beyond 192 minutes (κ-test, p = 6.0×10−4). Lastly, 90% of changes in occurred within the first 5 minutes. Together, these results show that idiothetic localization was achieved rapidly even when initially disoriented.
To determine whether the idiothetic localization described above could sustain spatially selective firing patterns similar to those of grid cells, a stochastic spiking model of grid cells was used [17] (SI Modelling and Analysis). Accurate localization was expected to result in multiple distinct, spatially regular activity peaks (modes).
Distinct grids were seen both when the animal was initially oriented (Fig. 2C, row 1) and initially disoriented (Fig. 2C, row 2). Autocorrelograms (Fig. 2C, right column) of the normalized firing fields showed spatial regularity similar to grid cells [13], [18], [19]. Of note was the rapid emergence of the grid pattern during the first 2 minutes (Fig. 2C, 0–2) when initially disoriented, consistent with the changes in Ip and V(θ) of Fig. 2B. These results show that it is plausible for neural correlates of successful idiothetic localization to be observed using arena size and timescales similar to rodent experiments.
A range of boundary properties were found to be compatible with idiothetic localization, including one axis of reflective symmetry (all arenas of Fig. 2C), arena concavity (T-maze arena), lack of vertices and straight edges (egg-shaped arena), or a circular outer arena boundary (void landmark arena). In all cases, the spatial information content and gridness indices (Table S1) demonstrated spatial specificity comparable to published rodent place and grid cell data [19], [20]. However, localization metrics including (Fig. 2D, left) and Ip distributions (Fig. 2D, right) varied with arena, showing that idiothetic localization performance depended on arena geometry. Noise level, additional allothetic boundary contact information and arena size also had graded effects on localization performance (Text S1 - Supporting Results, Fig. S1 and S2, Table S1).
The mechanism of localization from initial disorientation may be intuited by considering the mechanism of action of a particle filter. Among a large number of initially random pose hypotheses (represented by particles), some are close to the true pose while others are not (Fig. 3A, left). A poor initial pose estimate is more likely to result in an estimated trajectory which crosses a boundary in memory, compared to a good initial pose estimate (Fig. 3A, middle and right).
Over time, estimates cluster around the true pose, plus any rotationally symmetric poses. The latter occurs because in arenas with n-fold rotational symmetry (n-RS), there are n poses which are geometrically equivalent and consistent with the boundary map, making idiothetic localization to a unique pose impossible. Therefore the combination of boundary asymmetry and an internal model of boundary crossing sufficed for idiothetic localization.
Through careful inspection of the particle filter model, five further predictions related to idiothetic localization were made and demonstrated through simulations using the rodent HD error model. Firstly, local arena geometry was expected to affect the Ip value differently in arenas with the same degree of rotational symmetry, since within each of n sectors of an arena with n-RS, localization should be possible with performance depending on the sector's asymmetry. An adjusted Ip* for n-RS arenas was found to be positively correlated to the average rotational asymmetry (Fig. 3B, Text S1 - Modelling and Analysis), confirming that both local and global rotational asymmetry affected localization performance.
A second prediction was that an asymmetric interior barrier could replace rotational asymmetry in the traversable space, and was tested in a circular arena (Fig. 3C). Modelling boundary crossings, a circular traversable area (∞-RS) allowed idiothetic localization when an internal barrier was present. However, localization failed when estimated movements were represented as discrete steps which ignored barrier crossings (Text S1 - Modelling and Analysis), showing that the way that self-motion cues and arena memory information are combined can significantly affect performance. In this instance, being able to determine whether a boundary has been crossed was more important for the navigation system than merely determining whether it remained inside the arena. In the particle filter, the pose hypotheses (particles) which crossed any boundary were removed, even if they remained within the traversable space of the remembered arena. A related prediction was that intermittent use of a compass suffices for localization in empty circular arenas (Fig. S4), since a compass directly breaks rotational symmetry in any arena. In this way, allothetic compass information may be incorporated infrequently, while place information is maintained using idiothetic cues for most of the time. Importantly, no ‘reset’ of the position estimate was required – only breaking of rotational symmetry through the compass.
A third prediction was that the centre of a finite-sized navigating agent need not reach the arena boundary, if it used an internal model of its own perimeter. This was demonstrated using an elliptic agent using both random and thigmotactic (wall following) trajectories (Fig. 4A). Accuracy was significantly improved by using a thigmotactic movement strategy, demonstrating trajectory-dependence of localization performance.
A fourth prediction was that the test arena may vary slightly from the learned arena. Assuming a standard kite arena in memory, the test arena was linearly expanded in the X, Y or both X & Y directions, by 10% (Fig. 4B, S3C). Grid modes showed greater spatial specificity following X or Y expansion, than along both directions. Partial grid rescaling was observed [19], principally along the expansion direction. These results show that strict congruence between the learned and test arena was unnecessary, but that spatial specificity was affected by a disparity between the learned and actual boundary.
A final prediction was that the true pose at the beginning of a disoriented trial can be recovered by replaying self-motion estimates in reverse. In real time, the initial pose estimate was uniformly distributed over the arena in all directions. Following a period of localization, the final pose estimate was treated as the initial pose estimate of the same trajectory replayed in reverse, in an ‘offline’ manner. Fig. 5 shows that following reverse replay, pose estimates were substantially improved from real-time pose estimates during initial localization, which were optimal at the time. Assuming that a sequence of self-motion estimates can be stored and retrieved later, this simple strategy can significantly improve a past pose estimate retrospectively. Alternatively, an ‘online’ backward inference procedure can also be used to achieve retrospective localization for a chosen time, without storing self-motion estimates (Fig. S5, Text S1 – Modelling and Analysis, Text S1 - Supporting Results). The ability to accurately recover the starting pose implies that homing is possible using only idiothetic sensory cues, even when initially disoriented. In principle, direct homing can occur after an indefinite period of time, since both current pose and initial pose (‘home’) can be determined.
In terms of a particle filter, idiothetic localization can be seen as a consequence of a dynamic competition between increasing uncertainty due to iPI errors (Fig. 1A) and decreasing uncertainty due to culling of invalid hypotheses which cross a boundary (Fig. 3A). The rate of increase of uncertainty (diffusion-like particle cloud expansion) depends on the magnitude of intrinsic iPI errors (Fig. S1) and path structure (e.g., path tortuosity [16], arena size and shape). The rate of decrease of uncertainty (i.e., particle culling) depends on the interaction of arena shape (e.g., Fig. 2), size (e.g., Fig. S3) and path structure (e.g., Fig. 4). When the factors affecting the rate of increase and decrease in uncertainty are kept constant over a prolonged period of time, a dynamic equilibrium is reached which corresponds to the plateau seen in most functions.
The complex interactions between arena shape, size and path structure are briefly described. Intuitively, if relatively few pose hypotheses (particles) can account for a sequence of displacements, uncertainty tends to decrease. For example, a path taken from the acute to obtuse to a right angle corner of a kite arena cannot be accounted for by any particle trajectory except one which is close to the true trajectory. In contrast, there are infinite possible paths which span the diameter of a circular arena or follow its boundary, so there are always a range of particle trajectories which can account for any displacement sequence arising in a circular arena. Consequently, determining a unique location is impossible. In a given 1-RS arena, the path structure determines the relative uniqueness of the displacement sequences. Hence very few thigmotactic loops around a 1-RS arena (e.g., Fig. 4A) are required to uniquely determine location, within the limits of noise. Therefore, when rotational asymmetry is due to an outer boundary, thigmotaxis is an efficient strategy for localization. In contrast, a trajectory biased towards the center of such an arena takes longer for localization, and the agent spends relatively longer time periods poorly localized leading to a lower level of equilibrium performance. Thus large arenas tend to increase uncertainty since more time is spent in the centre of the arena. This effect is exacerbated by the nonlinear increase in positional uncertainty due to unconstrained iPI [16]. It is important to note that these effects are not applicable if the rotational asymmetry is due only to an interior arena structure such as a barrier or void. In this case, thigmotaxis is not sufficient for localization, while a trajectory bias towards the arena centre will improve localization compared to a random trajectory.
The magnitudes of the competing rates of increase and decrease in uncertainty also depend on the initial pose distribution (e.g., Fig. 2A, S2). For example, when initially disoriented, a large portion of pose hypotheses are grossly incompatible with self-motion cues given the known arena, resulting in a relatively high rate of decrease in uncertainty (particle culling) relative to the increase from iPI (diffusive expansion of particle cloud). Consequently, overall uncertainty decreases, leading to improved localization when initially disoriented. In contrast, when initially oriented, the majority of pose hypotheses remain compatible with self-motion information so the rate of particle culling is low relative to the diffusive expansion of the particle cloud due to iPI errors. Hence, there is an overall decay in localization performance when initially orientated. In general, the direction of change in localization performance (e.g., ) depends on whether the current pose uncertainty is above or below the equilibrium level. In turn, the equilibrium performance level depends on the interaction between iPI errors, arena size, shape and path structure.
By itself, rotational asymmetry is not sufficient for long term localization. Fig. 1B shows that even combining two levels of rotational symmetry breaking (compass+kite arena) is not sufficient for localization. The lack of an arena map led to unlimited growth of uncertainty in the position estimate, despite accurate orientation. In contrast, breaking rotational symmetry in combination with an arena map allowed long term localization (e.g., Fig. 2, Fig. S4). Hence it is the combination of rotational asymmetry and arena map which allows idiothetic localization.
However, an arena map is not always necessary if allothetic cues are available. When detected, the acute and obtuse corners of a kite arena are unique landmarks which can theoretically be used by a modular navigation system for localization [17], [21], without requiring a full boundary map in memory. The axis of the two unique landmarks provides the symmetry breaking information (hence orientation), while the point-like nature of each landmark provides location information. Similarly, rats and other animals can leave markings inside arenas which can potentially be used to break rotational symmetry and allow localization even in a circular arena, without a boundary map. Hence it is specifically the restriction to using only idiothetic cues which necessitates both the use of an arena map and presence of rotational asymmetry for successful localization.
It was shown previously that combining self-motion cues and arena memory significantly slowed the decay of pose estimates in a circular arena, when initially oriented. When allothetic boundary contact cues were included, the residual localization achieved could account for the prolonged stability of rodent place and grid fields observed in darkness despite an unstable head direction system [17]. It is now shown that in arenas with 1-fold rotational symmetry, localization decay can in fact be stopped altogether, that allothetic cues are not necessary, and the navigating agent can be initially disoriented. Important implications of these new findings are discussed below.
Using only idiothetic cues in conjunction with information already in memory potentially reduces computational load during navigation. Such a strategy allows a navigating agent to devote computational or attentional resources for processing allothetic sensory information for other tasks. Although precision is reduced, it is likely to be a low-risk strategy since occasional use of allothetic cues suffices to recover near-optimal localization (e.g., Fig. S2A, Fig. S4). There is evidence that animals may use allothetic sensory cues intermittently during navigation [22], [23]. If allothetic cues are used intermittently during localization by the hippocampal-entorhinal space circuit, two consequences may be observed. Firstly, cumulative spatial uncertainty may increase spatial firing variability beyond that expected from average firing rates. Over multiple passes through the same true location, positional uncertainty will cause variability in the estimated location, potentially reflected in variability in spike activity. This may contribute to the phenomenon of ‘overdispersion’ [24], [25] observed in CA1 place cells, whose firing fields are influenced by both allothetic cues [26], [27], and self-motion information such as via grid cells [8], [28]. A second possible consequence of cumulative spatial uncertainty is temporary divergence in the spatial code relative to allothetic landmarks. A temporary divergence may contribute to the multiple ensemble place codes which have been reported in rodent CA1, interpreted as alternating attention between distal landmark cues and self-motion cues [25]. However, the results of the present study suggest that using self-motion cues alone is likely to lead to degradation of the place code within 2–3 minutes (Fig. 1). Hence in the long term, attention to self-motion cues is not sufficient to account for a second stable place code, unless there is also intermittent compass information, rotational asymmetry in the arena, or both.
The rodent medial entorhinal cortex (mEC) contains both border cells and grid cells [13], [29], which raises the possibility that it could be a self-sufficient localization system. If mEC border cells encode a boundary spatial representation, then together with the putative grid cell based path integration system [8], [13], fusion of self-motion and boundary memory information should enable localization. If so, stable grid fields are expected to emerge in familiar arenas with 1-fold rotational symmetry, in the absence of vision, initially disoriented, and in a hippocampus-independent manner. However, in arenas with >1-fold rotational symmetry such as circular or square arenas, such a system would require supplementation with a compass cue, at least sporadically (e.g., Fig. S4). In rodents, the latter could be provided via the visually-stabilised head direction (HD) system [30]–[32], whose firing properties mature prior to place and grid cells during development [33], [34], as would be expected if HD cells provided important symmetry-breaking information for place and grid cells.
Boundaries occurring in natural environments rarely have >1-fold rotational symmetry, making it plausible that biological navigation systems may exploit this property for localization. It remains to be tested whether species which spend significant time in enclosed spaces [35]–[37] are more likely to have evolved mechanisms to use this localization strategy. It has been reported previously that the persistent stability of rodent place and grid cell firing in darkness starting from full orientation [12], [13] is likely to rely on the fusion of self-motion and boundary information [17]. Assuming the same type of information fusion occurs starting from full disorientation, the results of the current study suggest the emergence of stable place and grid fields should occur in arenas with 1-fold rotational symmetry.
It is worth noting that the localizing mechanism described here need not be restricted to bounded spaces with impassable physical barriers. For instance, this strategy is equally applicable if an animal's trajectory is limited by a few distinct landmarks, forming a virtual arena with 1-fold rotational symmetry in an otherwise open trajectory space. Familiar landmarks could thus be used to break rotational symmetry in the trajectory space without a physical barrier, and resulting in successful localization without knowledge of the distance or allocentric direction to those landmarks during a journey. However, the geometry of the virtual boundary must be known. In this hypothetical example, the trajectory is guided by allothetic cues, while localization uses idiothetic cues and information already in memory.
One prediction of idiothetic localization was that reverse replay of past information enabled retrospective improvement in localization (Fig. 5). That is, knowledge of current location improved when future information became available. Experimentally, replays of sequences of place cell activity corresponding to past behavioural trajectories have been reported during sleep [38], [39] and when awake [40], including in reverse temporal order [40], [41]. The modelling results here suggest that a possible role of hippocampal reverse replay may be to improve past estimates of location, which may in turn improve the accuracy of future path planning [42].
While the present study examined the information and computations which may be necessary and sufficient to be used in conjunction with a known boundary map for localization, future studies will need to address the acquisition of the boundary map itself. In robotics, particle filter methods have been used successfully to build boundary maps using only self-motion and boundary detection cues, starting from full disorientation – a Simultaneous Localization and Mapping (SLAM) problem [43], [44]. Similar methods may be used to investigate the factors which could affect the acquisition of a boundary map under biologically realistic conditions, and make predictions about localization performance when only imperfect maps are available.
From the findings in this study, it is proposed that spatial memory systems which can effectively combine idiothetic self-motion cues and boundary memory can determine location in familiar arenas with 1-fold rotational symmetry. If allothetic cues are stringently removed, localization necessarily demonstrates the fusion of idiothetic self-motion and memory-based boundary information. This prediction may be tested, for example, by using blindfolded human subjects passively led along random or thigmotactic trajectories. Where in vivo recordings are feasible, it may be possible to isolate the cells and circuits where the fusion of idiothetic self-motion and boundary memory information occurs. For instance, if stable mEC grid fields emerge from full disorientation under the cue-restricted conditions described, information fusion must occur either at or upstream of the medial entorhinal grid cells.
Together, the reported results reveal that detection of cues from the external world is not always necessary for localization, that bounded arenas are distinct from true open fields [9], [10], [45], [46], and that any information which breaks rotational asymmetry may be useful for localization. Furthermore, arena boundaries affect navigational difficulty in a size-, shape- and path- dependent manner, and need to be addressed during the design and interpretation of experiments which investigate the navigational abilities of animals in arenas. Finally, the results suggest that specific arena designs can be used to interrogate the combination of self-motion and memory information in the hippocampal-entorhinal space circuit, whose properties are influenced by environmental boundary information [18], [29], [47], [48].
It is known from recordings of Head Direction (HD) cells in rats, that error in the estimate of head direction increase steadily as the animal moves away from the place where it was first deprived of its vision [30], [31]. The accumulation of errors in head direction was modelled as a Wiener process [17], based on the reported drift in tuning functions of HD neurons without vision [30], [31]. Simulated rodents began each trial either oriented (i.e. with perfect initial pose information) or disoriented (no initial pose information).
A particle filter was used to approximate Bayes-optimal fusion of idiothetic self-motion and boundary information, to provide the best estimate of successive poses given noisy displacement inputs. The localisation performance, as revealed by this particle filter approach, was analysed across 103 random trials in each condition, using metrics developed previously to characterise instantaneous spatial accuracy and precision. Briefly, the median place stability index, , characterizes the particle filter's localization performance across trials, where 0.5 represents chance (uniform uncertainty within the arena), and 1 represents perfect localization. Similarly, the circular variance measures the particle filter's directional performance across trials, where 0 represents uniformly random heading estimates, and 1 represents no heading error.
The particle filter estimate of position was then used to simulate the firing activity of grid cells in medial entorhinal cortex, using a stochastic spiking model where spike probability decreased with estimated distance from the neurons' preferred firing locations [17]. The resultant spike patterns were analysed using standard time-averaged metrics developed to characterize place and grid cell activity [19], [20], including information content and gridness indices. See Text S1 - Modelling and Analysis for further details.
|
10.1371/journal.pgen.1007011 | Formation of a TBX20-CASZ1 protein complex is protective against dilated cardiomyopathy and critical for cardiac homeostasis | By the age of 40, one in five adults without symptoms of cardiovascular disease are at risk for developing congestive heart failure. Within this population, dilated cardiomyopathy (DCM) remains one of the leading causes of disease and death, with nearly half of cases genetically determined. Though genetic and high throughput sequencing-based approaches have identified sporadic and inherited mutations in a multitude of genes implicated in cardiomyopathy, how combinations of asymptomatic mutations lead to cardiac failure remains a mystery. Since a number of studies have implicated mutations of the transcription factor TBX20 in congenital heart diseases, we investigated the underlying mechanisms, using an unbiased systems-based screen to identify novel, cardiac-specific binding partners. We demonstrated that TBX20 physically and genetically interacts with the essential transcription factor CASZ1. This interaction is required for survival, as mice heterozygous for both Tbx20 and Casz1 die post-natally as a result of DCM. A Tbx20 mutation associated with human familial DCM sterically interferes with the TBX20-CASZ1 interaction and provides a physical basis for how this human mutation disrupts normal cardiac function. Finally, we employed quantitative proteomic analyses to define the molecular pathways mis-regulated upon disruption of this novel complex. Collectively, our proteomic, biochemical, genetic, and structural studies suggest that the physical interaction between TBX20 and CASZ1 is required for cardiac homeostasis, and further, that reduction or loss of this critical interaction leads to DCM. This work provides strong evidence that DCM can be inherited through a digenic mechanism.
| A molecular understanding of cardiomyocyte development is an essential goal for improving clinical approaches to CHD. While TBX20 is an essential transcription factor for heart development and its disease relevance is well established, many fundamental questions remain about the mechanism of TBX20 function. Principle among these is how TBX20 mutations associated with adult dilated cardiomyopathy circumvent (DCM) the essential embryonic requirement for TBX20 in heart development. Here we report using an integrated approach that TBX20 complexes with the cardiac transcription factor CASZ1 in vivo. We confirmed TBX20 and CASZ1 interact biochemically and genetically, and show mice heterozygous for both Tbx20 and Casz1 die, beginning at 4 to 8 weeks post birth, exhibiting hallmarks of DCM. Interestingly, the human mutant TBX20F256I bypasses the early essential requirement for TBX20 but leads to DCM. We report here that TBX20F256I disrupts the TBX20-CASZ1 interaction, ascribing clinical relevance to this protein complex. Further, by using quantitative proteomics we have identified the molecular pathways altered in TBX20-CASZ1-mediated DCM. Together, these results identify a novel interaction between TBX20 and CASZ1 that is essential for maintaining cardiac homeostasis and imply that DCM can be inherited through a digenic mechanism.
| Heart failure is a major cause of morbidity in the United States with more than 5 million people in the US living with this disease [1]. A major risk factor for developing heart failure is dilated cardiomyopathy (DCM). Clinically recognized as systolic dysfunction accompanied by dilation of one or both ventricles, DCM is a predominating cardiomyopathy and the most common disease requiring heart transplantation in the US [2, 3]; however, nearly half of DCM cases are of unknown etiology [4].
In efforts to understand the etiology of idiopathic DCM, mutations in over 50 genes including components of the contractile apparatus and cell cytoskeleton, as well as in factors involved in excitation-conduction coupling, have been identified as causative in DCM [5, 6]. However, few studies have explored the potential for aberrant transcriptional regulation of these factors to contribute to disease pathogenesis. In exception to this, recent studies have identified mutations in the T-box transcription factor TBX20 associated with DCM [7–9].
Results of genetic analysis and protein depletion studies are consistent with an essential role for TBX20 during the early stages of vertebrate heart development [10–17]. Hearts lacking Tbx20 show progressive loss of cardiomyocytes, failure of the heart to undergo looping and chamber formation, and defects in cardiomyocyte maturation [17–21]. In humans, loss-of-function mutations in TBX20 can cause dilated cardiomyopathy, atrial septal defects, or mitral valve disease, while gain-of-function mutations in TBX20 have been reported in patients with Tetralogy of Fallot (i.e., pulmonary outflow tract obstruction, ventricular septal defect, overriding aortic root and right ventricular hypertrophy) [7, 8, 22–24]. It has been further demonstrated that ablation of Tbx20 in adult mouse cardiomyocytes leads to the onset of severe cardiomyopathy leading to death within 1–2 weeks after Tbx20 loss [25].
While TBX20 is an essential transcription factor for heart development and its disease relevance is well established, many fundamental questions remain about the mechanism of TBX20 function. Principle among these is how TBX20 mutations associated with DCM circumvent the essential embryonic cardiac requirement for TBX20.
To elucidate the mechanisms by which mutations in TBX20 lead to human adult pathological states, we identified endogenous TBX20 cardiac protein-protein interactions by coupling a tagged endogenous allele of Tbx20 with unbiased proteomic analysis. Results from these studies revealed TBX20 interacts with the essential cardiac transcription factor Castor (CASZ1), a gene that was also recently linked to DCM [26]. We confirmed that TBX20 and CASZ1 interact biochemically and genetically, and we go on to show that while mice singularly haploinsufficient for Tbx20 or Casz1 are asymptomatic, mice heterozygous for both Tbx20 and Casz1 die, beginning at 4 to 8 weeks post birth, and exhibit cardiomyocyte hypertrophy, interstitial fibrosis, and severe DCM. Interestingly, the human mutant TBX20F256I bypasses the early essential requirement for TBX20 but leads to DCM. We report here that TBX20F256I disrupts the TBX20-CASZ1 interaction, ascribing clinical relevance to this protein complex. Further, by using quantitative proteomics we have identified the molecular pathways altered in TBX20-CASZ1-mediated DCM. Together, these results identify a novel interaction between TBX20 and CASZ1 that is essential for maintaining cardiac homeostasis. These findings imply that DCM can be inherited through a digenic mechanism.
To identify endogenous protein interactions that regulate TBX20 function, we introduced the Avitag, in-frame, to the carboxy terminus of mouse Tbx20 through homologous recombination in mouse embryonic stem cells (ESCs) (Tbx20Avi) (S1A and S1B Fig). Since the Avi-tag can be biotinylated through recognition of the Avi-tag sequence by the E. coli biotin ligase BirA [27, 28], we generated a lentivirus expressing BirA and transduced it into mouse Tbx20Avi/+ ESCs. After hygromycin selection, Tbx20Avi/+ ESCs that stably expressed BirA (S1C Fig) were differentiated into induced cardiomyocytes (iCM) using a serum-free differentiation method that routinely generates cultures containing >60% cardiomyocytes (Fig 1A) [29].
Expression analysis at each day of differentiation confirmed the Tbx20Avi:BirA ESCs recapitulated the wild-type cardiomyocyte differentiation program (Fig 1B). We further showed by Myosin Heavy Chain (MHC) expression and time lapse imaging that the Tbx20Avi:BirA ESCs differentiated into beating neonatal cardiomyocytes (Fig 1C, S1 Movie). Our analysis further verifies that Tbx20Avi expression recapitulates endogenous Tbx20 expression with highest levels in immature cardiomyocytes at day 4 and differentiated cardiomyocytes at day 7 (Fig 1B).
Published data has demonstrated a requirement for TBX20 in adult mice, with loss of TBX20 leading to abrupt cardiac failure [25]. Recently, mutations in TBX20 in humans were associated with DCM [7–9]. To delineate the mechanisms of how TBX20 DCM-associated mutations circumvent the essential requirements for TBX20 in cardiac development, we isolated and characterized the endogenous TBX20 cardiac protein interactome under physiological conditions from Tbx20Avi; BirA iCMs at day 7 of differentiation. As a control for non-specific interactions, identical affinity isolations were performed from BirA-negative iCMs. Proteins co-isolated from TBX20Avi affinity purifications (APs) were analyzed by an SDS-PAGE tandem mass spectrometry-based proteomics approach, as in [30]. TBX20 was detected in the AP from BirA-expressing iCMs, with 19 unique tryptic peptides covering 54% of the TBX20 sequence (out of a theoretical maximum coverage of ~75%) (Fig 1D and 1E; S2A and S2B Fig; S1 Table).
Identification of candidate high-confidence TBX20 interactions that have the potential to regulate cardiac functions was achieved using a multi-step bioinformatics approach based on the number of identified spectra per protein. First, interacting proteins identified by less than 10 spectra did not meet the identification requirement and were excluded from further analysis. Further, proteins identified in the BirA-expressing isolations were required to have at least a 4-fold increase in identified spectra over isolations from control iCMs. Due to the ascribed function of TBX20 as a critical cardiac transcription factor, we specifically focused on proteins with a nuclear or unknown subcellular localization. Finally, these interaction candidates were ranked by their AP enrichment (AP abundance versus whole cell abundance, S1 Table), which we have previously used to highlight the most prominent associations suitable for functional validation [31, 32]. Interestingly, the top 50 most enriched proteins in the TBX20 AP were predominately (32/50) annotated to Chromatin and Transcription gene ontologies (Fig 2A; S1 Table). Functional annotation of these proteins in the STRING database [33] revealed an interconnected network containing components of chromatin remodeling and RNA polymerase transcriptional complexes (including four components of the INO80 complex—Ino80, Actr5, Actr5, Nfrkb, and five components of the RNA Pol II mediator complex- Med13, Med14, Med17, Med19, Med27) (Fig 2A). These data suggest that TBX20 predominantly acts to regulate transcription in neonatal cardiomyocytes, likely via interactions with the INO80 and RNA Pol II mediator complexes.
In addition to identifying components of broadly expressed multiprotein chromatin machines, our analysis revealed the association of TBX20 with the essential cardiac transcription factor CASZ1 in BirA-expressing iCMs (25 unique peptides and 21% sequence coverage. As previously reported [34, 35], CASZ1 protein runs as three bands on at approximately 191kD, presumably due to post translational processing) (Fig 2B, S2C Fig). Surprisingly, this was the only developmentally-regulated cardiac transcription factor we found to interact with TBX20 in Day 7 cardiomyocytes. The low estimated cellular abundance of CASZ1 and the relatively high AP enrichment ratio (S1 Table) highlighted CASZ1 as a potential in vivo TBX20 interacting protein. We further confirmed the TBX20-CASZ1 interaction through reciprocal immuno-isolation of endogenous CASZ1 in cardiac nuclei from adult mouse hearts (Fig 2C) thus, verifying our ESC differentiation-based approach can successfully identify bona fide TBX20 interaction partners under physiological conditions. Since expression analysis and genetic fate mapping studies have shown CASZ1 is expressed only in cardiomyocytes and no other cardiac cell types [36], and since we were unable to identify this interaction at Day 4 of iCM differentiation, these studies imply the interaction between TBX20 and CASZ1 is temporally regulated and cardiomyocyte-specific.
Phylogenic analysis shows that TBX20 and CASZ1 are highly conserved across vertebrate orthologs [34, 35], suggesting that the TBX20-CASZ1 interaction may also be evolutionarily conserved. To confirm the interaction and to determine whether it is evolutionarily conserved, we injected X. laevis embryos with the Xenopus orthologous mRNAs of TBX20 and CASZ1. In parallel, we co-expressed murine versions of tagged TBX20 and CASZ1 proteins in HEK293 cells. Immunoaffinity purification of TBX20 protein complexes from both of these sources, followed by immunoblotting confirms the formation of a TBX20-CASZ1 interaction in human cells and in X. laevis embryos (S3A and S3B Fig). Taken together, our findings are supportive of an evolutionarily conserved role for the formation of a TBX20-CASZ1 protein complex in differentiated cardiomyocytes.
To determine the biological relevance of the TBX20-CASZ1 interaction, we tested for genetic interaction between Tbx20 and Casz1 by generating mice with cardiac-specific heterozygous loss of Tbx20 and Casz1 (Tbx20flox/+; Casz1flox/+; Nkx2.5Cre)[21, 36]. Compound heterozygous mice, hereafter referred to as Tbx20flox/+; Casz1flox/+, were born and appeared normal. However, beginning at 4 weeks of age we observed an increased incidence of death among Tbx20flox/+; Casz1flox/+ mice (2.7%) compared to the single heterozygotes (0%) (Table 1). This effect was amplified at later timepoints, with survival rates of 90.5% and 62.5% at 8 and 16 weeks of age, respectively, compared to a 100% survival rate in the single heterozygotes. Furthermore, we did not observe any overt phenotypes in Tbx20flox/+; Nkx2.5Cre or Casz1flox/+; Nkx2.5Cre mice. Since we were able to demonstrate that loss of Casz1 does not affect Tbx20 expression in adult heart tissue and that loss of Tbx20 does not affect Casz1 expression (S4A and S4B Fig), these studies are supportive of a genetic requirement for a functional interaction between TBX20 and CASZ1.
To determine the cause of the reduced survival rate we observe in Tbx20flox/+; Casz1flox/+ mice, we performed detailed physiological analysis, using echocardiography, of single and compound heterozygous mice. These studies revealed that cardiac function is significantly compromised in Tbx20; Casz1 compound heterozygotes compared to single heterozygotes. Compound heterozygotes exhibit significantly decreased ejection fraction and fractional shortening, increased left ventricular blood volume, and increased left ventricular diameter (Fig 3A–3D; Tables 2 and 3; S2 Table; S2–S6 Movies). Further, Tbx20flox/+; Casz1flox/+ heterozygous mice display dilated ventricles with a striking decrease in ventricular wall thickness compared to single heterozygotes (Fig 4A and 4B). These findings were observed in both male and female mice (S3 Table). Thus, Tbx20flox/+; Casz1flox/+ mice display defining anatomical features of DCM that progress to cardiac failure. Collectively, these findings suggest that the genetic interaction between Tbx20 and Casz1 is essential for normal cardiac homeostasis, and perturbation of this interaction leads to DCM.
One of the defining clinical features of severe DCM is an accumulation of myocardial collagen leading to interstitial fibrosis, a contributing and compounding factor in cardiac dysfunction [37–39]. To confirm that the severe cardiac dysfunction we observe in Tbx20; Casz1 compound heterozygotes is associated with advanced DCM, we examined collagen fibers and found robust collagen deposition in the interstitium of Tbx20flox/+; Casz1flox/+ hearts (Fig 4B and 4C). Despite the interstitial fibrosis and severely impaired systolic function, the fact that over half of these mice survive to adulthood with some degree of cardiac function led us to hypothesize that Tbx20; Casz1 compound heterozygous cardiomyocytes undergo compensatory pathological hypertrophy. To test this hypothesis, we measured cardiomyocyte cross-sectional areas and found that MF20-positive cells in compound heterozygotes were indeed increased in size relative to controls (Fig 4D and 4E). This data suggests that disrupting the TBX20-CASZ1 interaction leads to severe DCM and cardiac fibrosis. In response to this heightened cardiac stress, Tbx20; Casz1 compound heterozygote hearts appear to undergo pathological hypertrophy as an adaptive response.
Our data implies the TBX20-CASZ1 interaction is essential for normal cardiac homeostasis. To define the region of TBX20 that mediates interaction with CASZ1, we conducted immunoisolations with wild-type and deletion mutants in which either the T-Box or the C-terminus of TBX20 has been removed (Fig 5A). Immunopurifications of CASZ1 in the presence of wild-type TBX20, TBX20ΔT-box, or TBX20ΔC, show that the T-box domain is required for interaction with CASZ1, but that the C-terminus is dispensable (Fig 5A). In reciprocal studies, we find the four most amino-terminal zinc finger domains of CASZ1 are necessary for interaction with TBX20 (Fig 5B). The CASZ1-interacting region of TBX20, as well as the TBX20-interacting region of CASZ1, are highly conserved across species implying functional relevance to these regions (Fig 5C).
Recently, human TBX20 mutations have been identified that are associated with DCM; however, only one of these mutations, TBX20F256I, co-segregates in a dominant manner with complete penetrance in a family with DCM [9]. Moreover, DCM was found in all affected family members reported as healthy during health assessments performed when they were juveniles. The functional relevance of the F256I mutation is further underscored by the finding that the amino acid disrupted by this mutation is 100% conserved across all TBX20 orthologs and by the observation that no F256I mutations were identified in 600 control samples [9]. Interestingly, the F256I mutation associated with DCM lies within the TBX20 T-box domain, the region we found essential for interaction with CASZ1 (Figs 5A and 6A). To test if TBX20F256I perturbs the TBX20-CASZ1 interaction, we performed immunoaffinity purifications of CASZ1 in the presence of wild-type TBX20 or TBX20F256I. Interestingly, the F256I mutation significantly reduces the interaction with CASZ1 (Fig 6A). These data imply that the DCM mutation F256I may contribute to the development of cardiac disease by disrupting a critical physical interaction between TBX20 and CASZ1.
To gain a structural understanding of how the F256I mutation disrupts the TBX20-CASZ1 interaction, we conducted molecular modeling of the wild-type and TBX20F256I T-box domain (Fig 5B and 5C). The predicted structures were based on the range of fluctuations in the structure that occur over a period of 100 ns (S4 Movie). Three regions are highlighted which show conformational changes induced by the mutation (Fig 6B). Our models find F256 is not predicted to contact DNA but the conversion of phenylalanine to isoleucine at position 256 leads to steric clashes with the conserved T-box residues E258 and T259 (Fig 6C). The critical functional nature of this region of TBX20 is underscored by the complete conservation of amino acids at residues F256, E258, and T259 across 250 members of the T-box gene family (S5A and S5B Fig). Taken together, these findings imply that F256I leads to a conformational change across the surface predicted to interact with CASZ1, and that disruption of this interaction leads to alteration in DNA binding.
To determine the transcriptional consequences of TBX20F256I on the TBX20:CASZ1 interaction, we conducted transcriptional assays with TBX20, CASZ1 and TBX20F256I alone and in combination. Results demonstrate TBX20 synergistically acts with CASZ1 and that TBX20F256I significantly diminishes transcriptional activation by TBX20:CASZ1 (S6 Fig). These data together with our structural studies provide a mechanistic basis for how F256I disrupts TBX20:CASZ1 function.
To identify the molecular pathways altered in DCM haploinsufficient mutant (Tbx20flox/+; Casz1flox/+; Nkx2-5Cre) mouse hearts, we used quantitative multiplexed mass spectrometry to identify proteins with altered abundances relative to control hearts. Proteins were extracted from nuclear-enriched mouse cardiac fractions of mutant and control (Nkx2-5Cre) mice in duplicate and digested in-solution with trypsin. Peptides from each sample were labeled with different isobaric tandem mass tagging (TMT) reagents, pooled, fractionated, and analyzed by reverse phase nanoliquid chromatography coupled to a high resolution quadrupole Orbitrap tandem mass spectrometer. Using this strategy, 3164 proteins were identified and quantified based on their respective sequenced peptides and TMT reporter ions, respectively (S4 Table).
To define the TMT ratio threshold for differential relative abundance, protein abundance values were compared between biological duplicates (S7 Fig). For both the control and mutant replicates, the correlation of abundances was high (R2 = 0.99) and the majority of proteins had low dispersion from a 1:1 linear curve (S7A and S7B Fig), indicating low biological and technical variation. Curve-fit analysis of TMT abundance ratio histograms for the control and mutant biological duplicates showed that, on average, 90% of the ratios varied less than ±30% (S7 Fig). Based on this result, a relative abundance ratio of at least ±1.3-fold between mutant and control mice in both replicates were used to identify a protein as differential. From the total number of quantified proteins, 175 met this criterion, of which 86 and 89 were up and down-regulated, respectively (Fig 7C; S4 Table). Further verifying the role of the TBX20; CASZ1 interaction, 165 of the 175 of the proteins identified by this approach were encoded by a gene previously demonstrated to be a TBX20 target [40](S5 Table).
To generate an initial picture of potentially dysregulated pathways, the known functional connectivity among differential proteins can be determined using databases of annotated pathways and protein-protein interaction. Towards this goal, known functional associations among the 175 differential proteins were scored based on the STRING bioinformatics database [41], and the relational networks visualized in Cytoscape [42] (Fig 7D; S8 Fig). A high degree of interconnectivity was observed among the differential proteins as 127 of the 175 annotated proteins had at least one other connection and each protein on average was connected to 4.6 neighbors. Network clustering was performed to identify subsets of highly connected proteins, which likely share similar functions. Overall, there are 10 functional clusters containing at least 3 proteins, indicated by the color-coding in Fig 6C. To identify the most significant biological processes and pathways that are perturbed in Tbx20; Casz1 hypomorphic DCM hearts, we performed comparative Gene Ontology over-representation analysis of the differentially regulated proteins using ClueGO [43] (Fig 7E). Consistent with studies demonstrating an association between DCM and inflammation [44, 45], we found components of the pro-inflammatory response (i.e. complement activation) significantly up-regulated.
In addition to inflammation-associated proteins, our data further revealed a dysregulation of mitochondrial proteins known to be associated with impaired cardiomyocyte contractile function in DCM [46]. In line with these findings and the observation that reduced contractile force is linked to altered glycogen metabolism and cardiomyopathy [47–49], we found an over-representation of proteins associated with the glycogen metabolic pathway. We note that these were exclusively down-regulated proteins, represented by glycogen synthase (Gys1), glycogen phosphorylases (Pygm/Pygb), and phosphorylase kinase gamma 1 (Phkg1). Interestingly, proteins involved in glycogen regulation and in myosin-dependent muscle contractility were part of the same functional cluster (Fig 7D, yellow nodes); however, their individual abundances were down- and up-regulated, respectively (Fig 7D, circle vs. square nodes). Taken together, these data confirm at the protein level the DCM pathology in Tbx20; Casz1 hypomorphic DCM mice.
In addition to proteins previously reported to be associated with DCM, our analysis identified a distinct set of cell-cell adhesion proteins in Tbx20; Casz1 compound heterozygous hearts that were significantly overrepresented compared to whole genome annotation (Fig 7E; S8 Fig, yellow). These observations highlight the significant changes that are likely occurring in the extracellular and intracellular spaces and raise a key question- what are the signaling mediators that link these processes? To identify potential key mediators of TBX20-CASZ1-driven DCM, we constructed a gene-linked GO network for the Cellular Component ontology (S8 and S9 Figs). This network highlighted two interesting candidates, bone morphogenic protein 10 (Bmp10) and thrombospondin 1 (Thbs1), the former being a TGF-beta receptor ligand and the latter having roles in cell-cell adhesion as well as ER stress response [50]. Overall, this systems-level proteome view of DCM provides potential downstream targets and pathways that may be influenced as a result of Tbx20 and Casz1 haploinsufficiency and suggests a role for cell-cell adhesion in mediating DCM.
One in five adults free of cardiovascular disease by the age of 40 are at risk of developing congestive heart failure over their lifetime [51, 52]. Within this population DCM remains one of the leading causes of disease and death with nearly half of DCM cases genetically determined [53–57]. To date, most DCM mutations have been identified in genes coding for components of the contractile apparatus or the cell cytoskeleton or factors involved in excitation-conduction coupling. Though these studies have provided insight into the pathology of DCM, the transcriptional regulation of DCM is poorly understood. Recently, studies have identified mutations in the T-box transcription factor TBX20 in DCM patients [7–9]. However, these studies have not explained how DCM-associated mutations bypass early essential requirements for TBX20. Here, we demonstrate TBX20 and CASZ1 physically and genetically interact in the adult heart and establish that this interaction is essential for cardiac homeostasis. We further find that disruption of the TBX20-CASZ1 interaction in mice and humans leads to cardiomyopathy. Mice singly heterozygous for alleles of Tbx20 or Casz1 are asymptomatic, while Tbx20; Casz1 compound heterozygotes die post-natally, exhibiting systolic dysfunction, as well as ventricular dilation and interstitial fibrosis. These cardiac defects, in the absence of coronary artery disease or substantially abnormal load, are defining features of DCM as seen in human patients [6, 58, 59].
CASZ1 is a large para-zinc finger protein of unique structure and to date, there have been limited studies on the mechanisms of how CASZ1 regulates transcription [60–62]. These types of studies have been compromised by the lack of high-affinity high-specificity mammalian CASZ1 antibodies, precluding approaches such as ChIP-seq. It further remains unclear if CASZ1, as a para-zinc finger protein, directly binds DNA or is recruited via other transcription factors. Our structural studies favor a model by which the TBX20-CASZ1 interaction is required for DNA binding. This model predicts that the respective region of TBX20 that binds CASZ1 is near to or contributes to the DNA binding interface and has the potential to impact CASZ1 binding.
CASZ1 was first ascribed a role in vertebrate cardiovascular development in Xenopus [34, 62, 63]. Subsequent genetic studies in mammals uncovered that like TBX20, CASZ1 functions in the embryonic heart to control cardiomyocyte proliferation, with loss of CASZ1 leading to cardiac death by E12.5 [36, 64]. Our finding that Tbx20; Casz1 compound heterozygous mice die post-natally implies that CASZ1 has a second and later role in cardiac homeostasis. This model is supported by the recent finding that mutations in CASZ1, like TBX20, are associated with human DCM [26].
In these studies, the Nkx2.5-Cre driver was used to generate mice null for Casz1 and Tbx20. In all cases the Nkx2.5-Cre driver alone was used, with wild-type mice as a negative control in our physiological studies and in our quantitative proteomic studies. In contrast to previous reports [65], we could detect no significant changes in any cardiac function in Nkx2.5-Cre mice relative to wild-type mice. These finding may be due to genetic background, the sex on which the Cre driver was delivered to the offspring, or environmental variability as reported for other lines [66, 67]. Regardless of the reason for the variability, we did find the Nkx2.5-Cre driver had a high recombination efficiency reducing the levels of TBX20 and CASZ1 by half (S4 Table). Since, reducing CASZ1 expression by half had no detectable alteration in Tbx20 expression and vice versa, our data suggests a biochemical and genetic interaction between TBX20 and CASZ1. Our data further indicates that disruption of this complex leads to DCM in mice and humans.
A previously published model of DCM, the phospholamban R9C transgenic mouse [68], has also been studied by proteomic analysis [69, 70]. This model exhibits impaired calcium regulation in cardiomyocytes, accompanied by decreased cardiac contractility and premature mortality [68]. The GO-associated proteome changes that we found in the haploinsufficient mice share similarities with the Phospholamban R9C mice. Specifically, both mouse models show up-regulation of actin-myosin cytoskeletal networks and down-regulation of mitochondria-associated proteins involved in fatty acid oxidation. Interestingly, proteomic analyses performed on ventricular tissues from human patients with inflammatory DCM had similar findings [71]. Yet some functional protein classes in our Tbx20-CASZ1 haploinsufficient DCM mice were distinct, including an up-regulation of the complement system and greater coverage of down-regulated proteins in glycogen metabolic processes. While we found the Tbx20-CASZ1 haploinsufficient mice have evidence of differential regulation in calcium-binding proteins, not surprisingly, the Phospholamban R9C mice have more pervasive effects on calcium-dependent signaling, such as involving ER stress responses, though it is possible that these distinctions may be due to differences in the progression of the fibrosis associated with DCM.
One of the hallmarks of DCM is altered cardiomyocyte force transduction that is frequently associated with alteration in the composition or functions of intercalated discs- a cardiac-specific structure at the contact site between cardiomyocytes [72, 73]. Here, we observed a significant mis-regulation of proteins involved in cell-cell adhesion in heart tissue from Tbx20; Casz1 heterozygous mice. Moreover, these include three proteins which are encoded by genes that when mutated are causative to DCM- TTN, DES, and PDLIM3. Thus, our findings imply that the TBX20-CASZ1 complex acts, at least in part, to control the electrical and mechanical integration of neighboring cardiomyocytes.
The observation that the TBX20F256I mutation leads to a decreased association with CASZ1, along with the finding that patients heterozygous for a predicted TBX20 null mutation (TBX20Q195X) [23] also display DCM suggest that the TBX20F256I mutation may be acting in a haploinsufficient fashion. However, only two of the individuals within a single pedigree with the TBX20Q195X mutation display DCM while other individuals display a wide range of cardiac abnormalities [23]. Moreover, we have screened the Exome Aggregation Consortium (ExAc) reference set and have identified four variants in TBX20 in individuals that are asymptomatic. All variants lead to a premature stop codon in one of the TBX20 alleles and all would be predicted to be functionally null (introduction of stop codons into exons 2, 4, 7, and 8)[74]. Together, these findings imply that the function of the TBX20-CASZ1 complex in DCM is not dose dependent. Alternatively, individuals harboring the TBX20F256I mutation have a genetically sensitized background leading to a varying degree of penetrance that is determined by modifying genes that may be carried within the CASZ1 pathway.
In cardiovascular disease, genetic mutations often result in varying degrees of penetrance, and in extreme examples, the presence of a disease-causing mutation can be asymptomatic [75–80]. These phenomena have often been explained by the action of genetic modifiers in which one gene mutation is causative to CHD and a second mutation modifies the effect of the first. However, more recent studies suggest an alternate or additional mechanism by which complete penetrance is achieved in human disease states by genetic variation at one or more loci [81]. In digenic inheritance, two genetic mutations are required for the clinical phenotype with either mutations alone being asymptomatic. Our findings provide an example of digenic inheritance in DCM and suggest that mutations in TBX20 or CASZ1 could lead to susceptibility to DCM but in many cases are not in themselves causative. We would envision these findings are not restricted to TBX20 and CASZ1 but rather are applicable to other genes and other forms of congenital heart disease (CHD) and DCM, and predict that genome sequencing of familial CHD will ultimately reveal a spectrum of additional CHD susceptibility alleles.
The Tbx20Avi allele was created by introducing the biotin acceptor peptide (Avi) targeting cassette, similar to our previous study [82], in-frame to the terminal exon of Tbx20 in collaboration with the UNC Animal Models Core and the UNC BAC Core (Chapel Hill).
The Tbx20Avi; BirA cell line was generated by targeting a sequence containing the Avitag followed by a loxP-flanked neo cassette into the stop codon of exon 8 of a Tbx20a genomic fragment derived from a 129 Sv genomic BAC library. The targeting construct was linearized and electroporated into mouse embryonic stem cells (ESCs) of E14TG2a.4 origin. Targeted ESCs were placed under 250 μg/mL G418 selection for 7–10 days and G418-resistant ESC clones (n = 384) were screened for homologous recombination by Southern blot analysis. Three ESC clones were correctly targeted, and one of these clones was subsequently used to derive the Tbx20Avi/+; BirA cell line. Briefly, Tbx20Avi/+ ESCs were grown to approximately 40% confluence and transduced with 5 MOI Lenti-BirA for 8 hrs. Twenty-four hours following transduction, cells were placed under 200 μg/mL hygromycin selection for 4–5 days. Hygro-resistant Tbx20Avi/+ cells were subsequently used for cardiomyocyte differentiations.
Tbx20Avi/+; BirA ESCs were maintained on gelatin-coated dishes in a feeder-free culture system and differentiated [29] in serum-free (SF) media according to the Keller protocol. Briefly, ESCs were trypsinized and cultured at 75,000 cells/mL on uncoated petri dishes in SF medium without additional growth factors for 48 hrs. Two-day-old aggregated embryoid bodies (EBs) were dissociated and the cells reaggregated for 48 hr in SF medium containing 5 ng/mL human Activin A, 0.1 ng/mL human BMP4, and 5 ng/mL human VEGF (all growth factors purchased from R&D Systems). Four-day-old EBs were dissociated and 2 x 106 cells were seeded into individual gelatin-coated wells of a 6-well dish in StemPro-34 SF medium (Invitrogen) supplemented with 2 mM L-glutamine, 1 mM ascorbic acid, 5 ng/mL human VEGF, 20 ng/mL human bFGF, and 50 ng/mL human FGF10 (R&D Systems). Cardiomyocyte monolayers were maintained in this media for 4–5 additional days with cells typically beginning to beat 2 days after seeding onto gelatin (total of 7–8 days of differentiation).
For immunofluorescence of cardiomyocytes, four-day-old ES cell-derived EBs were dissociated and seeded into 8-well chamber slides precoated with 0.1% gelatin. Induced cardiomyocytes were fixed on day 7 of differentiation in 4% paraformaldehyde for 20 min at room temperature, washed (3 x 1X PBS), permeabilized in 0.1% Triton X-100 in 1X PBS for 10 min, and blocked (10% fetal bovine serum [FBS], 0.1% Tween 20 in 1X PBS) for 30 min. Anti-myosin heavy chain (Abcam) was applied overnight, followed by PBS washes (3 x 1X PBS), and incubation with goat anti-mouse Alexa 546 (Invitrogen) for 1 hr. Cells were incubated in DAPI (200 ng/mL in ethanol) for 30 min and visualized by confocal microscopy on a Zeiss 710.
Protein preparations, conjugation of magnetic beads and immunoaffinity purification and mass spectrometry were conducted as previously reported [82]. All results are from a minimum of two independent biological replicates. Briefly, immunoisolated proteins were resolved (~ 4 cm) by SDS-PAGE, and visualized by Coomassie blue. Each lane was subjected to in-gel digestion with trypsin and analyzed by nanoliquid chromatography coupled to tandem mass spectrometry as previously reported [83]. Tandem mass spectra were extracted by Proteome Discoverer (ThermoFisher Scientific, ver 1.4), and searched with the SEQUEST algorithm against a theoretical tryptic peptide database generated from the forward or reverse entries of the mouse UniProt-SwissProt protein sequence database (2013/08) and common contaminants (total of 43, 007 sequences). SEQUEST search results were analyzed by Scaffold (version 4.6.1, Proteome Software Inc) using the LFDR scoring scheme to calculate peptide and protein probabilities. Peptide and protein probabilities thresholds were selected to achieve ≤ 1% FDR at the peptide level based on LFDR modeling and at the protein level, based on the number of proteins identified as hits to the reverse database. The spectral counts assigned to proteins that satisfy these criteria and had a minimum of two unique peptides were exported to Excel for data processing.
Proteins identified by LC-tandem MS were filtered to exclude non-specific associations. Proteins were retained as specific interaction candidates if the proteins were assigned (1) at least ten spectral counts in the Tbx20Avi;BirA condition, and (2) were uniquely identified or had at least a 4-fold spectral count enrichment in the Tbx20Avi; BirA condition versus the control. Next, the subset of candidates assigned a nuclear or unknown UniProt subcellular localization were retained for calculation of enrichment index values, as previously described [31]. Briefly, the relative protein abundance within the affinity purification was calculated using the NSAF approach [84], then normalized by each protein’s respective cellular abundance estimated in the PAX database [85] (Mouse—whole organism, SC GPM 2014). Interaction candidates were ranked by their enrichment index and the top 50 proteins were analyzed by STRING [86] for interaction network analysis. Interactions with a combined STRING score of > 0.4 (medium confidence) were retained, exported, and visualized in Cytoscape (ver. 3.3). Proteins within the network were assigned into broad protein functional classes based on annotations in the UniProtKB database.
Western blots were probed with the following primary antibodies overnight at 4°C: mouse anti-V5 (Invitrogen) 1:5000; mouse anti-GFP (JL-8, Clontech) 1:10000; mouse anti-HA-HRP (Cell Signaling #2999) 1:1000, mouse anti-GAPDH (Millipore) 1:1000, goat anti-TBX20 (Santa Cruz Biotechnology) 1:600, rabbit anti-CASZ (Santa Cruz Biotechnology) 1:1000, and chick anti-BirA (Abcam) 1:2000. After being rinsed, blots were rinsed in the following secondary antibodies for 1 hr at room temperature: anti-IgG2a-HRP (Jackson Immunoresearch) 1:10000. Antibody-antigen complexes were visualized using an ECL Western Blotting Analysis System (Amersham).
Tbx20flox/+ mice were generously provided by Sylvia Evans (UCSD) [21]. The Casz1flox/+ mouse has been previously reported [36]. Histological sectioning and immunohistochemistry were done as reported except as noted [36]. All mice are on a mixed B6/129/SvEv/CD-1 background and all mouse experiments were performed according to the Animal Care Committee at the University of North Carolina, Chapel Hill.
Cardiac function was assessed in conscious 8–11 week-old Casz1flox/+; Nkx2.5Cre, Tbx20flox/+; Nkx2.5Cre, and Tbx20flox/+; Casz1flox/+; Nkx2.5Cre mice (5–10 mice per genotype) by thoracic echocardiography using VisualSonics Vevo 770 ultrasound system (Visual Sonics, Inc.). All imaging was done by trained technicians blinded to the genotypes of the animals. Briefly, a topical hair removal agent was used on the chest and abdomen of mice. The mice placed on a warmed table in the supine position for imaging. A 30 MHz pediatric probe used to capture 2-dimensional guided M-mode views of the long and short axes at the level of the papillary muscle. VisualSonics Analytic software was used to determine mean ventricular wall and interventricular septum thickness, as well as the left ventricle diameter from at least 3 consecutive cardiac cycles. Means were used to calculate ejection fraction and fractional shortening.
All statistical analysis performed using SAS JMP 10. Statistical significance between individual groups was calculated using Student’s T-test, while significance between more than 2 groups was calculated using ANOVA.
|
10.1371/journal.pcbi.1004491 | Perm-seq: Mapping Protein-DNA Interactions in Segmental Duplication and Highly Repetitive Regions of Genomes with Prior-Enhanced Read Mapping | Segmental duplications and other highly repetitive regions of genomes contribute significantly to cells’ regulatory programs. Advancements in next generation sequencing enabled genome-wide profiling of protein-DNA interactions by chromatin immunoprecipitation followed by high throughput sequencing (ChIP-seq). However, interactions in highly repetitive regions of genomes have proven difficult to map since short reads of 50–100 base pairs (bps) from these regions map to multiple locations in reference genomes. Standard analytical methods discard such multi-mapping reads and the few that can accommodate them are prone to large false positive and negative rates. We developed Perm-seq, a prior-enhanced read allocation method for ChIP-seq experiments, that can allocate multi-mapping reads in highly repetitive regions of the genomes with high accuracy. We comprehensively evaluated Perm-seq, and found that our prior-enhanced approach significantly improves multi-read allocation accuracy over approaches that do not utilize additional data types. The statistical formalism underlying our approach facilitates supervising of multi-read allocation with a variety of data sources including histone ChIP-seq. We applied Perm-seq to 64 ENCODE ChIP-seq datasets from GM12878 and K562 cells and identified many novel protein-DNA interactions in segmental duplication regions. Our analysis reveals that although the protein-DNA interactions sites are evolutionarily less conserved in repetitive regions, they share the overall sequence characteristics of the protein-DNA interactions in non-repetitive regions.
| Chromatin immunoprecipitation followed by high throughput sequencing (ChIP-Seq) is widely used for studying in vivo protein-DNA interactions genome-wide. The applicability of this method for profiling repetitive regions of the genome is limited due to short read sizes dominating ChIP-seq applications. We present Perm-seq, which implements a novel generative model for mapping short reads to repetitive regions of genomes. Perm-seq introduces a new class of read alignment algorithms that can combine data from multiple sources. We show with both computational experiments and the analysis of large volumes of ENCODE ChIP-seq data that utilizing DNase-seq derived priors in Perm-seq is especially powerful in mapping protein-DNA interactions in segmental duplication regions. This general approach enables the use of any number of histone ChIP-seq data alone or together with DNase data to supervise read allocation. Our large scale analysis reveals that although the protein-DNA interactions sites are evolutionarily less conserved in repetitive regions, they share the overall sequence characteristics of the protein-DNA interactions in non-repetitive regions.
| Chromatin immunoprecipitation followed by sequencing (ChIP-seq) has become a versatile high throughput assay for profiling of transcription factor (TF) binding and histone modifications. A typical ChIP-seq experiment generates millions of short reads (50–100 bps). The first step of any standard ChIP-seq data analysis pipeline involves mapping reads to a reference genome. In any given ChIP-seq experiment, a considerable fraction (5–30%) [1] of the reads can align to multiple locations on the genome (multi-reads) thereby creating ambiguity regarding their true origin. Although there have been some prior efforts in developing ChIP-seq specific mappers that can allocate multi-mapping reads to one of their mapping positions based on local counts of uniquely mapping reads [1–5] (uni-reads), the standard practice for ChIP-seq experiments is to either use only uniquely mapping reads or retain a conservative set of multi-mapping reads (e.g., with at most 2–3 mapping positions) and utilize one of the mapping positions randomly [6]. This bottleneck has serious downstream effects when characterizing regulatory elements common or specific to distinct cell types where, for example, cell-type specific characteristics that reside in repetitive regions are grossly under-represented. Similar observations have been made for MeDIP-seq analysis [7], where repetitive elements were severely underestimated in the traditional alignment and analysis of sequencing based data. This is a highly critical barrier especially to the advancement of analysis of large consortia (e.g., Encyclopaedia of DNA Elements (ENCODE)) data because significant fractions of eukaryotic genomes are composed of repetitive regions, e.g., more than half of the human genome. Genomic repeats play important roles in function and evolution of transcriptional regulatory networks [8, 9] making their functional annotation of highest biological importance. Segmental duplications played roles in creating new primate genes and contributed to human genetic variation [10]. Therefore, utilization of multi-mapping reads is especially important for characterizing regulatory activity in segmental duplications or LINE elements that harbor near-identical DNA sequences.
Current state-of-the-art approaches for allocating multi-mapping reads in both RNA-seq [11–17] and ChIP-seq [1–5] studies rely on utilizing read counts of the local neighbourhoods of the mapping positions. Therefore, these approaches have critical limitations when the local neighbourhood read counts of the mapping positions are highly similar. In order to resolve this bottleneck and improve specificity of multi-read allocation in ChIP-seq studies, we develop a novel strategy that utilizes data from DNase I hypersensitive sites sequencing (DNase-seq) experiments to inform read allocation in ChIP-seq. DNase-seq is a high-resolution assay for mapping active cis-regulatory elements across the genome [18, 19]. DNase-seq identifies broader regions of open chromatin which often exhibit transcription factor occupancy. Although these experiments do not provide specificity as to which regulatory factors occupy the captured accessible regions of the genome, there is a growing literature indicating their high predictive ability for identifying protein-DNA binding sites [20, 21]. Utilizing a large number of ENCODE ChIP-seq datasets from GM12878 and K562 cells, we show that DNase-seq has significant power for discriminating between the mapping locations of multi-reads with similar local ChIP-seq read counts. We develop a probabilistic model that utilizes DNase-seq derived priors (Perm-seq) and maps reads originating from highly repetitive regions with high accuracy. The Perm-seq framework is highly versatile and enables incorporation of multiple sources of data such as histone ChIP-seq for read allocation.
Our reanalysis of a large collection of ENCODE ChIP-seq datasets with Perm-seq identifies many novel protein-DNA interaction targets in highly repetitive regions of the genome, especially in segmental duplications. Our detailed analysis of these novel targets indicate that repetitive and non-repetitive modes of protein-DNA interactions share similar chromatin signatures. Although protein-DNA interaction sites in the repetitive regions are evolutionarily less conserved, they share sequence characteristics of the protein-DNA interaction regions identified by uniquely mapping reads. Furthermore, our analysis identifies that H2a.z ChIP-seq data performs as well as DNase-seq data for supervising allocation of multi-reads in highly repetitive regions.
In order to explore the discriminating power of DNase-seq data for allocating multi-mapping ChIP-seq reads, we identified all the reads that map to two different locations on chromosome 2 from an ENCODE Gata2 ChIP-seq experiment performed in Huvec cells [22]. Since all the available multi-read allocation methods utilize uni-read counts in local neighbourhoods of the mapping locations, we evaluated how different the local neighbourhoods are for the two locations that each read maps to. In addition, we also counted the number of DNase-seq reads in these neighbourhoods. This investigation revealed that the ChIP uni-read counts of the local neighbours of the two mapping locations are very similar with log base 2 fold changes smaller than 0.5, i.e., | log2(ChIP counts at location 1/ChIP counts at location 2)| < 0.5, for majority of the reads (70%, Fig 1(a)). However, DNase read counts together with ChIP read counts can discriminate between the two mapping locations with a log based 2 fold change of at least 0.5 for 49% of the reads, resulting in a 26% increase in the discrimination power. Similar trends in discriminative power are observed for a range of fold change thresholds (Table S1 in S1 Text).
We performed a similar analysis for four additional datasets (Fig 1(b), Fig. S1 in S1 Text): Atf3 and cFos in GM12878 and K562 cells, and observed that utilizing local ChIP read counts helps to distinguish between the two mapping locations of only 17–21% of the reads whereas DNase read counts alone were able to discriminate between two mapping locations for an additional 24–36% of the reads. For a small percentage of the reads (≤ 7%), the fold changes of DNase and ChIP read counts of the two mapping positions appear to be in opposite directions, i.e., DNase read counts in position 1 > DNase read counts in position 2, while ChIP read counts in position 1 < ChIP read counts in in position 2. This is attributable to the fact that signal measured by DNase-seq is not regulatory factor specific as well as to heterogeneity of cells the ChIP- and the DNase-seq experiments are conducted on.
Perm-seq models observed ChIP-seq read alignments conditional on DNase-seq data and significantly extends our prior work, CSEM [1] on multi-read allocation. Specifically, DNase-seq read counts are incorporated into a multinomial read generating distribution with a Dirichlet-multinomial regression model. The Dirichlet-multinomial regression model includes a log-linear prior that is a function of DNase-seq read counts. The Perm-seq framework involves three major components (Fig 1(c)). First, DNase-seq read counts are mapped to the reference genome with CSEM by taking into account multi-reads. In this step, each multi-mapping read is allocated to one of its mapping positions if the allocation probability is at least 0.9. This is quite a conservative mapping strategy for multi-reads; however it avoids generation of spurious DNase signals. Then, ChIP-seq reads are aligned to the reference genome with Bowtie [23] (using parameters -q -v 2 -a -m 99) while retaining both uni- and multi-reads. The next step involves building a log-linear model by regressing ChIP-seq uni-read counts on DNase-seq read counts. We utilize a data aggregation strategy to build such a model. Exploratory analyses of multiple ChIP-Seq datasets indicate that B-spline models provide very good fits (Fig. S2 in S1 Text) and capture how ChIP-seq read counts vary with DNase-seq signal. We use this estimated B-spline model to generate prior parameters for the Dirichlet-multinomial model and allocate multi-mapping reads to each of the mapping locations by weighting the evidence from ChIP-seq read counts of the local neighbourhoods and the corresponding DNase-seq priors. The output from Perm-seq is a file with all the uni- and multi-reads and includes final allocation probabilities for each of the mapping locations of the multi-reads. Perm-seq also provides functionality to convert this file into a BED file where each multi-read is allocated to its best mapping position with the largest allocation probability. Such a BED file can then be used by many standard peak calling methods. We used ENCODE’s uniform ChIP-seq processing pipeline [24], which is built on the peak caller SPP [25] and irreproducible discovery rate (IDR) [26] for determining the optimal number of peaks, in the analyses presented in this paper.
The overall effect of using prior information in ChIP-seq multi-read allocation with Perm-seq is an increase in the number of detected peaks compared to analyses that utilize only uni-reads (uni-read analysis) (Fig. S3 and Tables S2 and S3 in S1 Text) and a decrease in the number of peaks compared to analyses using multi-reads without prior information (CSEM). Analysis of 64 ChIP-seq datasets from GM12878 and K562 cells resulted in an average increase of 8.0% (with a standard deviation of 13.1%) in the number of peaks when comparing uni-read analysis with Perm-seq and a decrease of 4.6% (with standard deviation 9.8%) when comparing Perm-seq with CSEM. Fig 2(a) summarizes the numbers of peaks and peaks overlapping across the three different peak classes, namely, uni-read, CSEM, and Perm-seq in both the optimal and the relaxed mode of peak calling for transcription factor Ctcf. Optimal peak sets are obtained at an IDR of 2%. The relaxed peak sets are super sets of the optimal peaks and include both high signal peaks and regions that do not show any ChIP enrichment. We included comparisons with the relaxed peak sets to ensure that our results hold irrespective of the specific IDR threshold used for peak calling. We observed that 1320 peaks are identified by both Perm-seq and CSEM analyses, whereas Perm-seq identified 187 peaks that are not part of CSEM optimal list, and similarly, CSEM identified 774 peaks that are not part of Perm-seq optimal peak list. We evaluated peaks specific to CSEM and Perm-seq to assess whether utilizing the prior information is eliminating false positive and false negative peaks. Fig. S4, Fig. S5, and Fig. S6 in S1 Text display examples of each type of peak. We also provide circos plots of read allocation by Perm-seq and CSEM for a Perm-seq peak with reads distributed over four segmental duplication regions to elucidate how DNase information is guiding read allocation (Fig 2(b)). The three regions depicted in the circos plots span partially overlapping segmental duplications chr1:143,880,003–143,978,943, chr1:206,072,707–206,171,611, and chr1:143,880,003–144,005,301, chr1:120,872,119–249,250,621. Multi-read allocation without prior information distributes the set of multi-reads over these three regions because of the similarities in their local uni-read ChIP counts and fails to identify a peak in any of them. However, as depicted with the DNase-seq track, only one of these three regions has considerable DNase-seq signal indicating regulatory activity. Utilizing DNase-seq prior allocates most of the multi-reads to the region with high DNase-signal and successfully identifies a peak with a canonical Ctcf motif. Segmental duplication regions are defined as regions in which at least 1000 bps of the total sequence (containing at least 500 bps of non-repeat masked sequence) align with a sequence identity of at least 90% (percentage of matching bases out of the aligning bases) [27]. However, high sequence identity does not necessarily imply large fraction of non-unique, i.e., not uniquely mappable, sub-sequences (e.g., 50mers) within the aligning region. Even though ChIP-reads can map to non-unique regions within the segmental duplication, DNase reads can cover more of the uniquely mappable regions and thereby discriminate between two high sequence identity regions. Fig. S7 in S1 Text displays the maximum and the median of the run lengths of adjacent non-unique 50mers across all the segmental duplications as a function of the alignment sequence identity. Overall, only 14.7% of the segmental duplications have a maximum run length larger than 300 bps indicating that for a majority of the segmental duplication regions, there are many mappable positions within and in the immediate vicinity of the 300 bps window a ChIP-seq peak might cover. DNase reads aligning to such bases boost the signal for the ChIP-seq peak and discriminate between highly similar regions.
We further compared the peak sets from Perm-seq and CSEM with sequence analysis. We learnt the most enriched motif in the top 500 peaks of the intersection of the three peak sets (peaks from the uni-read, CSEM, and Perm-seq analyses) using the MEME suite [28] and evaluated the occurrence of this motif in the Perm-seq specific and the CSEM specific peak sets. Fig 2(c) displays motif occurrences in these two types of peak sets across 32 factors in GM12878 and K562 cells, respectively (Table S4 in S1 Text). On average, 13% more of the Perm-seq specific peaks have a motif occurrence compared to CSEM specific peak sets. Furthermore, 13 of the 64 Perm-seq specific peak sets exhibit at least 25% more motif occurrence compared to CSEM specific peak sets, indicating that much larger fractions of the Perm-seq only peak sets are supported by the sequence analysis. Overall, the differences in motif occurrences of the two methods in both cell types are significant with Wilcoxon-rank-sum test (p-values < 0.001). Fig 2(d) extends this comparison on the Ctcf ChIP-seq data from GM12878 to include a Gibbs-based approach [2], which is another multi-read allocation method similar to CSEM, and Lonut [4], which is the most recently published multi-read allocation method. Overall, we observe that Perm-seq reduces numbers of peaks without motif and increases the numbers of peaks with the motif.
After studying the accuracy of Perm-seq for multi-read allocation, we sought to understand the broader impacts of incorporating multi-mapping reads in ChIP-seq analysis. In what follows, the comparisons are focused on (i) peak sets identifiable by both the uni-read and prior-enhanced multi-read analysis with Perm-seq (Common) and are obtained by intersecting optimal peak lists from uni-read and Perm-seq analyses; (ii) peak sets identifiable only with the prior-enhanced multi-read analysis with Perm-seq (Perm-seq-only), i.e., Perm-seq optimal peaks not overlapping the uni-read optimal peaks; and finally, (iii) the subset of the Perm-seq-only peaks that are not overlapping the uni-read relaxed peak sets (Perm-seq-exclusive). Overall, our Common set is highly comparable to the peak sets from the ENCODE project which utilized Bwa [6] instead of Bowtie [23] for aligning the reads.
We first annotated these peak sets in terms of segmental duplication regions (Fig 2(e) and Fig. S11 in S1 Text) and observed that both Perm-seq-only and Perm-seq-exclusive peak sets identify significantly more peaks in segmental duplication regions compared to the Common peak sets. For example, more than 60% (90%) of the Perm-seq-only (Perm-seq-exclusive) Ctcf peaks reside in segmental duplications in contrast to only 2% of the Common Ctcf peaks in segmental duplications. The largest percentage of Common peaks located in segmental duplications is only 3% across all the TFs. In contrast, for all the TFs except for Usf2, Jund, and Max, more than 30% of the Perm-seq-only peaks reside in segmental duplications. Similarly, at least 50% of all but Nfe2 Perm-seq-exclusive peaks overlap with segmental duplications. These results emphasize that characterizing regulatory activity in segmental duplications requires utilization of multi-reads.
Notably, Perm-seq-exclusive peak sets contain many peaks that reside in promoter regions of genes in segmental duplication regions. Our Common peak set has very similar annotation characteristics to the ENCODE peak sets [24] and that Perm-seq-exclusive peak sets enhance the ENCODE peak sets by the addition of many peaks in segmental duplication regions.
In Fig. S12 and Fig. S13 in S1 Text, we compared enrichment of Common and Perm-seq-exclusive peaks for other classes of repetitive elements including short and long interspersed elements (SINE and LINE retrotransposons), DNA transposons, and long terminal repeat elements (LTRs). In contrast to enrichment for segmental duplications, enrichments of peaks for these other repetitive element classes were comparable between the Common and Perm-seq-exclusive peaks for a majority of the TFs.
Next, we explored Pol2 peaks further to assess whether Perm-seq-only peaks contributed to interpretation of RNA-seq results in GM12878 and K562 cells. Fig 2(f) displays transcripts per million (TPM) on the log base 2 scale versus the rank of 3818 genes with TPM greater than 50 either in GM12878 (2957 genes) or K562 (2921 genes) cells. Genes with a Common peak in their promoters (defined as 5000 bps downstream and 500 bps upstream of the transcription start site) are depicted in green, where as genes with a Perm-seq-only peak are depicted in blue. A total of 38 and 58 genes with low mappable promoters had Pol2 peaks only identifiable with the Perm-seq analysis in the GM12878 and K562 cells, respectively. Notably, CCL4, a chemotactic cytokine, is one of such genes which is specifically expressed in GM12878 (TPM values for K562 and GM12878 are 0 and 707.27, respectively, Fig. S14 in S1 Text). It has been recently shown that CCL4 was induced in EBV-infected B cells and was expressed at high levels in all EBV-immortalized lymphoblastoid cell lines [31] (i.e., GM12787).
We considered a subset of the peaks across all TFs that resided in the vicinity of genes (within -10000 bps upstream of the transcription start site and +10000 bps downstream of the transcription end site) and classified these genes into three classes as having (i) Common peaks, (ii) Common and Perm-seq-only peaks, (iii) Perm-seq-only peaks. The mean numbers of genes in each of the three categories were 4226.5, 289.5, and 217.4 for GM12878 and 4589.4, 403.9, and 190.5 for K562. We then compared RNA-seq data for these three classes of genes (Fig 3(a)) and observed that they have comparable expression levels, indicating that genes with only Perm-seq-only peaks are transcriptionally similar to genes that harbor Common peaks. We further evaluated the percentage of genes residing in segmental duplications within each class (Fig 3(b)) and observed that a significantly larger percentage of the genes with a Perm-seq-only peak resided in segmental duplications.
Fig 3(c) depicts (and S2 Text & S3 Text) average histone modification profiles (H3k9ac, H3k27me3, H3k4me1, H3k4me3) for the [-2000 bps, +2000 bps] window anchored at the peak summit for the three groups of peaks. We observe that Peak-seq-only peaks have almost identical profiles to those of the Common peaks. Peak-seq-exclusive peaks, in general, show lower signal for all the modifications consistent with the fact that ChIP-seq signal for these peaks are also lower; however, their overall average modification profiles match closely to those of the Common and Perm-seq-only peaks.
We next sought to understand whether protein-DNA interactions followed similar sequence-specific features in non-repetitive (Common peak sets) and repetitive (Perm-seq-exclusive peak sets) modes of binding. We performed de novo motif analysis of both sets of peaks using the MEME suite [28] (Fig. S15 in S1 Text). De novo analyses on the Perm-seq-exclusive peak sets identified motifs that matched the most significant motifs (denoted as M1 in the MEME analysis) from the de novo motif analyses of the Common peaks for 78.1% and 71% of the factors in GM12878 and K562 cells, respectively. The factors for which M1 motifs of the Common peaks were not identified de novo in the Perm-seq-exclusive sets are Bclaf1, Chd2, Ets1, Nr2c2, Sin3a, Taf1, and Tbp for GM12878 and Bcl3, Bclaf1, Chd2, Ets1, Nr2c2, Pol3rg, Sin3a, Taf1, and Tbp for K562. The majority of these factors are either chromatin modifiers (Chd2, Sin3a), general Pol2 associated factors (Taf1, Tbp), or the sizes of their Perm-seq-exclusive peak sets are smaller than 500 (Nr2c2, Polr3g, Ets1 (GM12878), Bclaf1 (K562), Bcl3). For 65.6% (GM12878) and 53.1% (K562) of the factors, the most significant motifs (M1s) from the Common peak sets were identified as the M1 motif in the Perm-seq-exclusive analysis.
We next compared the occurrences of the M1 motifs of the non-repetitive mode of each factor in the peak sets. We used all the Perm-seq-exclusive peaks and the subset of the Common peaks that were not utilized in the de novo motif analysis, i.e., test peaks from Fig. S15 in S1 Text. Similar results were obtained when Common peaks used in the de novo motif analysis were included (data not shown). Overall, we observed that Perm-seq-exclusive peaks tend to have lower enrichment for the M1 motif (Fig 3(d)). To evaluate whether this was a consequence of the overall lower ChIP signal of the Perm-seq-exclusive peaks compared to Common peaks, we considered the subset of Common peaks that matched the signal strength of the Perm-seq-exclusive peaks. Fig 3(d) indicates that when the signal strength is taken into account, the enrichment of the M1 motif tends to become more comparable in the non-repetitive and repetitive modes. We also compared the information contents of the M1 motifs of the Common peaks with their best matching motifs from the Perm-seq-exclusive peaks (Fig. S16 in S1 Text). This comparison revealed that the overall information contents of the de novo learnt canonical motifs are highly correlated: 0.74 (0.73) when using Common peaks versus 0.80 (0.94) when using the subset of the Common peaks that match the overall signal of the Peak-seq-exclusive peaks in GM12878 (K562) cells. Correlation calculations excluded Zbtb33 for both cell types and Srf for K562 as outliers. For these factors, the motif logos from Perm-seq-exclusive peaks revealed extended versions of the logos from Common peaks and had high information content flanking regions (Fig. S17 and Fig. S18 in S1 Text).
Having established that binding sites from the repetitive mode exhibit similar sequence contents to those of the binding sites in the non-repetitive mode, we evaluated evolutionary conservation of the binding sites from the two groups. Specifically, we took all the best matches to the M1 motif of the Common peak set from both of the Common and Perm-seq-exclusive peak sets and analyzed their phyloP score [32] distributions using the pre-computed phyloP scores from the UCSC Genome Browser. Fig 4(a) compares the phyloP conservation scores averaged over the individual binding sites within each group with an empirical cumulative distribution function plot for transcription factor Usf1. The observed pattern indicates that Perm-seq-exclusive binding sites are overall less conserved. A similar result holds for the majority of the factors (S4 Text & S5 Text). A Wilcoxon-rank sum test for each of the 32 factors revealed that only four (Jund, Spi1, Srf, Tbp) and six (Max, Nfe2, Polr2a, Rest, Sin3a, Spi1) factors do not have significantly different conservation levels (adjusted p-value larger than 0.05 according to the Benjamini-Hocberg false discovery rate (FDR) control [33]) between the Common and Perm-seq-exclusive binding sites in GM12878 and K562 cells, respectively. We next considered position-specific (nucleotide level) phyloP scores. Fig 4(b) displays average position-specific phyloP scores for binding sites of Common and Perm-seq-exclusive peaks of Usf1. Although the overall nucleotide-level conservation scores for the Perm-seq-exclusive sites are lower than those of the Common sites, the patterns of the mean profiles are very similar and highly correlated (Pearson correlations are 0.81 and 0.85 in GM12878 and K562 cells, respectively). Furthermore, nucleotides within the motif (between vertical dashed lines) tend to have higher scores than those in the adjacent non-motif positions for the Perm-seq-exclusive binding sites. We next tested the correlations between average position-specific phyloP profiles of the Common and Perm-seq-exclusive peaks (Fig 4(c)). All but 5 factors (Chd2, Polr3g, Srf, Tbp, Zbtb33) have significantly correlated mean profiles of position-specific phyloP scores between the repetitive and non-repetitive modes of binding in both of the cell lines. Of the five factors with low correlation, Tbp is conserved in moderate levels in both GM12878 and K562 but does not pass the significance cut-off in GM12878 (adjusted p-value of 0.051). Profiles for the other four factors exhibit starker differences for the Common and Perm-seq-exclusive peaks. Neither Chd2 nor Polr3g are sequence-specific factors and their M1 motifs from the Common set are not identified de novo from the analyses of the Perm-seq-exclusive peaks. Serum response factor Srf is a sequence-specific transcription factor. The canonical motif for this factor is identified as the forth significant motif in GM12878 and as the first in K562 Perm-seq-exclusive peak sets. Although the correlations between conservation profiles do not achieve the significance cut-off, they exhibit reasonable correlation (0.41 in K562 with an adjusted p-value of 0.06 and 0.26 in GM12878 with an adjusted p-value of 0.17). Furthermore, there are not notable differences between the canonical motifs identified from the repetitive and non-repetitive regions of binding (Fig. S17 in S1 Text). Zbtb33 is a transcriptional regulator with bimodal DNA-binding specificity and is known to bind to methylated CGCG and the non-methylated consensus site TCCTGCNA [34]. The position-specific conservation profiles between the Common and Perm-seq-exclusive peak sets exhibit significant lack of correlation, especially at the CGCG core (Fig 4(d)). The “G” nucleotide of the first CpG in the core shows accelerated evolution with a negative phyloP score in K562 and the sequence logos of both of the motifs from GM12878 and K562 Perm-seq-exclusive peaks reveal a degenerate position at this location of the “CGCG” core (Fig. S18 in S1 Text). This raises the possibility of Zbtb33 interaction with unmethylated or partially methylated CGCG. To assess this, we calculated the fraction of the Common peaks that exhibited methylation at the CGCG core. Comparisons with the ENCODE Reduced Representation Biosulphite Sequencing (RRBS) data revealed that only 26.3% and 17.2% of CGCGs were methylated in the Common peaks with at least one CGCG core, in GM12878 and K562 cells respectively. This result supports Zbtb33 interaction with unmethylated CGCGs and the possibility of a more degenerate CGCG core for the Zbtb33 motif in the repetitive mode.
The core of the Perm-seq methodology relies on the innovation of incorporating DNase-seq data to derive priors for read allocation probabilities with a Dirichlet-multinomial regression model. This regression framework is highly versatile and can easily accommodate incorporation of other epigenomic data for prior construction. Various histone modifications contribute different putative functions to gene regulation. For example, acetylation of H3k27 and H3k9 are associated with gene activation, trimethylations of H3k27 and H3k9 are linked to repression and H3k4me1, H3k4me2, and H3k4me3 are correlated with functional enhancers/promoters in various cell types [35, 36].
Perm-seq incorporates multiple histone datasets as additional covariates in the Dirichlet-multinomial regression model and employs a variable selection procedure with Group Lasso [37] to select the histones that are most relevant for read allocation (work-flow in Fig. S19 of S1 Text).
We reanalyzed a subset of the ENCODE ChIP-seq datasets (Atf3, Ctcf, Rest, Sin3a, Egr1, Ep300 from GM12878 cells) by utilizing 11 histone ChIP-seq datasets (H2a.z, H3k27ac, H3k27me3, H3k36me3, H3k4me1, H3k4me2, H3k4me3, H3k79me2, H3k9ac, H3k9me3, H4k20me1). In this set, Atf3 and Ctcf function as both transcriptional activators and repressors, Rest and Sin3a are repressors, and Egr1 and Ep300 are activators [38]. Table S6 in S1 Text lists the final set of histone marks that were selected for each TF. We observe that histone variant H2a.z and modifications H3k27ac and H3k27me3 significantly contributed to prior construction for at least five of the six datasets. Overall, inclusion of multiple histone data in prior construction resulted in an increase of numbers of peaks for Atf3, Ctcf, and Ep300 and a decrease for the rest of the factors compared to Perm-seq analysis with only DNase-seq (Table S7 in S1 Text). On average, the change in the numbers of peaks with the inclusion of histone datasets were much smaller (0.99%) compared to the overall change in the numbers of peaks with and without prior information (4.6%) and with and without multi-reads (8%).
We performed detailed analysis of Perm-seq peaks obtained with DNase alone and DNase and Histone ChIP-seq and did not observe significant differences between the motif occurrence percentages in these peak sets. Fig 5(a) displays the DNase and Histone modification profiles of Ctcf peaks in the Perm-seq analysis. Utilizing histone ChIP-seq in addition to DNase-seq eliminates peaks with low DNase-seq read counts and lacking histone support. The majority of Ctcf Perm-seq-specific (DNase) peaks are located within low DNase and low histone signal regions, with an average DNase read count of 18.94 and average histone read count of 5 (Fig 5(a) (upper panel)). Incorporation of histone datasets identifies novel peaks supported by histone profiles. For the Perm-seq-specific (DNase+Histone) peaks, the average DNase read count is 30.18 and average H2a.z count is 19.51. Further investigation of these regions indicate that additional support from the histone profiles enable discrimination between mapping positions with similar DNase-seq profiles. Fig. S20 in S1 Text provides circos plots of read allocation by Perm-seq-specific (DNase) and Perm-seq-specific (DNase+Histone) for a Perm-seq-specific (DNase+Histone) peak with reads distributed over two segmental duplication regions and elucidates how additional histone information is guiding read allocation. The two regions depicted in the circos plots span partially overlapping segmental duplication regions (chr1: 83,647,856–83,955,427 and chr7: 76,280,701–76,575,579). Multi-read allocation by Perm-seq using only DNase data fails to distribute the set of multi-reads to a single region because of low DNase read counts in both of the regions and, hence, does not identify a peak in any of them. However, as depicted with the H2a.z and H3k27ac tracks (the second and third tracks), only one of these two regions has considerable H2a.z, a variant of histone H2a associated with regulatory elements within dynamic chromatin [39], and H3k27ac, mark of active regulatory elements. Utilizing histone information allocates majority of the multi-reads to the region supported by histone information and identifies a peak with a motif.
We next explored Egr1 Perm-seq (DNase) specific peaks. Most of these peaks show both strong DNase and histone signals (Fig. S21 in S1 Text), yet they are not identified by the Perm-seq (DNase+Histone) analysis. We compared the DNase and histone signals of these peak regions with the signal in the next best alternative mapping regions (mapping regions with the second largest allocation weights compared to peaks). In Fig. S21 of S1 Text, we observe histone signal both in the next best alternative mapping regions (left panel) and peak regions (right panel). Furthermore, very few peak regions have DNase and histone signal consistently higher than those of the next best possible mapping regions, leading to Perm-seq analysis with both DNase and Histone to miss these regions.
Understanding the nature of protein-DNA interactions in highly repetitive regions of the genome remains a major challenge in deciphering the human transcriptional regulatory code. To expand the capability of ChIP-seq experiments for profiling repetitive regions of genomes, we developed a prior-enhanced read mapping method named Perm-seq that capitalizes on DNase-seq and histone ChIP-seq data and read counts of local neighbourhoods of mapping locations to discriminate between multiple alignment positions of the same read. Perm-seq analysis of a large collection of ENCODE ChIP-seq datasets from GM12878 and K562 cells revealed that Perm-seq leads to higher sequencing depths and, more importantly, identification of novel targets in segmental duplication regions as well as other repetitive regions. We found that these novel targets, to a large extent, share the same sequence characteristics of the known canonical binding sites of the studied TFs. Conservation analysis indicated that canonical binding sites in these regions are evolutionary less conserved compared to their counterparts in non-repetitive regions; however, the patterns of nucleotide-level conservations correlate well between the repetitive and non-repetitive region binding sites. By integrating ChIP-seq data of Pol2 with RNA-seq data, we found that many highly expressed genes originally thought not have Pol2 occupancy in the promoter regions with uni-read analysis of ChIP-seq data indeed have Pol2 peaks revealed by Perm-seq analysis.
Perm-seq assumes that DNase-seq data harbors information regarding the nature of the protein-DNA interactions and the prior building step learns the specifics of this relationship. For most transcription factors, protein-DNA interactions occur in regions of accessible chromatin. However, a number of proteins, e.g., KRAB-associated factors) are observed to bind to inaccessible chromatin [40]. Exploratory plots such as Fig. S2 of S1 Text indicate whether DNase-seq priors are informative. If this plot does not reveal a systematic relationship between the ChIP and DNase counts, the prior building step generates non-informative priors and, hence, Perm-seq is not expected to improve over methods that do not utilize priors. We also explored the use of multiple histone ChIP-seq datasets in addition to DNase-seq for constructing priors. Although we observed that histone ChIP-seq, in addition to DNase-seq, could further discriminate between the mapping positions, the gain due to using additional histone ChIP-seq was not as large as that of the gain due to using DNase-seq as prior as opposed to no priors. Furthermore, H2a.z ChIP-seq performed as a comparable alternative to DNase-seq for prior construction.
The DNase-seq samples we used have moderate sequencing depths (approximately 61M aligned reads for Huvec, 63M for GM12878, and 49M for K562). We observe that, as expected, the discriminative power of DNase-seq attenuates with lower depths as illustrated in Fig. S23 of S1 Text. Therefore, we recommend prior construction with high quality DNase-seq data with at least 15–20M reads as recommended by the ENCODE consortium [24]. Finally, we focused on read allocation for TF ChIP-seq experiments. The overall framework should also be useful for multi-read allocation of ChIP-experiments for punctuated histone marks.
The current implementation of the Perm-seq pipeline is available at https://github.com/keleslab/permseq.
The initial step of Perm-seq is to align reads to the reference genome with a standard aligner that can report both uni- and multi-mapping reads. We utilized Bowtie [23] for this task and retained reads with at most 2 mismatches and fewer than 100 reported alignments. These initial aligned set of reads were then modelled with the Perm-seq generative model to probabilistically allocate multi-mapping reads to their mapping locations using DNase-seq data as a prior. We next provide mathematical details of the Perm-seq model.
Let N be the number of reads, i.e., sequencing depth, and M be the number of genomic positions. Let i = 1, …, N denote the index for reads, j = 1, …, M for positions. Let Zi = (Zi1, …, ZiM) be the true origin indicator for the i-th read. If the i-th read is generated from j-th position, then Zij = 1 and Zij′ = 0, for j′ ≠ j. We assume that {Zi, i = 1, …, N} are independent and from a multinomial distribution with parameter vector π = (π1, …, πM). Here, π is the density function for generating reads, and specifically, P(Zij = 1) = πj. In this generative model, Zis are not directly observable since multi-reads can align to multiple locations on the reference genome. Hence, Zi are hidden random variables. As a result of alignment, we observe binary vectors Yi = (Yi1, …, YiM) where Yij = 1 if i-th read is aligned to j-th position on the genome and Yij = 0 otherwise. If the i-th read is a uni-read, then ∑ j = 1 M Y i j = 1. If the i-th read aligns to k different positions with k > 1, then ∑ j = 1 M Y i j = k. For both uni- and multi-reads, ∑ j = 1 M Z i j = 1. Defining Li = {j:Zij = 1} as the actual position read i originates from and hij = Pr(Yi ∣ Li = j), we have P r ( Y i , Z i ) = ∏ j = 1 M ( h i j π j ) Z i j. Under the assumption that the true origin Li is one of the observed alignment positions, we have hij = Yij.
To utilize DNase-seq data, we assign a Dirichlet prior distribution to π with the following density function:
f D ( π ; γ ) = Γ ( γ + ) ∏ j = 1 M Γ ( γ j ) ∏ j = 1 M π j γ j - 1 , (1)
where Γ(.) is the Gamma function, γ + = ∑ j = 1 M γ j, and ∑ j = 1 M π j = 1. Then, we define intermediate variables Sj, Sj ≥ 0 as the pseudo counts added to position j, j = 1, ⋯, M and set γj = Sj + 1, which enforces γj ≥ 1 to avoid negative component values in the maximization (M) step in estimation of the allocation probabilities.
DNase-seq data informs the prior parameters with a Dirichlet-multinomial regression by assuming that Sj depends on DNase read counts xj via the following log-linear model with B-splines:
S j ∣ x j ∼ exp [ β 0 + β 1 S P ( x j ) ] , j = 1 , … , M , (2)
where SP(.) is a vector of piece-wise linear B-spline basis functions and β1 are the set of associated parameters. Let x = (x1, …, xM). Then, we can rewrite Eq (1) by conditioning on x as:
f D ( π ; β 0 , β 1 , ∣ x ) = Γ { M + ∑ j = 1 M exp [ β 0 + β 1 S P ( x j ) ] } ∏ j = 1 M Γ { 1 + exp [ β 0 + β 1 S P ( x j ) ] } ∏ j = 1 M π j exp [ β 0 + β 1 S P ( x j ) ] . (3)
When incorporating multiple histones ChIP-seq datasets, we first convert histone read counts into trinary enrichment indicators indicating low, median, and strong signal strengths. Specifically, for a given histone ChIP-seq dataset, positions with read counts smaller than 90-th percentile of the overall read count distribution are set as non-enriched (0), positions with read counts smaller than 99-th percentile but larger than 90-th percentile are set as moderately-enriched (1), and positions with read counts larger than the 99-th percentile are set as enriched (2). Unlike DNase-seq which measures the extent of chromatin accessibility, histone ChIP-seq identifies regions exhibiting histone modifications, i.e., the true underlying states in a histone ChIP-seq experiment is modified versus unmodified as opposed to a continuous scale. The inclusion of a moderately-enriched group is due to the fact that the multi-read only peaks, in general, exhibit lower signal for all the modifications but their average modification profiles match to those of common peaks that can be identified by both uni- and multi-reads as illustrated in Fig 3(c) and S2 Text & S3 Text.
Let Nh denote the number of histone ChIP-seq datasets and define hj = (h1j, …, hNh j), where hij is the dummy variable of i-th histone’s trinary indicator at j-th position. Let h = (h1, …, hM) and β2 be a vector valued parameter associated with the histone ChIP-seq datasets. Then, Eq (3) can be extended as:
f ( π ; β 0 , β 1 , β 2 , x , h ) = Γ { M + ∑ j = 1 M exp [ β 0 + β 1 SP ( x j ) + β 2 h j ] } ∏ j = 1 M Γ { 1 + exp [ β 0 + β 1 SP ( x j ) + β 2 h j ] } ∏ j = 1 M π j exp [ β 0 + β 1 SP ( x j ) + β 2 h j ] . (4)
Theoretically, π, β0, and β1 can be estimated simultaneously via the Expectation(E)-Maximization(M)-Smoothing(S) algorithm [41]. However, since prior parameters β0 and β1 do not have closed form estimators, simultaneous estimation of all three parameters require repeated numerical optimization in the maximization step of the EMS algorithm. To avoid such high computational cost, we first estimate the prior parameters β0 and β1 and then estimate π while keeping these prior parameter estimates fixed. Our data-driven computational experiments indicate that this estimation procedure can reliably estimate the read density (Fig. S10(d) in S1 Text).
When estimating β0 and β1, we first fit the log-linear model (2) using data from positions with only uni-reads with a data aggregation strategy. In the case of uni-reads, Sj can be replaced by the actual ChIP read counts. We group positions 1, …, M with the same DNase read counts and average Sj within each group. When using both DNase and histone data, Sj is averaged across positions with the same DNase read counts and combinatorial histone patterns. To avoid the potential estimation bias from regions with ultra-low DNase read counts, we group positions with 0, 1, 2 DNase read counts together. To further decrease computational complexity, we pool positions with ultra-high DNase read counts (larger than the 99.95-th percentile of the DNase read count distribution) together. Then, we put knots at the 90, 99, and 99.9-th percentiles of the averaged DNase read counts and fit the log linear model using these aggregated data. This set of knot points was well supported by our large scale analysis of ENCODE data. When multiple histone ChIP-seq datasets are available for prior construction, we employ Group Lasso [37] on the aggregated data to select the most relevant histone datasets. Once we obtain the estimates β 0 ^ and β 1 ^, we use the EMS algorithm to estimate π as follows. Let π(t) denote the estimate of π from the t-th iteration,
E-step: Taking expectation of Zij conditional on observed read alignments Y, we obtain
z i j ( t ) ≡ E π ( t ) [ Z i j | Y = y ] = π j ( t ) ∑ j ′ ∈ R i π j ′ ( t ) 1 ( j ∈ R i ) ,
where Ri is the set of positions that the i-th read aligns to and 1(.) is the indicator function. M-step: In the M-step, we obtain an intermediate estimate of πj denoted by μ j ( t + 1 ):
μ j ( t + 1 ) = ∑ i N z i j ( t ) + exp [ β 0 ^ + β 1 ^ S P ( x j ) ] N + ∑ j = 1 M exp [ β 0 ^ + β 1 ^ S P ( x j ) ] . S-step: In the S-step, we smooth μ j ( t + 1 ) with a moving average smoother to obtain π j ( t + 1 ):
π j ( t + 1 ) = 1 2 w + 1 ∑ j ′ = j - w j ′ = j + w μ j ′ ( t + 1 ) ,
where w is the half size of the smoothing window. The moving average smoother was also adapted by CSEM where 2w + 1 was set to the average library size. This choice was well supported by the computational experiments in CSEM. Estimation with the extended histone ChIP-seq datasets follow the same procedure with the density of π given in Eq (4).
All of the data files used in the analysis are listed in Supplementary Tables S8–S10. After processing of all the fastq files with Perm-seq, we assigned each multi-read to its best mapping position if the allocation probability at that position exceeded 0.5, i.e., multi-reads with maximum allocation probability less than 0.5 were discarded. Although this approach is more conservative than utilizing all the multi-mapping reads and all the positions that they are allocated to, it makes the final alignment files compatible with the ENCODE’s ChIP-seq uniform processing pipeline. We used the peak calling and IDR parameters adopted by the ENCODE pipeline and set the IDR to 0.02. This resulted in both optimal peak lists and relaxed peak lists for each factor in both the uni-read and Perm-seq analyses. The Common peak set for each TF is defined as the peaks common to uni-read and Perm-seq optimal peak lists. Perm-seq-only peaks include those that are in the Perm-seq optimal list but not in the optimal list of the uni-read analysis. To identify the set of Perm-seq-only peaks that are extremely unlikely to be identified with the uni-read analysis, we extended the uni-read optimal list with nopt peaks to a total of 5nopt peaks by taking the top 5nopt relaxed peaks. Then, we defined the Perm-seq-exclusive peak sets as the subsets of the Perm-seq-only peak sets that were not part of these relaxed uni-read peak list.
Sequence analysis of the Common and Perm-seq-exclusive peak sets included de novo motif finding with the MEME Suite [28] using ± 50 bps of summit of the top 500 peaks. For each peak set, we identified the top 5 motifs (M1 to M5) with the smallest E-values (Fig. S15 in S1 Text). Then, the rest of the peak sets were scanned with the fimo tool of the MEME Suite using the default p-value cut-off of 1e-4 for the occurrences of the identified motifs within sequences spanning ± 150 bps of the peak summits. For evaluating the degeneracy of the motifs identified from Perm-seq-exclusive peaks, we generated a subset of the Common peaks that matched the Perm-seq-exclusive peaks in signal strength reported by the peak caller SPP. This set included all the Common peaks with peak signal strengths within the minimum and third quantile of the signal strength of the Perm-seq-exclusive peaks.
We manually compared the most significant motif (M1) of the Common peak set with the five motifs (M1–M5) from the Perm-seq-exclusive peak set for each TF. If the M1 motifs of the Common peak sets were not identified in the Perm-seq-exclusive motif sets, we considered the canonical motifs (as defined in literature for each factor) identified de novo in the Common peak sets. The factors for which neither the M1 nor the canonical motif from the Common peak sets were identified the among the M1–M5 of the Perm-seq-exclusive peak sets were excluded from the information content analysis. Then, for each motif pair (M1 or canonical from the Common peak set and the corresponding motif from the Perm-seq-exclusive peak set for each TF), we identified the longest sub-motifs with the highest sequence similarity and extended the flanking bases of the sub-motifs to either the full motif lengths or the first non-degenerate position that did not match between the motifs. The information content of each extended sub-motif was calculated [42] for comparison.
Histone ChIP-seq datasets were processed with CSEM and reads with allocation probability larger than 0.9 were utilized after being extended to the average library size of 200 bps to calculate aggregation profiles. For each genomic coordinate around the ± 2000 bps of the peak summit, the number of extended reads within a 151 bps window were averaged to generate a smooth aggregation profile. Each profile was then normalized to 1 million reads and the profiles for a given peak set were averaged coordinate-wise. The profile plots were generated using the Segvis R package (https://github.com/keleslab/Segvis).
The perm-seq algorithm is implemented as an R package named permseq and is freely available from https://github.com/keleslab/permseq. The raw datafiles listed in Tables S9, S10, and S11 in S1 Text are available at the ENCODE portal of the UCSC Genome Browser (http://hgdownload.cse.ucsc.edu/goldenpath/hg19/encodeDCC/). Perm-seq results are available at ftp://ftp.cs.wisc.edu/pub/users/kelesgroup/encode2-perm-seq-peaks/.
|
10.1371/journal.pgen.1000578 | Positive Epistasis Drives the Acquisition of Multidrug Resistance | The evolution of multiple antibiotic resistance is an increasing global problem. Resistance mutations are known to impair fitness, and the evolution of resistance to multiple drugs depends both on their costs individually and on how they interact—epistasis. Information on the level of epistasis between antibiotic resistance mutations is of key importance to understanding epistasis amongst deleterious alleles, a key theoretical question, and to improving public health measures. Here we show that in an antibiotic-free environment the cost of multiple resistance is smaller than expected, a signature of pervasive positive epistasis among alleles that confer resistance to antibiotics. Competition assays reveal that the cost of resistance to a given antibiotic is dependent on the presence of resistance alleles for other antibiotics. Surprisingly we find that a significant fraction of resistant mutations can be beneficial in certain resistant genetic backgrounds, that some double resistances entail no measurable cost, and that some allelic combinations are hotspots for rapid compensation. These results provide additional insight as to why multi-resistant bacteria are so prevalent and reveal an extra layer of complexity on epistatic patterns previously unrecognized, since it is hidden in genome-wide studies of genetic interactions using gene knockouts.
| Understanding the nature of genetic interactions, known as epistasis, is crucial in biology. The strength and type of epistasis is relevant for the evolution of sex, buffering of genetic variation, speciation, and the topography of fitness landscapes. While epistasis between gene deletions has been the recent focus of research, interactions between randomly selected alleles, which are of the greatest evolutionary interest, have not. We have studied the strength and type of epistasis amongst alleles that confer antibiotic resistance and have found that: in an antibiotic-free environment, the cost of multiple resistance is smaller than expected—a signature of pervasive positive epistasis amongst alleles that confer resistance to antibiotics; epistatic interactions are allele specific; a significant fraction of resistant mutations can be beneficial in certain resistant genetic backgrounds; some double resistances entail no measurable cost; and some allelic combinations are hotspots for rapid compensation. Overall, our findings provide added reasoning as to why multi-resistance is so difficult to eradicate. Importantly, our results of allelic-specific epistasis reveal an extra layer of complexity on epistatic patterns previously unrecognized.
| Epistasis occurs when the phenotypic effect of a mutation in a locus depends on which mutations are present at other loci. When the phenotype of interest is fitness, the existence of such genetic interactions can constrain the course of evolution. The strength and form of epistasis is relevant for the evolution of sex, buffering of genetic variation, speciation and the topography of fitness landscapes. While epistasis between gene deletions [1],[2] has been the focus of recent research, interactions between randomly selected alleles, which are of the greatest evolutionary interest has not [3]. Since mutations that confer antibiotic resistance are known to affect bacterial fitness, levels of epistasis amongst such mutations may determine how multiple resistance evolves. Such knowledge can be used to understand and predict what type of resistance mutations are likely to be segregating in microbial populations [4]. To understand and predict the evolution of multiple resistance it is of key importance to know how the fitness of sensitive and resistance bacteria is affected in different environments, particularly both in the presence and in the absence of drugs. Recent studies have shown that interactions exist amongst pairs of antibiotics, i.e. resistance to one drug affects the action of another drug [5],[6]. In particular the combination of pairs of drugs has been studied and the combinations have been characterized as additive, synergistic, antagonistic or suppressive. Importantly it has been found that in certain drug combinations (suppressive) one of the antibiotics may render the treatment more effective against its resistant mutant than against the wild type [6]. However, we are lacking data on genetic interactions amongst single nucleotide mutations conferring antibiotic resistance in a drug free environment, i.e. the cost of multiple resistance. With the occurring increase in frequency of multiple resistant bacteria and the public health problems associated with it, knowledge on the type and strength of epistasis is of most importance in understanding the evolution of multiple resistance and, ultimately, the planning of new strategies for human intervention.
One can think that a way to halt the spread of resistance to a given antibiotic is to stop the use of that antibiotic. In the absence of the antibiotic for which resistance has been acquired, antibiotic-resistance mutants have a fitness cost when compared to sensitive bacteria [7],[8]. However, when infection is not resolved, a common strategy is to continue treatment with a different antibiotic to which the infecting bacteria are still susceptible. Unfortunately this strategy has led to rapid increase in multiple-drug resistance and not to the loss of resistance to the first treatment, as it would be desired [9]. This raises the possibility that multi-resistant mutations are not independent. If so, when resistance first develops the following question should be asked: if a pathogenic strain is resistant to antibiotic X, which antibiotic should be administered as a second treatment? Clearly, the strategy will depend not only on knowledge about the fitness costs of single resistance mutations but also on the level of genetic interactions – epistasis - between the alleles that underlie those phenotypes. When epistasis exists, it can be positive (antagonistic or alleviating) or negative (synergistic or aggravating). If strong negative epistasis amongst drug resistance alleles is found, then the cost of multiple resistances is high and one can expect multi-drug resistant microbes to be counter selected and disappear very rapidly in the absence of either drug. On the contrary, a much more worrying scenario is the existence of positive epistasis, for it implies that the expected time for elimination of multiple resistance, even if antibiotic pressure is inexistent, will be much longer and multiple resistant bacteria are expected to accumulate in the population.
Here we quantify the degree of genetic interactions on cellular fitness in an antibiotic free environment for point mutations which confer resistance to commonly used antibiotics of three different classes. We have focused on resistances to: (i) the quinolone nalidixic acid, which inhibits DNA replication by binding to DNA gyrase; (ii) rifampicin, which belongs to the rifamycins class of antibiotics that bind to the β-subunit of RNA polymerase thereby inhibiting transcription; and (iii) streptomycin, an aminoglycoside that binds to the ribosome and inhibits elongation of protein synthesis [10]. We find that epistasis is allele specific and that the vast majority of allelic combinations exhibit positive epistasis. The costs of double resistance are therefore smaller than what one would expect if they were independent. Interestingly we found several cases of sign epistasis, which implies that mutations conferring resistance to a new antibiotic are compensatory, i.e. alleviate the cost of resistance present on another locus.
To study the degree of epistatic interactions amongst alleles that confer antibiotic resistance we started by selecting a series of Escherichia coli spontaneous mutants resistant to commonly used antibiotics of 3 different classes: nalidixic acid, rifampicin, and streptomycin (Methods). From a panel of 120 sequenced clones that carry a single nucleotide change, we obtained 19 different classes of clones (Table S1) with spontaneous mutations in gyrA, rpoB and rpsL, which are the correspondent common target genes of resistance to nalidixic acid, rifampicin, and streptomycin, respectively. Resistant bacteria with mutations in the same amino acids as those collected here are segregating in microbial populations [4],[11]. We should notice that our procedure for isolating clones carrying antibiotic resistance requires that viable colonies of resistant bacteria can be formed and detected. Any mutations that can arise and cause very high fitness costs cannot therefore be accounted for in this study. Given that highly deleterious mutations are unlikely to segregate in natural populations [4], genetic interactions amongst such mutations are likely to be of less clinical importance, unless these can be very easily compensated for.
We determined the cost of resistance of each of the clones in the absence of antibiotics. This was performed by measuring relative competitive fitness of the resistant strains against a marked sensitive strain through a competition assay in liquid LB medium without antibiotics [12] (Methods). Figure 1A shows the distribution of fitness costs of each of the mutants. The genotype of each mutation as well as its fitness costs are provided in Table S1. On average the cost of antibiotic resistance was 9%. A Kolmogorov-Smirnov test does not reject the exponential distribution (P = 0.85), a commonly used model to describe the distribution of deleterious mutations in several organisms [13]. Mutations that confer resistance to nalidixic acid had a lower average fitness cost (3%) than streptomycin or rifampicin (13% and 9%, respectively). Similar levels of fitness cost were found previously in resistant mutants of Mycobacterium tuberculosis [4],[14] and E. coli [15],[16].
To study the fitness cost of double resistance and measure epistasis we constructed by P1 transduction all possible pairwise combinations (103) between different resistance alleles. These correspond to 5×11 streptomycin/rifampin, 5×3 streptomycin/nalidixic acid and 11×3 rifampin/nalidixic acid possible combinations. We measured the fitness of the obtained double resistant by competition assays, determined its cost and compared it with the cost that would be predicted if there was no epistasis. Figure 1B shows that the average cost of double resistance is less than twice the average cost of a single resistance (compare with Figure 1A), indicating positive epistasis. Fitness of double resistance was used to calculate pairwise epistasis. Pairwise epistasis, ε, between locus A and B can be measured as follows [17]. If WAB is the fitness of the wildtype, WAb, WaB are the fitnesses of each of the single mutants and Wab that of the double mutant then:When the value of the costs is small then this measure becomes very similar to the difference between the cost of double resistance and the sum of the costs of each resistance.
Strikingly, the majority of mutations show positive epistasis. 68% of the points are above the line in Figure 2A and from the clones that show significant epistasis 15% show negative epistasis and 42% show positive epistasis. The later ones correspond to clones where the cost is less than the sum of the costs of each resistance (see also Figure 2C for the specific combinations of alleles showing significant positive epistasis).
It has been suggested from theoretical modeling of RNA secondary structures and from studies in digital organisms – computer programs that mutate and evolve- [18] that epistasis and the fitness effect of mutations may be correlated, such that mutations with larger effects are more epistatic. In our data we find significant correlations between the strength of epistasis (deviation from zero in absolute value) and the costs of the mutations (Pearson's correlation r = 0.61, P<0.001), which support this theoretical prediction. We also find a marginally significant correlation between the value of epistasis, ε, and the cost of the mutations (Pearson's correlation r = 0.19, P = 0.06).
Figure 2B shows the distribution of the ε values, the median is significantly positive (median = 0.025, Bootstrap 95% CI [0.016; 0.032]), and its value corresponds to about 1/3 of the average cost of each single mutation. From the allelic combinations for which there is significant epistasis (53%), only 27% give rise to negative epistasis, whereas 73% of these combinations result in positive epistasis. Epistatic interactions were observed more frequently between mutations in gyrA and in rpsL, with high frequency of positive epistasis but also extreme cases of negative epistasis (synthetic sub-lethal, Figure 2C). Combination of gyrA and rpoB mutations produced the lowest frequency of negative epistasis (6%), which suggests that the sequential prescription of antibiotics leading to these resistances may easily result in multi-resistance development.
Focusing on the mutations we notice that the interactions are not gene but allele specific. For example R529H and H526L mutations in rpoB notably showed very high negative epistasis frequency, whereas four other mutations (D516V, H526N, I572F and S512F) in the same gene showed no negative epistasis regardless of the combination, and D516V and I572F mutations revealed high frequency of positive epistasis. The latter are potential candidates to segregate in natural populations. The rpoB H526D mutation is a specific example where its epistasis is highly dependent on the particular allele of the second mutation: rpoB H526D interacts positively with rpsL K43N, but negatively with rpsL K43T. Interestingly, the rpoB H526D mutation has been found in multi-drug resistant M. tuberculosis strains and also in this organism its cost was shown to depend on the genetic background [4]. This suggests two important predictions: that we should find resistance alleles with strong positive interaction segregating at higher frequencies (such as rpoB H526D/rpsL K43N) in natural populations and that these genetic interactions should also apply to microorganisms other than E. coli.
Recently it has been shown that an important type of epistasis, which is known as sign epistasis [19], may constrain the evolution of resistance to high penicillin concentrations [20]. Sign epistasis happens when the sign of the fitness effect of a mutation (deleterious or beneficial) is itself epistatic, i.e. sign epistasis exists when a mutation is deleterious on some genetic backgrounds but beneficial on others. This form of epistasis, if common, may give rise to multiple peaks on the fitness landscape. In the context of antibiotic resistance, the implication is simple: if sign epistasis is pervasive then it will be much more difficult to move from a scenario of multiple drug resistance to a scenario where all the bacteria are sensitive, even if antibiotic selection pressure is stopped. Experimental evolution studies have shown that, within hundreds of generations, mutations which compensate for the cost of antibiotic resistance (compensatory) are more likely to occur [16],[21],[22] than revertants [14]. This suggests that sign epistasis might have an important role for the evolution of antibiotic resistance. Within the allelic combinations studied, 12% of our clones showed unexpected sign epistasis between drug resistance alleles. This corresponds to double mutants that have a fitness bigger than the fitness of at least one of the single mutants (Figure 3), and here it means that the mutation conferring resistance to a new antibiotic is beneficial (compensatory) when in a genetic background that contains a mutation conferring resistance to a different antibiotic. This is the worst possible scenario for the host and the best possible for the microbe. Given that a particular mutation was just selected by application of an antibiotic, evolution by natural selection makes it likely that the fixation of a mutation conferring resistance to another antibiotic will occur, even if selective pressure is not applied. Specifically, we find that the same mutation conferring streptomycin resistance (rpsL K88E, Figure 3) can be compensated by different mutations that confer rifampicin resistance showing that rifampicin treatment should be avoided in patients infected with rpsL K88E streptomycin resistant mutants. Given these results, knowledge of both the clinical history of patient antibiotic use as well as the specific genotypes associated with a given resistance is recommended for predicting the optimal clinical outcomes.
Another worrying class of clones that we found corresponds to combination of double resistant mutations that entail no significant cost. These are 6 out of the 103: rpsL K43R/gyrA D87G, rpsL K43R/gyrA D87Y, rpoB H526N/gyrA D87G, rpoB H526N/gyrA D87Y, rpoB D516V/gyrA D87Y, rpsL K43R/rpoB H526N. These mutants have no disadvantage against the wild-type. Although this constitutes a small percentage, it is nevertheless of extreme importance since for these double mutants little, if any, compensation is required for restoring the competitive ability of the wild-type. Given the known association between the cost of resistance mutations measured in the laboratory and the frequency at which they are found in clinical settings has been demonstrated, at least in M. tuberculosis [4], we predict that those combinations of double mutations are the ones which are more likely to be found. Future studies are planned to test this prediction, although it should be noted that the target for resistance may vary between species and environmental conditions [23]. In the opposite extreme we found five clones for which the transduction efficiency was very low (combinations shown in black, Figure 2C) which might correspond to combinations of mutations that must entail high fitness cost.
To make predictions about which resistance alleles are likely to be segregating it is important to study the frequency at which they spontaneously arise. To query which of the P1 transducted mutants are likely to naturally occur and if these also show pervasive epistasis we collected hundreds of spontaneous double resistance mutants. From 289 clones that were sequenced, we obtained 76 different genotypes, whose frequencies are given in Figure 4. We performed a χ2-square test to investigate the effects of genetic background on the spectrum of mutations that spontaneously arise. We observe that there are significant differences between the types of new resistance mutations that appear in certain resistant backgrounds and the wild-type sensitive background (Table S2).
We measured the fitness of each of the different spontaneous double resistant clones and compared it with the fitness in the corresponding clones constructed by P1 transduction. 67 out of 72 spontaneous clones were not different from the P1 constructed mutants (Figure S2), but surprisingly, five double resistant clones (rpoB S531F/gyrA D87Y; rpsL K43T/rpoB H526Y; rpsL K88R/rpoB H526D; rpsL K43R/rpoB S531F; rpoB I572F/rpsL K88R) had a significantly higher fitness (Wilcoxon test, P<0.01). These five spontaneous mutants have a higher fitness because they must have acquired an extra mutation during their isolation which is compensatory. To show that this is in fact the case we measured the fitness of independent clones carrying the same resistance mutations. For three of these haplotypes (rpsL K43T/rpoB H526Y, rpsL K88R/rpoB H526D, rpsL K43R/rpoB S531F) we measured the fitness of spontaneous clones which were obtained applying the reversed antibiotic selection procedure. (i.e. the clones were now isolated by selecting first for rifampicin and secondly for streptomycin). A Wilcoxon-test revealed that the fitness values of these new independent clones were not different from the corresponding P1 clones, showing that the original spontaneous clones carry a compensatory mutation. Given the reduced number of generations in the procedure of generating spontaneous double resistant mutants, observing a compensatory mutation is only probable if such mutation has a very strong effect (sc). This is because only with a large sc it will not be stochastically lost (probability of fixation ∼2sc), and can fix in such short time (time to fixation ∼1/sc) [24]. Indeed for the five mutants mentioned above, the estimated fitness effect sc is very large, on average 0.09 (with the corresponding effects of each compensatory mutation 0.07; 0.13; 0.11; 0.07; 0.06). Adaptive mutations of such strong effect emerging and fixing so rapidly in bacterial populations under such small effective population size, is surprising given previous estimates of effects of beneficial mutations (on average 0.01) [25]. A strong compensatory mutation must also have occurred in the spontaneous clones carrying rpsL K43T/gyrA D87G, rpsL K43N/gyrA D87G, rpsL K43T/gyrA D87Y and rpsL K43N/gyrA D87Y mutations, which were determined as synthetic sub-lethals by P1 transduction.
We calculated the ε values for the 67 spontaneous clones where we do not have evidences for extra compensatory mutations and obtained the same trend as with the double mutants constructed by P1 transduction (58% showed significant epistasis from which 74% showed positive epistasis and 7% that had no significant cost (Figure S3)). These results indicate that positive epistasis is also pervasive in spontaneous mutants supporting the relevance of our results. Additionally, there is evidence from epidemiological studies that in M. tuberculosis environmental resistant isolates more than 96% of the strains resistant to rifampicin have at least one mutation in rpoB, 52 to 59% of the streptomycin resistant strains have mutations in rpsL and 74 to 94% of the strains resistant to ofloxacin or levofloxacin (quinolones like nalidixic acid used here) have mutations in gyrA [26]. Also, the same type of mutations have been isolated in E. coli [27] and in Salmonella enterica where 42% of the isolates showed substitutions in gyrA at position S83 and 35% at position D87 [28], that we show here to exhibit epistasis. Thus we expect the traits observed to be relevant for the evolution of multi-drug resistance acquired during treatment of infectious agents.
Given the importance of epistasis in a variety of biological features such as sex and recombination, buffering of genetic variation, speciation, and the evolution of genetic architecture [29], recent studies have focused on measuring genome-wide levels of epistasis. However, typically only deletions or knockout mutations were studied [1],[2],[30],[31]. Here, albeit focusing in a small number of genes, we measured epistasis on fitness at the scale of the allele given that these may be the most common type of mutations segregating in natural populations and because it also allows us to study mutations in essential genes. Our results not only show the significance of these interactions for the evolution of antibiotic resistance but also reveal an extra layer of complexity on epistatic patterns previously unpredicted, since it is hidden in genome-wide studies of genetic interactions using gene knockouts.
The data obtained here revealed an average level of positive epistasis, which differs strongly from the results of the degree of epistasis among slightly deleterious mutations caused by random transposon insertions in E. coli, where on average no epistasis was found [30]. Positive epistasis has also been found in HIV-1 isolates [32]. Since the data in that study was obtained by sampling mutants from natural populations, comparisons between the results obtained and those found here should be taken carefully. We note nevertheless that some bias towards positive epistasis that may be present in the HIV study is not present in our study, since we constructed all possible pairwise combinations of double mutants. It remains to be explored whether the type of mutations (single nucleotide changes, deletions or tranpositions) can affect the pattern of genetic interaction which can be observed. We predict that it can since we show that the type of interactions is not gene but allele specific. Positive epistasis was also detected when studying interactions between rifampin and streptomycin resistant mutants in Pseudomonas aeruginosa [33]. The pattern in E. coli presented here and the results found in P. aeruginosa (even though having a different genetic basis) [33] indicate that the presence of positive epistasis amongst antibiotic resistance mutations is not species specific. Furthermore epistatic interactions involving fluoroquinolone resistance mutations in gyrA have also been found in Streptococcus pneumoniae [34].
Although we have studied different classes of antibiotics that affect different cell targets, our finding of pervasive epistasis can be reflecting the fact that these target genes are part of the fundamental flow from DNA to RNA to protein, and thus can be considered to be working in the same pathway. Because these affect highly conserved cell processes they should be relevant in many different organisms. An interaction between some of these genes has been described for specific traits, namely propagation of bacteriophage T7 (interaction between rpsL and rpoB) and mitomycin C resistance in E. coli (between rpoB and gyrA) [35],[36].
Our data, both of spontaneous and P1 transducted double resistant clones, also indicates the presence of sign epistasis in the cost of multi-drug resistance involving rifampin, streptomycin and nalidixic acid, that is, a small fraction of double resistant clones showed a higher fitness than at least one of the corresponding single resistant mutants. Sign epistasis implies that the fixation of one mutation (for example by strong selection pressure of a given antibiotic) may alter the adaptive path in both number and type of subsequent beneficial mutations. Sign epistasis was also previously found in the context of resistance to the antibiotic cefotaxime [20]. In this system, of the 120 possible mutational paths from the low resistance to high resistance, only 18 can actually occur due to the occurrence of sign epistasis. Although there are not many examples in the literature [19], this one clearly shows the power of sign epistasis in constraining protein adaptation.
Another interesting example in the context of this work is bacterial adaptation to the cost of resistance through the acquisition of new compensatory mutations. In such an adaptive process it was observed that adaptive mutations which reduce the cost of resistance to streptomycin in E. coli [37] and Salmonella [12] are deleterious in the streptomycin-sensitive background and therefore constitute an example of sign epistasis.
In this work we have determined the fitness effect of mutations in the absence of antibiotics. Future studies should focus on epistatic interactions when bacteria grow in the presence of antibiotics, a condition already shown to be relevant to the evolution of resistance [38],[39].
Our results highlight the importance of determining the costs of single and multiple resistances and, accordingly ordering allele combinations by the degree of epistasis they exhibit. Given that at the present time, infections are likely to be caused by microbes that carry resistance to at least one drug, the strategy expected to give the best outcome is one in which the next drug is the one leading simultaneously to the resistant mutant with the biggest cost and strongest negative epistasis. Another approach to prevent the evolution of multidrug resistance would be to use drug combinations (at certain concentrations) that select for sensitive bacteria. This is a plausible scenario since it has been shown that when competing in the presence of the two drugs, sensitive bacteria outgrow resistant [6]. Our finding of pervasive positive epistasis suggests one possible explanation for the difficulty of eradicating multi-drug resistance in organisms like M. tuberculosis, for which current treatments involve combinations of the same drugs as studied here.
The strains used were Escherichia coli K12 MG1655 and Escherichia coli K12 MG1655 Δara. All clones with antibiotic resistant mutations were derived from the ancestral strain Escherichia coli K12 MG1655. Escherichia coli K12 MG1655 Δara, was used as reference for the competition fitness assay. The two strains are distinguishable by phenotypic difference due to a deletion in the arabinose operon: ara+ and Δara give rise to white and red colonies, respectively, in tetrazolium arabinose (TA) indicator agar [12]. The clones were grown at 37°C on plates containing Luria-Bertani (LB) supplemented with agar and the respective antibiotics. The antibiotic concentrations were 100 µg/ml for rifampicin, 40 µg/ml for nalidixic acid and 100 µg/ml for streptomycin. To estimate fitness costs competitions were performed during 24 hours in 50 ml screw-cap tubes containing 10 ml of LB medium at 37°C, with aeration (orbital shaker at 230 RPM). To estimate the frequency of each strain, in the beginning and by the end of the competition, Tetrazolium Agar (TA) medium containing 1% peptone, 0.1% yeast extract, 0.5% sodium chloride, 1.7% agar, 1% arabinose and 0.005% Tetrazolium chloride was used. All sets of dilutions were done in MgSO4 at a concentration of 0.01 M.
The measure of two-locus epistasis requires the construction of double mutants from single mutants, in order to obtain clones with the same mutations alone and in combination. The antibiotics chosen to isolate the mutants were rifampicin, nalidixic acid and streptomycin. Sets of 40 single clones resistant to each antibiotic were obtained by growing independent cultures of Escherichia coli K12 MG1655, plating in Luria-Bertani (LB) agar medium supplemented with each antibiotic and randomly selecting the clones after 24 hours of incubation at 37°C. Each clone was streak plated, and a single colony was grown in a screw-cap tube with 10 ml of LB medium supplemented with the respective antibiotic and stored in 15% glycerol at −80°C. Generalized transduction of the resistance mutations with bacteriophage P1 was performed as described previously [40]. These mutations were obtained by isolating spontaneous resistant clones and then using these as donor or recipient strains for the construction of the double mutants. Whereas in the great majority of clones the transduction efficiency was high, in 5 combinations of clones it was extremely low. These clones were termed synthetic sub-lethals and are indicated on Figure 2C. For two combinations of double resistance mutations, three independent clones were assayed for fitness. No significant differences were observed (Kruskal-Wallis test) between these clones.
Three single spontaneous clones, resistant to each antibiotic, were used to put back the resistance on the wild-type sensitive background through P1 transduction. We then measured in 5 fold replicate competitions the fitnesses of each spontaneous mutant and the corresponding single P1 transducted clone. Pairwise comparisons, by Wilcoxon test, revealed no significant differences between the single resistance clones constructed in different ways.
We exposed the single resistant clones to a second antibiotic to select for spontaneous mutants resistant to two antibiotics. These were obtained by culturing each clone carrying resistance independently and plating in Petri-dishes with the two antibiotics. Clones with the double resistance were picked randomly. All spontaneous resistant clones were tested for a mutator phenotype and those few with evidence of high mutation rate were disregarded.
The main target genes for resistance to rifampicin, nalidixic acid and streptomycin are rpoB, gyrA and rpsL, respectively. To know the mutations that confer the obtained resistances, each target was amplified and then sequenced. The primers used to amplify the portion of the rpoB gene encoding the main set of mutations conferring resistance to rifampicin were: 5′-CGTCGTATCCGTTCCGTTGG-3′ and 5′-TTCACCCGGATAACATCTCGTC-3′; for gyrA gene, which encodes for the main set of mutations conferring resistance to nalidixic acid, 5′-TACACCGGTCCACATTGAGG-3′ and 5′-TTAATGATTGCCGCCGTCGG-3′; for rpsL gene, that encodes for the mutations conferring streptomycin resistance, 5′-ATGATGGCGGGATCGTTG-3′ and 5′-CTTCCAGTTCAGATTTACC-3′. The same primers were used for sequencing straight from the PCR product.
To measure fitness cost of the resistance mutations a competition assay was done. The resistant mutants were competed against a reference strain, Escherichia coli K12 MG1655 Δara in an antibiotic free environment, in an approximate proportion of 1∶1. To do so, we grew both resistant and reference strains in LB liquid medium for 24 hours at 37°C with aeration. Accurate values of each strain initial ratio were estimated by plating a dilution of the mixture in TA Agar plates. Competitions were performed in 50 ml screw-cap tubes containing 10 ml of LB liquid medium by a period of 24 hours at 37°C with aeration. By the end of this competition process, appropriate dilutions were platted onto TA agar plates to obtain the final ratios of resistants and reference strains. The fitness cost of each mutant strain- i.e the selection coefficient - was estimated as the per generation difference in Malthusian parameters for the resistant strain and the marker strain [12], discounted by the cost of the Δara marker. The fitness cost was estimated as an average of four and five independent competition assays for P1 and spontaneous resistant clones respectively. No correlation was observed between the cost of the resistance mutations and the frequency at which they arose (Figure S1).
Pairwise epistasis, ε, can be measured assuming a multiplicative model in which case: ε = WABWab−WAbWaB, where Wij is the fitness of the clone carrying alleles i and j and capital letters represent the wild-type sensitive alleles. Error (σε) of the value of ε is then estimated by the method of error propagation:Whenever the value of ε was within the error we considered that alleles a and b did not show any significant epistasis (we indicate such combinations as white boxes labeled no epistasis in Figure 2C).
From the distribution of values of ε, provided in Figure 2B, we calculated the median value of ε and its 95% CI by bootstrap where we took 1000 samples. We tested normality of the distribution by a Shapiro-Wilk normality test (P = 0.024), and a Wilcoxon test for the location in zero resulted in a P value of P = 0.0006.
To query about the presence of sign epistasis in the data we made pairwise comparisons between the fitness of each double resistance clone and its corresponding single resistance clones using a Wilcoxon test to assess if the fitness of the double resistant was higher than the fitness of any of the single resistance clones. The P values are indicated in Figure 3 for those combinations that provided significant results, at 5% confidence level, after Bonferroni correction (n = 2 comparisons). For some of the double resistant clones created by P1 transduction that indicated the presence of sign epistasis- see Figure 3- we also obtained spontaneous double resistant mutants, that equally indicated evidence for sign epistasis.
Epistasis is sometimes calculated assuming an additive model [41]: ε = cab−(ca+cb), such that it measures the deviation of the cost of carrying double resistance from the sum of the costs of each resistance. Since the values of the majority of the cost are small, applying the multiplicative or the additive model leads to the same conclusions, i.e. those combinations of alleles that lead to positive (negative) epistasis under the multiplicative model, also lead to positive (negative) epistasis under the additive model. A Kolmogorov-Smirnov comparing the distributions of ε under multiplicative model with ε under the additive model results on P = 0.9. The median value in the distribution of epistasis calculated under the additive model is 0.025 with bootstrap 95% CI [0.015; 0.036] and a Wilcoxon test for location strongly supports the presence of positive epistasis: P = 0.00002. This shows that the same pattern occurs applying either the multiplicative or the additive models.
To perform the statistical analysis we used the free software R: http://www.r-project.org.
|
10.1371/journal.pntd.0000995 | Assessing the Impact of Misclassification Error on an Epidemiological Association between Two Helminthic Infections | Polyparasitism can lead to severe disability in endemic populations. Yet, the association between soil-transmitted helminth (STH) and the cumulative incidence of Schistosoma japonicum infection has not been described. The aim of this work was to quantify the effect of misclassification error, which occurs when less than 100% accurate tests are used, in STH and S. japonicum infection status on the estimation of this association.
Longitudinal data from 2276 participants in 50 villages in Samar province, Philippines treated at baseline for S. japonicum infection and followed for one year, served as the basis for this analysis. Participants provided 1–3 stool samples at baseline and 12 months later (2004–2005) to detect infections with STH and S. japonicum using the Kato-Katz technique. Variation from day-to-day in the excretion of eggs in feces introduces individual variations in the sensitivity and specificity of the Kato-Katz to detect infection. Bayesian logit models were used to take this variation into account and to investigate the impact of misclassification error on the association between these infections. Uniform priors for sensitivity and specificity of the diagnostic test to detect the three STH and S. japonicum were used. All results were adjusted for age, sex, occupation, and village-level clustering. Without correction for misclassification error, the odds ratios (ORs) between hookworm, Ascaris lumbricoides, and Trichuris trichiura, and S. japonicum infections were 1.28 (95% Bayesian credible intervals: 0.93, 1.76), 0.91 (95% BCI: 0.66, 1.26), and 1.11 (95% BCI: 0.80, 1.55), respectively, and 2.13 (95% BCI: 1.16, 4.08), 0.74 (95% BCI: 0.43, 1.25), and 1.32 (95% BCI: 0.80, 2.27), respectively, after correction for misclassification error for both exposure and outcome.
The misclassification bias increased with decreasing test accuracy. Hookworm infection was found to be associated with increased 12-month cumulative incidence of S. japonicum infection after correction for misclassification error. Such important associations might be missed in analyses which do not adjust for misclassification errors.
| Hookworm, roundworm, and whipworm are collectively known as soil-transmitted helminths. These worms are prevalent in most of the developing countries along with another parasitic infection called schistosomiasis. The tests commonly used to detect infection with these worms are less than 100% accurate. This leads to misclassification of infection status since these tests cannot always correctly indentify infection. We conducted an epidemiological study where such a test, the Kato-Katz technique, was used. In our study we tried to show how misclassification error can influence the association between soil-transmitted helminth infection and schistosomiasis in humans. We used a statistical technique to calculate epidemiological measures of association after correcting for the inaccuracy of the test. Our results show that there is a major difference between epidemiological measures of association before and after the correction of the inaccuracy of the test. After correction of the inaccuracy of the test, soil-transmitted helminth infection was found to be associated with increased risk of acquiring schistosomiasis. This has major public health implications since effective control of one worm can lead to reduction in the occurrence of another and help to reduce the overall burden of worm infection in affected regions.
| Polyparasitism is a common feature in parasite endemic regions, which includes most developing countries [1], [2]. High prevalence of co-infection with soil-transmitted helminths (STHs), which include roundworm (Ascaris lumbricoides), whipworm (Trichuris trichiura), and hookworm (Ancylostoma duodenale and Necator americanus), and Schistosoma spp. has been reported [3], [4]. Together, these infections correspond to an estimated 43.5 million disability-adjusted life years (DALYs) lost annually [5], [6]. Schistosomiasis and STH infections are associated with conditions of poverty, such as poor hygiene, lack of safe water, inadequate sanitation and factors such as water management systems, age, gender, and farming related activities [4], [5], [7]–[14].
Laboratory studies suggest that infection with one helminth may influence the outcome of infection with another helminth [15]. Positive cross-sectional correlation and synergism between schistosome and STH infections have been reported [2], [3], [6], [16]–[18]. Immunosuppressive effect of STH has been reported, particularly with hookworm infections [19], [20]. The influence of STH infection on risk of infection with schistosomes has not been epidemiologically investigated. One challenge faced by investigators is the use of a less than perfect diagnostic test. The outcome, exposures, confounding variables, or any combination of these can contain errors [21]–[23]. Error in identification of infection status occurs when the test used to identify the infection is not 100% accurate, or not a ‘gold’ standard test [21], [24], [25].
Schistosoma japonicum and STH infections are most commonly detected by examining a stool sample under the microscope for the presence of parasitic eggs. Variation from day-to-day in the excretion of S. japonicum and STH eggs in human feces has been reported [26]–[29]. Collecting stool samples over consecutive days has been shown to improve the sensitivity of coprological tests like Kato-Katz [29], [30]. However, in practice, an unequal number of stool specimens per subject are collected as it is difficult to collect the desired number of stool samples from each subject. This produces potential complications in diagnosing S. japonicum and STH infections as the sensitivity and specificity of the diagnostic tests vary according to the number of stool samples examined [31], [32].
The purpose of this study was to show the impact of adjusting for misclassification error in estimating the effect of STH infections on the 12-months cumulative incidence of S. japonicum infection. Measuring such impact will contribute to a better understanding of the association between STH and schistosomiasis.
The research was approved by the institutional review board (IRB) of the Brown University in the United States and the IRB of the Research Institute for Tropical Medicine in the Philippines. The data analysis component of the study was reviewed and approved by the University of Oklahoma Health Sciences Center IRB. The chiefs of all villages were asked permission for the village to be included in the study. In addition, all eligible participants were asked for their consent to participate. Only those individuals who provided written informed consent were included. Written informed consent for individuals below 18 years old was obtained and provided by parents or legal guardians.
We used data from a longitudinal study conducted between January 2004 and December 2005 in the province of Samar, the Philippines. The main purpose of the original study was to assess the effect of water and animal management systems on the transmission of S. japonicum infection. The design of the baseline study was described elsewhere [33]. A brief summary is given below.
Seventy-five out of 134 villages endemic for S. japonicum in Samar in 2002 were eligible for participation [33]. The inclusion criteria were safety and accessibility of the field team, location and number of households in each village. Twenty-five primarily rain-fed villages and 25 villages with some form of man-made irrigation system were selected.
Eligible households were those of at least five members and where at least one member was working full time in a rain-fed farm in “rain-fed” villages and at least 50% of the time in a man-made irrigated farm in “irrigated” villages. A maximum of 35 eligible households were randomly selected from each village using the following procedure. A list of 50 random numbers was created (one list per village). Eligible households were allocated consecutive numbers and visited in the order chosen at random. If a household refused to participate, the next available household was asked to participate. When 35 or fewer households were eligible in a village, they were all invited to participate in the study. At most six individuals including at least one full-time rice farmer were selected at random from each household.
An individual-level interview included questions on age, gender, and occupation. Participants were asked to provide one stool sample (morning or first) per day for three consecutive days. Each participant provided between one and three stool samples. If a participant provided a stool sample on one of the three days but was unable for any reason to provide stool samples on other days, that person was still considered as a stool sample provider. Stool envelopes (of wax paper and book paper) with popsicle sticks were distributed to participants a day before the actual stool collection. At least thumb-size stool samples were submitted. Portions from different parts of the stool were taken to fill up the template. Although consistency of the stool sample was not recorded, only pasty to formed stool could be accommodated in the stool envelopes. Stool samples were processed 2–3 h after collection. Two slides were prepared from each stool sample. All slides were placed in a styrofoam box with cold packs inside at the end of each collection day. At the end of each collection week all slides were brought to a designated laboratory and transferred to a refrigerator. The time delay between stool sample processing and microscopic reading associated with day one stool collection (provided by 99.45% of participants) ranged from less than 24 hours to as long as 20 days with a median of 4 days (inter-quartile range: 2–6 days). Stool samples were examined for the presence of eggs of S. japonicum and the three STHs. No distinction between N. americanus and A. duodenale eggs was made, although prior reports from the Philippines found exclusively Necator spp. infections [34]. The Kato-Katz technique was used to detect the helminth eggs in stool samples [35]. The number of eggs per gram of stool (epg) was counted for S. japonicum. Although the eggs of each of the STHs were originally documented qualitatively in five response categories (0, + through ++++), STHs were considered as dichotomous variables (observed infected or uninfected) since the researchers were particularly interested in this association. Also, since the infection of interest of the original study was schistosomiasis, the semi-quantitative ascertainment of STH infection may not have been as accurate as that for schistosomiasis. Laboratory technicians were blinded to the identity of the provider of the stool sample they were preparing and reading and did not know if two stool samples were from the same participant (two consecutive day's sample).
Details about the mass treatment have been published elsewhere [36]. Briefly, following the baseline data and stool collection, all residents who were ≥5 years of age at the time and living in the 50 study villages were offered praziquantel. Praziquantel was administered in two equal split doses to give each individual a total of 60 mg/kg. The split doses were administered 4 hours apart with the first dose usually between 9 am and noon. All participants who provided baseline stool samples had been notified of their test results before treatment was offered. Before mass treatment, community preparation was implemented and an effort was made to ensure all cases found to be positive for S. japonicum were treated. Despite these efforts, the village-level participation proportion varied from 16% to 81% [36]. The parasitological test results were shared with the local ministry of health and the national schistosomiasis control team and it was decided to treat villagers positive to STH at the end of the whole study, that is, after the 12-months follow-up. This approach was approved by both IRBs.
All of the study participants were asked to provide three stool samples over three consecutive days 12 months after the mass treatment. All individuals who provided at least one stool sample were considered as follow-up stool sample providers. Stool samples were processed and examined in the same manner and by the same people as at baseline.
Some of the participants who provided the baseline stool samples did not participate in the mass treatment program. Moreover, not all participants provided stool samples during the follow-up survey. The 12-month cumulative incidence of S. japonicum infection/reinfection following mass treatment can only be calculated among the “at-risk” participants who provided at least one stool sample at baseline and follow-up and received treatment. For the purpose of this study, we assumed 100% efficacy of praziquantel for the treatment of schistosomiasis.
As mentioned earlier, we obtained between one and three stool samples on consecutive days from each participant at baseline and follow-up. This introduces individual variations in the sensitivity and specificity of the Kato-Katz to detect infection. To take this variation into account, and to adjust for the village-level clustering of infection, we used a Bayesian latent class hierarchical cumulative-logit regression model based on a method described by Joseph and others (1995) and adapted to our problem (1, 2, or 3 days of sampling) for S. japonicum in animals and in humans in the Philippines [25], [33], [37], [38].
The probability of any single test being positive is the sum of the probability of a true positive result and the probability of a false positive result. If P is the total probability of a positive test, then, from the properties of diagnostic tests, we have
When there is more than one test per person, the properties of multiple tests can be modeled using probability P as the probability parameter of a binomial distribution, assuming that the tests are independent from each other [37]. In the absence of a ‘gold’ standard test, the true status of each subject is unknown, and hence can be considered as ‘latent data’. According to Bayes' theorem, the joint posterior distribution is proportional to the product of the likelihood function and prior distribution, from which all inferences can be obtained. The posterior distribution is not directly available, but inferences about each parameter are available using a Gibbs sampler algorithm, as has become standard in Bayesian analysis. The unknown true infection status for each subject can be estimated once the sensitivity and specificity have been estimated.
The main outcome of interest here is the probability distribution of the true S. japonicum infection category at follow-up. S. japonicum epg counts were grouped into three categories namely: uninfected (0 epg), light infection (1 to 100 epg) and moderate to heavy infection (over 100 epg) [33]. With a three-category outcome variable, classification errors must be further subdivided. For example, when a participant who is truly negative tests positive, there are two possible errors and 1-specificity or the false positive rate must be divided into light or moderate/heavy misclassification errors. The exposure of interest is the probability distribution of true STH infection status (for a particular STH) classified as positive or negative. Separate models were carried out for each of the three STHs.
Each hierarchical model consists of three levels, as follows: the first level includes one intercept parameter for each village and independent variables for age, sex, occupation, and one of the STHs under study. At the second level of the hierarchical model, the intercept parameters from each of the 50 villages are modeled as a linear regression to account for the clustering of infection within village. At the third level, prior distributions were specified for all parameters. Uniform (uninformative) prior distributions on the range from 0 to 1 (parameters of the beta distribution: α = 1, β = 1) were used for sensitivity and specificity of all three STH infections. For S. japonicum, prior specificity mean (SD) for one stool sample was based on our previous work and set to 94.7% (4.0%), and prior sensitivities (SD) for detecting light infection and moderately to heavy infection were set to 54.1% (10.1%) and 75.3% (15%), respectively [33].
The above model was modified to construct three additional models: one model accounted for misclassification error in outcome but not in exposure, one accounted for misclassification error in exposure but not in outcome, and another one did not account for any misclassification error. For models where misclassification error was not accounted for, an individual with any stool sample positive for a particular STH was considered as infection positive for that STH. For S. japonicum, epg per participant (intensity of infection) was obtained by averaging the epg of all stool samples collected from a participant, which is the most commonly used method for calculating overall epg per participant [1], [39], [40].
We assumed conditional independence between subsequent tests in our model, meaning in practice that when more than one sample was available from a subject, the test results are independent from each other, conditional on the person's true infection status. In other words, the probability of a positive (or negative) test depends only on the true status, and once this true staus is known, does not depend on any test results from other days. This assumption seemed reasonable, and simplifies the statistical model compared to a model that might account for any between-day dependencies.
WinBUGS software (version 1.4.3, MRC Biostatistics Unit, Cambridge, UK) was used to implement the Gibbs sampler algorithm. Posterior medians of random samples derived from marginal posterior densities were used as point estimates, reported with 95% Bayesian credible intervals (BCI). The programs written in WinBUGS are available upon request to the authors.
Of the 5624 individuals who agreed to participate in the study at baseline, 2276 (40.5%) constitute the group “at-risk”. The “at-risk” group and those who were not treated with praziquantel or did not provided any stool sample during the follow-up (“not at-risk” group) are compared in Table 1. A higher proportion of people in the “at-risk” group had a positive schistosomiasis test at baseline (23.5%) as compared to those in the “not at-risk” group (10.5%). Because of this discrepancy, there were more rice farmers in the “at-risk” group than in the ‘not at-risk’ group (50.2% vs. 40.9%), since rice farming is associated with S. japonicum infection. Having been positive at baseline, however, did not have an impact on the probability of providing a stool sample at follow-up among those people who did receive treatment (75.9% vs. 76.1%).
Figure 1 displays the OR estimates for the exposure variable (STH infection) from models with and without correction for misclassification error. The OR estimates (95% BCI) for hookworm infection changed from 1.28 (0.93, 1.76) without any adjustment for misclassification error to 2.13 (1.16, 4.08) when both exposure (hookworm infection) and outcome (S. japonicum infection) were corrected for misclassification error. For A. lumbricoides and T. trichiura, the OR changed from 0.91 (0.66, 1.26) to 0.74 (0.43, 1.25) and 1.11 (0.80, 1.55) to 1.32 (0.80, 2.27), respectively. Correction for misclassification error in either exposure or outcome gave intermediate estimates. However, only adjusting for misclassification error in S. japonicum had a larger impact on the OR estimates and their 95% BCI than only adjusting for the misclassification error in the STH. In general, misclassification error-adjusted estimates were further away from the null value and had wider confidence intervals than non-adjusted estimates. In addition, the impact of adjusting for misclassification error on OR estimates and their 95% BCI was larger for hookworm which had the lowest sensitivity and specificity values.
Table 2 provides OR estimates for covariates from respective STH models, with and without adjustment for misclassification error. For all three STH models, misclassification error-unadjusted OR estimate for >40 year-old individuals (reference: ≤10 years) was approximately 1.5 times that found in the exposure and outcome misclassification-adjusted model. Also, for all three STH models, OR estimates for males (reference: females) from the misclassification error-adjusted model were considerably different from OR estimates found in the unadjusted model. In general, both exposure and outcome misclassification error-adjusted ORs, and only outcome-adjusted ORs were similar whereas misclassification error-unadjusted ORs and only exposure-adjusted ORs were similar. The estimated 95% BCI from models adjusting for misclassification error in the outcome variables, with or without adjustment for misclassification error in the STH, were wider than those from models without adjustment of the outcome variable. Adjusting for misclassification error of STH only did not impact the width of the 95% BCI of the ORs of other variables in the model.
To our knowledge, this is the first longitudinal study to estimate the effect of STH infection on the 12-month risk of S. japonicum infection in a population where both of these infections are endemic. In addition, this study minimizes several potential biases by including adjustment for misclassification error in both dependent and independent variables, varying sensitivity and specificity of both tests depending on the numbers of samples available, accounting for clustering between individuals within villages, and taking care of other possible confounders. The adjusted model suggests that hookworm infection is associated with increased 12-month risk of S. japonicum infection following treatment with praziquantel. The two other STH studied did not have an important effect on the risk of infection with schistosomiasis.
Although our analysis included only about one third of the baseline participants from 50 villages, the longitudinal sample size was large enough for this analysis. When comparing individuals included in and excluded from the analysis, we found more rice farmers in the ‘at-risk’ group than in the ‘not at-risk’ group. This is because more males were treated than females (56.4% vs. 43.6%), and because more rice farmers were infected with S. japonicum at baseline. A larger proportion of individuals infected with S. japonicum at baseline received treatment [36]. However, this did not have an impact on the probability of providing a stool sample at follow-up among those people who did receive treatment. So, the use of the “at-risk” group of participants is unlikely to introduce selection bias and to affect the validity of our estimates.
Our results show that OR estimates for all three STHs are pulled away from the null value when the OR estimates are adjusted for misclassification error. This effect of non-differential misclassification has long been recognized, although this is not always the case when exposure and outcome variables are dependent, a discrete variable assumes more than two values, or there is misclassification error in the confounding variable [21], [23], [41], [42].
The effect of misclassification on the OR estimates of the association between STH and the risk of S. japonicum infection differed for the three STHs under study. The magnitude of impact of misclassification error depends on the sensitivity, specificity, and true prevalence of the variable(s) of interest. The relative change in the OR estimates between the unadjusted model and the model adjusting for misclassification error of STH and S. japonicum was larger for hookworm than the other STHs. This is likely to be due to the considerably lower sensitivity (single stool sample) of the Kato-Katz for hookworm as compared to that for A. lumbricoides and T. trichiura [43].
Two studies have reported estimates of cross-sectional association between hookworm infection and infection by another schistosome species (S. mansoni). Keiser and others (2002) reported an OR of 2.25 (95% CI: 1.31, 3.85) from their study conducted among 325 school children in Côte d'Ivoire [2]. Fleming and others (2006) reported an OR of 2.95 (95% CI: 2.19, 3.98) from a study conducted among 1332 individuals in Brazil [17]. Their results, which did not adjust for misclassification error, could be due to the cross-sectional nature of their study, which could increase the association between the prevalences of hookworm and schistomiasis. It is also possible that the association between hookworm and schistosomiasis is larger for S. mansoni than for S. japonicum or that the Kato-Katz performs better for the diagnosis of S. mansoni, thus reducing the effect of misclassifaction error. Moreover, temporality of the association could not be ascertained because of the cross-sectional design of these studies. Longitudinal design of our study allowed us to assess the impact of hookworm infection on the incidence of schistosomiasis japonica, after adequate adjustment for misclassification error. Even though the OR may overestimate somewhat the relative risk, these measures are likely to be reasonably close in our study since the risk of re-infection was in the order of 13%.
Important changes in OR estimates for other covariates were also observed. The OR estimates for covariates when only S. japonicum data (outcome) were adjusted for misclassification error were very close to the OR when both S. japonicum and STH data were adjusted. In contrast, the OR estimates for covariates when only STH data (exposure) were adjusted for misclassification error were very close to unadjusted OR estimates. This is because the strength of the association between the covariates and S. japonicum infection was considerably larger than the confounding effect of STH infections. Nevertheless, even correction for misclassification error in the outcome variable only was capable of changing estimate of effect of some of the covariates on the risk of S. japonicum infection. This has important implications for the assessment of the confounding effect of these variables and their association with the risk of S. japonicum infection. We also observed wider confidence intervals for all misclassification error-adjusted ORs. This results directly from incorporating uncertainty in estimating infection status [21], [44].
The largest impact of misclassification error was observed for the association between hookworm and S. japonicum, which was negligible in the unadjusted model and important on the adjusted one. Several authors have provided numerical examples in their publications showing larger effects of joint misclassification of both exposure and outcome [22], [41], [45]. For A. lumbricoides and T. trichiura, OR point estimates indicate a negative and a positive relationship, respectively, but of a smaller magnitude.
The efficacy of praziquantel for the treatment of schistosomiasis has been reported to range between 71% and 99% in published literature [46], [47], [48]. However, more recent papers have reported an efficacy of praziquantel for the treatment of schistosomiasis around 96% [46], [47]. The “at-risk” group size is likely to be affected by a lower efficacy as treatment with praziquantel does not completely cure everyone who has the infection. In our study, we assumed 100% efficacy of praziquantel for the treatment of schistosomiasis and decided not to adjust for a lower efficacy of praziquantal. This would have required yet another level of uncertainty for only a small proportion of the population (the efficacy is very high), and is unlikely to have changed our conclusions. Another limitation of this study is that our model assumes conditional independence of test results within each individual given the latent true infection status which is always uncertain. To assess conditional dependence we first have to build a more complex model assuming that there is at least some dependence. This allows examination of the size of the dependence parameter and whether or not its use is meaningful [49]. Exploring such a complex model is beyond the scope of this paper. However, several authors have noted that overlooking conditional dependence does not substantially change parameter estimates [49]–[51]. Our results were adjusted for risk factors most often reported to be associated with schistosomiasis, and often shared with hookworm, such as age, gender, occupation, and the village where people live. Although some additional unmeasured confounding factors may explain the observed association, such factors would need to have a very strong relationship with both hookworm and schistosomiasis to modify our conclusion.
Our data suggest that hookworm infection is associated with increased 12-month cumulative incidence of S. japonicum infection. Such important associations might be missed in analyses which do not adjust for misclassification errors. Our findings have important implications for control of these infections in regions where these worms are co-endemic. Effective control of one helminth can lead to reduction in incidence of another and help to reduce the overall burden of helminthic infection in affected regions.
|
10.1371/journal.ppat.1002988 | Identification of Two Legionella pneumophila Effectors that Manipulate Host Phospholipids Biosynthesis | The intracellular pathogen Legionella pneumophila translocates a large number of effector proteins into host cells via the Icm/Dot type-IVB secretion system. Some of these effectors were shown to cause lethal effect on yeast growth. Here we characterized one such effector (LecE) and identified yeast suppressors that reduced its lethal effect. The LecE lethal effect was found to be suppressed by the over expression of the yeast protein Dgk1 a diacylglycerol (DAG) kinase enzyme and by a deletion of the gene encoding for Pah1 a phosphatidic acid (PA) phosphatase that counteracts the activity of Dgk1. Genetic analysis using yeast deletion mutants, strains expressing relevant yeast genes and point mutations constructed in the Dgk1 and Pah1 conserved domains indicated that LecE functions similarly to the Nem1-Spo7 phosphatase complex that activates Pah1 in yeast. In addition, by using relevant yeast genetic backgrounds we examined several L. pneumophila effectors expected to be involved in phospholipids biosynthesis and identified an effector (LpdA) that contains a phospholipase-D (PLD) domain which caused lethal effect only in a dgk1 deletion mutant of yeast. Additionally, LpdA was found to enhance the lethal effect of LecE in yeast cells, a phenomenon which was found to be dependent on its PLD activity. Furthermore, to determine whether LecE and LpdA affect the levels or distribution of DAG and PA in-vivo in mammalian cells, we utilized fluorescent DAG and PA biosensors and validated the notion that LecE and LpdA affect the in-vivo levels and distribution of DAG and PA, respectively. Finally, we examined the intracellular localization of both LecE and LpdA in human macrophages during L. pneumophila infection and found that both effectors are localized to the bacterial phagosome. Our results suggest that L. pneumophila utilize at least two effectors to manipulate important steps in phospholipids biosynthesis.
| Legionella pneumophila is an intracellular pathogen that causes a severe pneumonia known as Legionnaires' disease. Following infection, the bacteria use a Type-IVB secretion system to translocate multiple effector proteins into macrophages and generate the Legionella-containing vacuole (LCV). The formation of the LCV involves the recruitment of specific bacterial effectors and host cell factors to the LCV as well as changes in its lipids composition. By screening L. pneumophila effectors for yeast growth inhibition, we have identified an effector, named LecE, that strongly inhibits yeast growth. By using yeast genetic tools, we found that LecE activates the yeast lipin homolog – Pah1, an enzyme that catalyzes the conversion of diacylglycerol to phosphatidic acid, these two molecules function as bioactive lipid signaling molecules in eukaryotic cells. In addition, by using yeast deletion mutants in genes relevant to lipids biosynthesis, we have identified another effector, named LpdA, which function as a phospholipase-D enzyme. Both effectors were found to be localized to the LCV during infection. Our results reveal a possible mechanism by which an intravacuolar pathogen might change the lipid composition of the vacuole in which it resides, a process that might lead to the recruitment of specific bacterial and host cell factors to the vacoule.
| Legionella pneumophila, the causative agent of Legionnaires' disease, is an aerobic Gram-negative pathogen that multiplies intracellularly in human phagocytic cells and in freshwater protozoa [1], [2]. The bacteria enter the cells by phagocytosis and reside within a unique phagosome, known as the Legionella containing vacuole (LCV), that grows in size and changes its membrane lipids composition during infection [3]. During the onset of infection, the LCV does not fuse with the host cell lysosomes nor become acidic, but instead the bacteria actively recruit secretory vesicles to the LCV and establish a replication niche [4], [5]. For the formation of the LCV, the bacteria utilize the Icm/Dot type IVB secretion system by which they translocate effector proteins that manipulate host cell processes during infection (for reviews see [6], [7]). A very similar Icm/Dot type IVB secretion system was also found in the obligate intracellular pathogen Coxiella burnetii, the etiological agent of Q-fever [8]–[11]. Similar to L. pneumophila, the Icm/Dot secretion system of C. burnetii was shown to be required for intracellular growth [8]. However, the intracellular lifestyle of these two pathogens is completely different [12], [13].
Currently, about 300 Icm/Dot dependent effectors have been identified in L. pneumophila [6] using a variety of bioinformatics and genetic screens [14]–[19]. Several of the effectors were shown to influence different host cell processes, and some of these processes are targeted by several effectors (for reviews see [7], [20]). Six effectors were found to subvert host cell vesicular trafficking by manipulating the host small GTPase Rab1: SidM/DrrA was shown to recruit Rab1 to the LCV and it activates Rab1 by functioning both as a Rab1-GEF (GDP/GTP exchange factor) and as a Rab1-GDF (GDI [GDP dissociation inhibitor] displacement factor) [21], [22]. SidM/DrrA was also shown to AMPylate Rab1 thus keeping it in its active state on the LCV [23], and the effector SidD was shown to deAMPylate Rab1 and to counteract the AMPylation of SidM/DrrA [24]. In addition, AnkX was shown to phosphocholinate Rab1, thus keeping it in its active state on the LCV [25], and Lem3 was found to dephosphocholinate Rab1 and to counteract the phosphocholination mediated by AnkX [26], [27]. An additional L. pneumophila effector, LidA was reported to bind Rab1 and render it active when bound to GDP or GTP [28], [29] and to tether endoplasmic reticulum (ER) derived vesicles to the LCV [30], while the effector LepB was shown to inactivate Rab1 by functioning as a Rab1-GAP (GTPase activating protein) [22]. Three L. pneumophila effectors (LubX, AnkB and LegU1) have been shown to be involved in ubiquitination of host cell proteins; LubX possesses two eukaryotic U-box domains and it was shown to ubiquitinate the host cell cycle protein Clk1 and the L. pneumophila effector SidH [17], [31]. AnkB possess a eukaryotic F-box domain and it was shown to functionally mimic eukaryotic F-box containing proteins and it exploit the host ubiquitination machinery via the conserved eukaryotic processes of K48-linked polyubiquitination and the proteasome machineries in order to generate free amino-acids for the bacteria [32]–[36]. LegU1 was also shown to mediate the ubiquitination of the host chaperone protein BAT3 involved in the regulation of the ER stress response [37]. Five other L. pneumophila effectors, including Lgt1/2/3, SidI and SidL were shown to target the host translational machinery and block protein synthesis [38]–[40] and two additional effectors, LegK1 and LnaB, were shown to activate the host cell NF-kB pathway [41], [42]. These observations clearly indicate that important host cellular processes are targeted by more than a single effector during L. pneumophila infection.
Beside the effectors described above, several L. pneumophila effectors were shown to manipulate phospholipids. Four L. pneumophila effectors, VipD and its paralogs VpdA, VpdB and VpdC, are homologues to phospholipase A (PLA), patatin-like, enzymes [43], [44]. PLA enzymes hydrolyze the carboxylester bonds at the carbon-1 or carbon-2 positions of phospholipids and generate fatty acids and lysophospholipids [45]. VipD was shown to possess a PLA enzymatic activity in a yeast model [44], VipD, VpdA and VpdC were reported to cause lethal effect on yeast growth when expressed, and VipD and VpdA were shown to cause secretory defects in yeast [15]. Another L. pneumophila effector, LegS2, was shown to act as a sphingosine-1-phosphate lyase (SPL), an enzyme that catalyze the irreversible degradation of sphingosine-1-phosphate, which is an important lipid secondary messenger, to phosphoethanolamine and hexadecanal [46]. Beside the effect on lipid composition, several L. pneumophila effectors (SidC, SidM/DrrA and SdcA) were shown to anchor to the LCV by specific binding to phosphatidylinositol-4 phosphate (PI4P) [47], [48], and other effectors (LidA, SetA and LpnE) were shown to preferentially bind phosphatidylinositol-3 phosphate (PI3P) [47], [49], [50].
Other bacterial pathogens have also been shown to manipulate host cell's phospholipids. Similar to L. pneumophila, Salmonella enterica resides in a unique phagosome known as the Salmonella containing vacuole (SCV) during infection. The S. enterica effector SseJ possesses a PLA and glycerophospholipid-cholesterol-acyltransferase activities. SseJ is localized to the SCV membrane where it esterifies cholesterol in order to promote infection [51], [52]. Another S. enterica effector involved in phospholipids manipulation is SopB (also known as SigD). SopB mediates the accumulation of PI3P on the SCV and affects multiple processes during the course of infection, including bacterial invasion, SCV formation and maturation [53]–[55]. SopB was shown to mediate PI3P accumulation by the recruitment of Rab5 to the SCV. Rab5 in-turn recruits and/or activates Vps34 which is a phosphatidylinositol (PI) 3-kinase that phosphorylates PI to produce PI3P [54]. Another example of phospholipids manipulation by a pathogen was shown in Mycobacterium tuberculosis which also replicates intracellularly in a phagosome [56]. The bacteria secrete the PI phosphatase SapM that specifically dephosphorylates PI3P to PI and lowers the levels of PI3P on the phagosomal membrane, thereby blocking phagosome fusion with late endosomes and lysosomes [57].
To date, L. pneumophila effectors were shown to be involved in the host cell's phospholipids regulation in two main aspects; i. Direct degradation of phospholipids by phospholipases (such as VipD). ii. Anchoring of effectors to the LCV via specific PIs (such as SidM/DrrA). In this work we present a novel strategy used by L. pneumophila to manipulate host cell phosphatidic acid (PA), a main component in the host cell phospholipids biosynthetic pathway. We found that the L. pneumophila effector LecE manipulates the PA biosynthetic pathway by activating the host PA phosphatase protein family which results in the conversion of PA to diacylglycerol (DAG). We also found that another L. pneumophila effector, LpdA, a phospholipase-D (PLD) enzyme, generates PA in mammalian cells and in this way it supplies additional substrate (PA) to the PA phosphatase which is activated by LecE. These findings suggest that L. pneumophila specifically manipulates the phospholipids composition of their phagosome to result in a successful infection.
L. pneumophila and C. burnetii utilize the Icm/Dot type-IVB secretion system to translocate a large number of effector proteins into host cells [6], [58], [59]. The Icm/Dot complex components of these two bacteria are conserved [9] but their intracellular infection process is completely different [7], [60]. Similarly to their intracellular lifestyle, the effector proteins translocated by these two bacteria are different and only few of them share sequence motifs such as ankyrin domains [61]. However, the high conservation of the Icm/Dot system in both bacteria in terms of sequence homology and gene organization might suggest that these bacteria also share similar effector proteins. To test this hypothesis, we performed several genomic searches aiming at the identification of proteins that show a similar phyletic distribution as the Icm/Dot proteins. We searched for genes present in the available Legionella and C. burnetii genomic sequences, discarding genes that are also present in other closely related bacteria such as Escherichia coli and Pseudomonas aeruginosa (e.g. house keeping genes). This analysis resulted with the identification of seven proteins (Table 1). To determine the involvement of these proteins in pathogenesis, we constructed CyaA fusions for the L. pneumophila homologs of these genes and examined them for translocation into host cells and five of them were found to encode for effector proteins (Fig. 1A). We named these proteins Lec for Legionella effectors with homologs in Coxiella (Table 1). During the course of our study, four of the L. pneumophila proteins (lpg1692, lpg1717, lpg2546 and lpg2552) [17], [19] and two of the C. burnetii proteins (CbuA0006 and Cbu_0410) [59], [62] were shown by others to translocate into host cells during infection. With the aim of identifying the function of these genes, they were cloned under the control of the galactose-regulated promoter (GAL1 promoter) and expressed in the yeast Saccharomyces cerevisiae. We found that LecE causes strong lethal effect on yeast growth when ectopically expressed and this lethal effect was found to be more pronounced at 37°C in comparison to 30°C (Fig. 1B and data not shown). At this point we decided to focus on LecE and to explore its function.
The LecE protein (Lpg2552) is 555 amino acids long, and is predicted to contain at least six hydrophobic domains which are most likely associated with membranes after translocation into host cells. To examine the involvement of LecE in L. pneumophila intracellular growth we constructed a deletion substitution mutant in the gene encoding for LecE and examined it for intracellular growth in Acanthamoeba castellanii and HL-60 derived human macrophage. Similarly to most of the L. pneumophila effectors, the deletion of lecE had no effect on the intracellular growth in both hosts (Fig. 2A, and data not shown). In addition, similarly to other L. pneumophila effectors, the translocation signal of LecE was found to be located at the C-terminus, since a CyaA fusion of the 92 C-terminal amino acids of LecE was found to translocate into host cells with a similar efficiency like the full length protein (Fig. 2B).
To identify the cellular target of LecE we decided to use a S. cerevisiae high-copy number genomic library, and look for colonies that grow in the presence of LecE, at 37°C, under inducing conditions (media supplemented with galactose). Several colonies where isolated (see Materials and Methods), and most of them did not produce a full-length LecE (data not shown), however one suppressor colony produced a full length LecE protein and the yeast cells were able to grow under LecE inducing conditions (Sup13 in Fig. 3A). The library plasmid present in this suppressor colony was isolated and reintroduced into a yeast strain containing the galactose inducible lecE gene and similar suppression was obtained. Sequencing of the two edges of the plasmid insert revealed the genomic region responsible for the suppression observed (Fig. 3B). Several subclones that were constructed (Fig. 3B) indicated that the dgk1 gene is the gene responsible for the suppression effect. To further confirm the results obtained, we cloned the dgk1 gene under the GAL1 promoter and both lecE and dgk1 containing plasmids were introduced into yeast. As can be seen in Fig. 3A, Dgk1 over expression showed clear suppression of the lethal effect caused by LecE.
Dgk1 is a diacylglycerol-kinase enzyme that catalyzes the formation of phosphatidic acid (PA) from diacylglycerol (DAG) and counteracts the phosphatase activity of the enzyme Pah1 on PA (Fig. 4A) [63]. The activity of Pah1 has been shown to be dependent on its phosphorylation state, and it was shown to be active when de-phosphorylated [64]. The kinase-cyclin complex Pho85-Pho80 has been shown to phosphorylate Pah1 thus inactivating it [65] and the Nem1-Spo7 phosphatase complex has been shown to dephosphorylate Pah1 and activate it [64]. It is important to note that over expression of Dgk1 was found before as a single suppressor in two screens: i) In a screen aimed at identifying yeast suppressors that can rescue the lethal effect caused by the over expression of Pah1-7P (a constitutively dephosphorylated and therefore active Pah1) [63] and ii) In a screen aimed at identifying yeast suppressors that can rescue the lethal effect caused by the over expression of the yeast Nem1-Spo7 phosphatase complex that dephosphorylates and therefore activates Pah1 [63]. In both screens, Dgk1 over expression suppresses a highly active Pah1 enzyme, what might indicate that this is also the outcome of the over expression of the L. pneumophila effector LecE.
The Dgk1 suppression of the LecE lethal effect can be explained in several ways: i) LecE might inhibit the function of Dgk1, in this case higher levels of Dgk1 will result in some Dgk1 that will be left active in the cells; ii) LecE might directly activate the function of Pah1, in this case higher levels of Dgk1, which performs the opposite enzymatic reaction, will suppress the effect of Pah1 activation by LecE. There are also two indirect ways by which Pah1 might be activated by LecE: iii) LecE might activate the Nem1-Spo7 phosphatase complex, that activates Pah1, and in this way it might activate Pah1 indirectly, and iv) LecE might inhibit the Pah1 kinase-cyclin complex Pho85-Pho80 that inactivates Pah1 and in this way it might activate Pah1 indirectly. v) An additional possibility might be that LecE itself possesses an enzymatic activity like Pah1 (PA phosphatase) and Dgk1 suppresses the effect of LecE simply because it performs the opposite enzymatic reaction. To sort between these possibilities we used several yeast deletion mutants and strains over expressing relevant yeast genes and the results of these analyses are presented in Fig. 4B, C, D, E and Fig. S1.
If LecE inhibits the function of Dgk1 then we would expect that a deletion mutant in dgk1 will be lethal to yeast, however it is known that a deletion in dgk1 is viable and show no yeast growth defects ([66] and Fig. 4B). In addition, when we over expressed LecE in the dgk1 deletion strain the lethal effect of LecE was even stronger in comparison to the effect on wild-type yeast (Fig. 4B) indicating that LecE causes its lethal effect also in the absence of Dgk1, therefore it is not possible that the lethal effect observed in the wild-type strain occurred due to inhibition of Dgk1 activity. Moreover, the result showing that LecE caused a stronger lethal effect in the dgk1 deletion mutant, in comparison to its lethal effect in the wild-type yeast (Fig. 4B), supports the possibility that LecE activates the opposite reaction which is catalyzed by Pah1. If LecE activates the function of Pah1 then its expression in a pah1 deletion mutant is expected to result with suppression of the LecE lethal effect because its target protein will be missing. Thus, LecE was over expressed in a pah1 deletion mutant and the result obtained was very clear, the deletion in the gene encoding for pah1 almost completely eliminated the lethal effect of LecE (Fig. 4C), clearly showing that Pah1 is required in order for LecE to cause its lethal effect on yeast growth. In addition, when LecE and Pah1 were over expressed together the lethal effect of LecE was enhanced, even though Pah1 by itself had no effect on yeast growth (Fig. 4C). The combined results indicate that Pah1 is activated by LecE and that this activation causes the observed LecE lethal effect on yeast growth. The fact that Pah1 was required for LecE to cause its lethal effect on yeast growth also indicates that lecE does not encode for a PA phosphatase enzyme by itself (the Pah1 activity) since in this case the deletion in pah1 should have had no effect on the lethal effect caused by LecE.
In order to test whether LecE directly activates the Pah1 function or indirectly by targeting the Pah1 regulators, the relations between LecE and the Nem1-Spo7 phosphatase complex that activates Pah1 and the kinase-cyclin complex Pho85-Pho80 that inactivates Pah1, were examined. To examine if LecE activates the Nem1-Spo7 phosphatase complex, LecE was over expressed in the nem1 deletion mutant (nem1 encodes for the catalytic subunit of the phosphatase complex) or together with the Nem1-Spo7 phosphatase complex. As shown in Fig. 4D, a deletion in nem1 weakly suppressed the lethal effect caused by LecE while the over expression of LecE together with the Nem1-Spo7 phosphatase complex enhanced the lethal effect compared to LecE or the phosphatase complex by themselves. Both results indicate that the Nem1-Spo7 phosphatase complex is not targeted by LecE, but that both LecE and the Nem1-Spo7 phosphatase complex perform a similar function that results with the activation of Pah1 (see below). According to this hypothesis, when nem1 is missing some of the Pah1 protein remains inactive and therefore a weak suppression effect was observed, while when LecE was over expressed together with the Nem1-Spo7 phosphatase complex they both activate Pah1, resulting with an enhanced lethal effect. An additional way for LecE to indirectly activate Pah1 is to inhibit the function of the Pah1 kinase-cyclin complex Pho85-Pho80 that was shown to phosphorylate Pah1 and in this way inactivate it [65]. To examine this possibility, LecE was over expressed together with the Pho85-Pho80 kinase-cyclin complex or in the pho80 deletion mutant. As shown in Fig. 4E, the over expression of the Pho85-Pho80 kinase-cyclin complex completely suppressed the LecE lethal effect on yeast growth while in the pho80 deletion mutant the lethal effect of LecE was enhanced (comparable results were obtained when LecE was over expressed in the pho85 deletion mutant, data not shown). The enhanced lethality of LecE in the pho80 and pho85 deletion mutants indicates that the Pho85-Pho80 kinase-cyclin complex is not targeted by LecE. In addition, the suppression of the LecE lethal effect by the Pho85-Pho80 kinase-cyclin complex indicates that its function is opposite to the one of LecE. In conclusion, the analyses performed in the yeast system strongly indicate that LecE directly activates Pah1.
To further validate the possibility that LecE functions similarly to the Nem1-Spo7 phosphatase complex, we directly compared the effect of LecE and the Nem1-Spo7 phosphatase complex on yeast growth (Fig. 5). We found that over expression of LecE or the Nem1-Spo7 phosphatase complex are both lethal to yeast growth and their lethal effect was suppressed by over expression of Dgk1 (Fig. 5A) and by a deletion of the gene encoding for Pah1 (Fig. 5B). In addition, to test whether the Pho85-Pho80 kinase-cyclin complex also suppresses the lethal effect of the Nem1-Spo7 phosphatase complex on yeast growth, a different yeast strain (W303) that allowed the introduction of four plasmids, was used. Since this strain grows slowly at 37°C it was incubated at 30°C where the LecE lethal effect was less pronounced (see above). Similarly to the over expression of Dgk1 and the deletion of pah1, the over expression of the Pho85-Pho80 kinase-cyclin complex also suppressed the lethal effect on yeast growth of both LecE and the Nem1-Spo7 phosphatase complex (Fig. 5C). These results indicate that LecE directly activates Pah1 similarly to the Nem1-Spo7 phosphatase complex.
As indicated above, the S. cerevisiae dgk1 gene encodes for a DAG kinase enzyme that catalyzes the formation of PA from DAG. Unlike the DAG kinases from bacteria, plants, and animals, the yeast enzyme utilizes CTP, instead of ATP, as the phosphate donor in the reaction [67]. Point mutations of conserved residues within the Dgk1 CTP transferase domain were shown before to result in a loss of DAG kinase activity [67]. To determine if the enzymatic activity of Dgk1 is required for the suppression of the lethal effect caused by LecE, we generated two point mutations (R76A and D177A) in Dgk1 that were shown before to abolish the DAG kinase activity [67] (Fig. 6A). As can be seen in Fig. 6B, both mutated Dgk1 proteins were unable to suppress the lethal effect of LecE in comparison to the wild-type Dgk1, indicating that an enzymatically active Dgk1 is required for suppression. The same result was also obtained for the lethal effect caused by the over expression of the Nem1-Spo7 phosphatase complex (Fig. 6C), further demonstrating the similar function of LecE and the Nem1-Spo7 complex.
The S. cerevisiae Pah1 belongs to a highly conserved family of proteins, called lipins. This novel family of Mg+2-dependent PA-phosphatase enzymes catalyze a fundamental reaction in lipid biosynthesis, namely the dephosphorylation of PA to DAG. Lipins are highly conserved throughout the eukaryotic kingdom and exhibit similar overall primary organization [68]. They are relatively large proteins (close to 100 kDa) and contain a conserved amino-terminal domain (N-LIP) of unknown function, and a carboxy-terminal catalytic domain (C-LIP) harboring an invariable HAD-like phosphatase motif, the DXDXT motif [68]–[70].
To determine if the enzymatic activity of Pah1 is required for LecE to cause its lethal effect on yeast growth, we generated a point mutation (D398E) in the conserved DXDXT motif of Pah1 (Fig. 7A) that was shown before to be critical for the PA phosphatase activity of Pah1 [71]. To determine the outcome of this mutation on yeast cells in relation to LecE, we first constructed an HA-tagged wild-type Pah1 and introduced it into yeast containing a deletion in the pah1 gene and LecE. The introduction of the HA-tagged Pah1 restored the lethal effect of LecE on yeast cells (Fig. 7B), however when the mutated HA-tagged Pah1 (D398E) was introduced instead of the wild-type Pah1 protein the lethal effect of LecE was not restored, indicating that an enzymatically active Pah1 is required to be present in the yeast cells in order for LecE to cause its lethal effect (both the wild-type and mutated HA-tagged Pah1 proteins were expressed in the yeast cells examined, Fig. 7D). Like in the case of the mutated Dgk1, a similar result to the one obtained with LecE was also obtained with the over expression of the Nem1-Spo7 phosphatase complex (Fig. 7C), further demonstrating the similar function of LecE and this complex.
The results described thus far, clearly demonstrate that LecE requires the presence of an enzymatically active Pah1 protein in the yeast cells in order to cause its lethal effect on yeast growth, and this requirement is identical to the one of the Nem1-Spo7 phosphatase complex. However, the mechanisms of action by which effectors activate host cell factors are often different than the ways by which these host factors are activated naturally (see Introduction). To further determine the mechanism of activation of Pah1 by LecE, we examined the size of the Pah1 protein in yeast cells over expressing the LecE effector in comparison to the over expression of the Nem1-Spo7 phosphatase complex. Western analysis showed a clear reduction in the size of the Pah1 protein when the Nem1-Spo7 phosphatase complex was over expressed in yeast but no change in the apparent molecular weight of Pah1 was observed when LecE was over expressed (Fig. S2). These results indicate that LecE activates Pah1 in a different way than the Nem1-Spo7 phosphatase complex and it does not function as a phosphatase of Pah1.
It was shown recently that sometimes several L. pneumophila effectors affect the same host cell processes during infection (see Introduction). To determine if there are additional effectors that affect PA and DAG levels, we examined seven additional effectors (Table 2) that according to their sequence homology and/or sequence motifs are expected to be involved with or were shown to function in phospholipids biosynthesis [44], [46]. We reasoned that yeast deletion mutants in specific host factors (such as dgk1 and pah1) can be used in order to uncover additional effectors that target the same cellular process (for example, other effectors that caused lethal effect on yeast growth might be suppressed by the same yeast strains). Moreover, effectors that originally show no lethal effect on wild-type yeast might cause lethal effect when they will be over expressed in the relevant yeast deletion mutants. Such a result might reveal effectors that target the same cellular process (both effectors might activate or one of them might activate and the other inhibit the same process), during L. pneumophila infection. For this purpose we cloned the seven effectors listed in Table 2 under the control of the galactose-regulated promoter (GAL1 promoter) and expressed them in wild-type yeast (Fig. 8 and Fig. S3). Three of these effectors (VipD, VpdA and VpdB) caused strong lethal effect on yeast growth and one effector (LegS2) caused a moderate lethal effect on yeast growth when expressed in wild-type yeast and they were not suppressed by the deletions in dgk1 or pah1. However, interestingly, when lpg1888 (an effector containing a PLD domain, that we named LpdA, see below) was expressed in wild-type yeast no lethal effect was observed, but when it was expressed in the dgk1 deletion mutant clear lethal effect was observed, indicating that this specific yeast genetic background exposed the function of the effector.
Eukaryotic enzymes containing a PLD domain where shown before to convert phosphatidylcholine (PC) to PA and free choline [72], [73], and the yeast Spo14 is a known PLD enzyme (Fig. 4A). Thus the results obtained with LpdA can be explained in the sense that in the absence of Dgk1 there is no enzyme that can phosphorylate DAG back to PA and under these conditions the activity of LpdA was observed.
LpdA was shown before to translocate into host cell, as part of a large screen, and its translocation level was very low (only 5% of the cells show indication for translocation) [19]. Therefore, we fused LpdA to the CyaA reporter and examined its translocation into host cell (Fig. 9A). Our analysis confirms that LpdA translocates into host cells, its translocation levels were low in comparison to the other effectors examined in this study (Fig. 1A), but no translocation was observed from an Icm/Dot deletion mutant (Fig. 9A). To investigate the relations between LecE and LpdA we constructed a single deletion mutant in lpdA as well as a double deletion mutant of lecE and lpdA and examined the intracellular multiplication of these mutants in A. castellanii. As can be seen in Fig. 9B, no intracellular growth phenotype was observed for the single or double deletion mutants, as was shown before for most of the deletion mutants in L. pneumophila effectors.
To further explore the relations between LpdA and LecE we expressed both proteins together in yeast. This analysis resulted with an additive effect on yeast growth, both effectors together were more lethal to yeast in comparison to LecE by itself (Fig. 9C), indicating that both effectors function in the same direction (LpdA by itself caused no yeast growth defect (Fig. 9C)). Our results reveal two conditions under which LpdA lethal effect on yeast growth can be observed: i) in a dgk1 deletion mutant (Fig. 8) and ii) when LecE was expressed in the yeast cells (Fig. 9C). Importantly, both these conditions have the same outcome on the yeast cell since in the first condition the yeast cell cannot convert DAG into PA and therefore DAG probably accumulates in the yeast cell. In the second condition there is high activity of Pah1 due to the expression of LecE that also leads to the accumulation of DAG. Thus, the results obtained with LpdA further supports the function of LecE as a Pah1 activator.
LpdA was suggested to encode for a phospholipase-D due to sequence homology to eukaryotic (fungal) PLD enzymes. The PLD protein family is conserved from yeast to human and it comprises a conserved catalytic core (HxK(x)4D) [74]. To determine if LpdA encodes a functional PLD enzyme we generated two point mutations (K165R and K376R) in two conserved lysine residues located in both predicted PLD conserved catalytic cores (Fig. 10A). We then used the LpdA lethal effect observed in the yeast dgk1 deletion mutant (Fig. 8) in order to examine these two mutants. As can be seen in Fig. 10B, over expression of the wild-type LpdA in the dgk1 deletion mutant caused lethal effect on yeast growth and this effect disappeared when the two LpdA mutants were used, and the yeast growth with these two mutants was similar to the one of the empty vector. These mutations did not detectably affect the stability of LpdA in yeast (Fig. 10C), suggesting that the loss of toxicity was very likely due to the abolishment of the enzymatic activity of LpdA. Due to these results Lpg1888 was named LpdA for Legionella Phospholipase D.
As indicated above, LpdA enhances the lethal effect caused by LecE on yeast cells (Fig. 9C). To determine if this enhancement also requires the PLD activity of LpdA, LecE was expressed together with LpdA and it's two mutants (K165R and K376R) in yeast cells. As can be seen in Fig. 10D, the enhancement of the LecE lethal effect by LpdA requires its PLD activity, and the mutations in the PLD active site almost eliminated the enhancement of the lethal effect caused by LpdA. Also in this analysis the two mutations did not detectably affect the stability of LpdA (Fig. 10E).
The results obtained from the yeast analysis of LecE indicated that the function of this effector probably results in an increase in DAG levels in cells (Fig. 4). To determine if LecE affects DAG levels in-vivo in mammalian cells a system based on a DAG fluorescence biosensor was employed using live-cell imaging. The LecE effector was fused to the mCherry fluorescent protein (Cherry-LecE) and was ectopically expressed in COS7 cells together with a PKC-C1-DAG binding domain fused to GFP (GFP-DAG) that was validated before as a specific DAG sensor in several systems [75]–[78]. When the GFP-DAG sensor was expressed in COS7 cells it exhibited two localization patterns: in 59% of the cells the sensor was diffusely distributed throughout the cell but was also concentrated in a membranal peripheral nucleus area, while in 41% of the cells the GFP-DAG sensor showed a completely diffuse distribution (Fig. 11A, B). In contrast, when the GFP-DAG sensor was expressed together with Cherry-LecE its distribution changed and in 88% of the cells the GFP-DAG sensor was mostly concentrated in the membranal peripheral nucleus area (Fig. 11B). Moreover, a similar intracellular distribution was also obtained for Cherry-LecE (Fig. 11C). Importantly, the Cherry-LecE induced changes of the GFP-DAG sensor was significant (p<value 0.007, Student's t-test; Fig. 11B). In addition, the effect of Cherry-LecE on the accumulation of the GFP-DAG sensor in the peripheral nucleus area was examined. In this analysis, the GFP-DAG sensor concentration at the peripheral nucleus area was significantly enriched in cells expressing Cherry-LecE in comparison to cells expressing the GFP-DAG sensor by itself, (2.35 fold, p<value 1.2×10−10 in Student's t-test; Fig. 11D). As a control, GFP was expressed in the presence or absence of Cherry-LecE, and no alterations to the mixed cytosolic and nuclear distribution of GFP were observed upon Cherry-LecE co-expression (Fig. 11E and data not shown). This result further supports the conclusion that Cherry-LecE influences the distribution of the GFP-DAG sensor specifically. In addition, the specific localization of Cherry-LecE was examined and it was found to be localized to the cis-Golgi apparatus as it co-localized with GFP-KDEL-Receptor (GFP-KDELR) a well established cis-Golgi marker (Fig. 11F) [79]. Notably, a previous work done with a different GFP-DAG sensor (based on the PKD-C1-DAG binding domain) found it to be localized to the Golgi in HeLa cells [80]. The combined results presented demonstrate that LecE induces alternations in DAG content in COS7 cells and show co-localization of the ectopically expressed effector and its lipid product to the same sub-cellular compartment, the cis-Golgi.
An analogous approach to the one described above was also applied to address the functionality of LpdA in mammalian cells and its ability to influence the levels and distribution of PA. For that purpose, LpdA was fused to the mCherry fluorescent protein (Cherry-LpdA) and ectopically expressed in COS7 cells together with GFP fused to a PA-binding domain from the yeast Spo20 SNARE protein (GFP-PA). This domain was previously shown to function as a sensitive and specific PA sensor in mammalian cells [81]. As shown in Fig. 12A (on the left), when the PA sensor was expressed by itself it accumulated in the cell nucleus. Several studies have shown before a similar accumulation of the GFP-PA sensor in resting cells, and it was found to be not specific [81], [82]. In striking contrast, when Cherry-LpdA was expressed together with GFP-PA sensor it induced a punctuate distribution of GFP-PA throughout the cell's cytoplasm (Fig. 12B), suggesting an effector-dependent generation of PA. Of note, the GFP-PA-labeled structures were highly mobile, resembling intracellular vesicles. The Cherry-LpdA effector itself showed a diffuse pattern in the cells with some punctuate distribution as well (Fig. 12B). Importantly, when the GFP-PA sensor was expressed together with the LpdA PLD mutant, Cherry-LpdA-K165R, no change in the distribution of the GFP-PA sensor was observed (Fig. 12C), what indicates that the PA production in the cells depended on the PLD activity of the Cherry-LpdA. In addition, it was demonstrated before that PA is usually dephosphorylated to DAG in-vivo [83], [84]; thus, we examined the distribution of the GFP-DAG sensor (described in the previous section) when co-expressed with Cherry-LpdA, and found that it was also localized to motile puncta (Fig. 12D), in sharp contrast to its distribution pattern when expressed alone (Fig. 11A and Fig. 12A on the left). Importantly, when GFP was expressed with or without Cherry-LpdA it showed a diffuse distribution in the cells with some concentration in the cell nucleus (Fig. 12E and data not shown). The combined results presented indicate that the changes observed with LpdA were specific to the PA and DAG sensors. These results substantiate LpdA as a PLD enzyme in the cells, where it generates PA which is further converted to DAG.
To determine where in the host cell LecE and LpdA perform their function during L. pneumophila infection, we constructed plasmids that over express these effectors in L. pneumophila as a fusion to a myc-tag at their N-terminus, and infected U937-derived human macrophages with a wild-type L. pneumophila containing these plasmids and used confocal fluorescence microscopy to visualize the two effectors during infection. As can be seen in Fig. 13, both effectors were found to be localized to the LCV during infection. Only intracellular bacteria show a signal with the anti-myc antibody directed against the effectors.
Thus we conclude that LecE and LpdA are both localized to the LCV, where they probably manipulate the phagosome phospholipids composition during infection.
Up to date about 300 effector proteins were identified in L. pneumophila and the function of only several of them was uncovered. Effectors were found to affect diverse host cell processes which include vesicular trafficking, apoptosis, ubiquitination, translation and others [7], [85]. In several cases, pairs of effectors were found to function together and one effector was found to counteract the function of another effector. The SidM/DrrA effector was found to AMPylate the host cell small GTPase binding protein Rab1, thus keeping it in an active state which cannot be inactivated by host cell factors [23], and the effector SidD was found to reverse this modification by deAMPylation of Rab1 [24], [86]. Another pair of effectors also involved in Rab1 activation was described recently - AnkX and Lem3. AnkX was found to phosphocholinate Rab1 and Lem3 was found to reverse this modification [25]–[27]. An additional effector that was found to affect another effector is LubX. LubX contains an E3 ubiquitin ligase domain and it was found to specifically target the bacterial effector protein SidH for degradation by the host cell proteasome [31], thus affecting the time during infection when SidH is present in the host cell and performs its function. In this manuscript, we described a new pair of effectors that might function together – LecE and LpdA. This pair of effectors is different from the three pairs described above since both effectors function in the same direction and do not counteract the function of one another. The effector protein LpdA was found to contain a functional PLD domain and these enzymes were shown before to convert PC to PA and free choline [87]. The second effector – LecE was found to activate the yeast lipin homolog (Pah1) which converts PA to DAG (Fig. 14). Both these effectors were found to be localized to the LCV during infection thus the combined lipid biosynthetic reactions that might occur on the LCV will include conversion of PC into PA (by LpdA) and then conversion of PA to DAG (by LecE activated PA phosphatase) a process which is expected to result in changes of the lipid composition of the LCV that can affect its fate in the host cell as well as the host proteins and bacterial effectors that will be recruited to the LCV (see below). These results indicate that pairs or groups of L. pneumophila effectors function together and additional such effectors are expected to be found.
The way by which LecE activates Pah1 is currently not known. The natural activation of Pah1 in yeast occurs via dephosphorylation, but our results indicate that this is not the way by which LecE activates Pah1 (Fig. S2). An important result regarding the mode of activation by LecE comes from the finding the over-expression of Pho80-Pho85 in yeast suppresses the lethal effect caused by LecE. This result suggests that the activity of Pho80-Pho85 is dominant on the activity of LecE, therefore it might be that LecE cannot perform its function when Pah1 is fully phosphorylated (the expected state of Pah1 after Pho80-Pho85 over-expression). We hypothesize that LecE activates Pah1 by modifying one of its amino acids (as was shown for SidM and AnkX in the case of Rab1) or by directly binding to it.
Identification of pairs or groups of effectors that influence the same or related host cell processes is very important for the ability to understand the function of the enormous number of effectors translocated by L. pneumophila during infection. The approach that we used in this study, which led to the identification of LpdA as an effector that function with LecE, can help to discover such pairs and/or groups of effectors. Our approach takes advantage of yeast genetics as a tool to identify such groups of effectors. This approach can be applied in a very broad way in order to study effector proteins. When an effector that causes lethal effect on yeast growth is found and a yeast suppressor is identified, other effectors that cause yeast lethal effect and might be suppressed by the same yeast suppressor can be identified in case that they affect the same host factor in a similar way (activation or inactivation). For example the lethal effect on yeast growth caused by AnkX was found to be completely suppressed by over expression of the yeast Ypt1 protein (the yeast homolog of Rab1) [27], it is possible that other effectors that cause lethal effect on yeast growth will be suppressed by over expression of Ypt1, thus leading to the identification of the cellular process they affect. An even more interesting situation is the one described in this manuscript. LpdA causes no lethal effect on wild-type yeast, but when it was expressed in a yeast dgk1 deletion mutant clear lethal effect was observed. In this way, not only effectors that cause lethal effect on wild-type yeast can be sorted into functional groups but also effectors that cause no lethal effect on yeast growth can be sorted, since their effect can be uncovered by using different yeast genetic backgrounds. In the case of LecE and LpdA, over expression of Dgk1 suppresses the lethal effect of LecE and a deletion of dgk1 uncovered the lethal effect of LpdA thus indicating that both effectors function in the same direction. Our approach can also be expanded to other host cell processes expected to be affected by L. pneumophila effectors (or any other pathogens). For example, yeast deletion mutants or strains over expressing genes related to trafficking (such as vps) or authophagy (such as apg) can be used to screen the collection of L. pneumophila effectors, both the ones that cause lethal effect on wild-type yeast as well as these that have no effect on wild-type yeast growth. In this way pairs or groups of effectors that affect similar host cell processes can be uncovered.
The results presented in this study uncover another aspect of the involvement of phospholipids in L. pneumophila infection of host cells. It was shown before that several L. pneumophila effectors (SidC, SidM/DrrA and SdcA) specifically bind PI4P on the LCV [47], [48], [88]. However, it is known that PC constitutes the major phospholipid in eukaryotic membranes [89]. The results presented in this study show that the combined activity of the LpdA and LecE effectors is expected to result in the conversion of PC to DAG on the LCV. In addition, it was shown before that the presence of PI4P on the LCV is strongly dependent on the activity of the enzyme PI 4-kinase IIIβ (PI4KIIIβ) that converts PI into PI4P [47]. One way to recruit PI4KIIIβ to the LCV is by the activity of Arf1 [90], however it was shown before that RalF that recruits Arf1 to the LCV is not required for SidC decoration of the LCV [47]. Another, major way to recruit PI4KIIIβ to membranes is by the action of protein kinase-D (PKD). The recruitment of the latter to membranes is mainly mediated by its two DAG C1-binding domains [91]. The activation of PKD also requires phosphorylation by protein kinase-C (PKC) which is also recruited to membranes by DAG [92]. Thus, one way to increase the levels of PI4P on the LCV is by generating higher levels of DAG by the function of LpdA and LecE. The higher levels of DAG will result in the recruitment of PKC and PKD to the LCV, then PKC may phosphorylate PKD that will lead to the recruitment of PI4KIIIβ to the LCV that in turn will generate PI4P from PI. It is important to note that this is probably not the only way by which the LCV can recruit PI4KIIIβ since this enzyme can also be recruited from Golgi derived vesicles that fuse with the LCV.
The results presented in this study uncovered an additional layer in the complex interaction between the L. pneumophila phagosome and the host cell, and show that changes in phospholipids composition are manipulated by L. pneumophila effectors in many ways to result with successful infection.
The L. pneumophila wild-type strain used in this work was JR32 [93], a streptomycin-resistant, restriction-negative mutant of L. pneumophila Philadelphia-1, which is a wild-type strain in terms of intracellular growth. In addition, mutant strains derived from JR32, which contain a kanamycin (Km) cassette instead of the icmT gene (GS3011) [11], the lpg2552 gene (RV-L6-45) (this study), a gentamicin (Gm) cassette instead of the lpg1888 gene (RV-L10-71) (this study), and a double lpg2552/lpg1888 deletion (RV-L10-77) (this study) were used. The E. coli strains used were MC1022 [94] and DH5α. The S. cerevisiae wild-type strains used in this work were BY4741 (MATa his3Δ leu2Δ met15Δ ura3Δ) [95] and W303 (MATa leu2-3,112 trp1-1 can1-100 ura3-1 ade2-1 his3-11,15) [96]. In addition, mutant strains derived from BY4741, which contain a G418 cassette instead of the pah1 gene (RV-L8-59) (this study), the dgk1 gene [97] (a kind gift from Prof. Martin Kupiec, Tel-Aviv University) and the nem1 gene (RV-L8-54) (this study) were used. Plasmids and primers used in this work are listed in Table S1 and S2.
The pMMB-cyaA-C vector [98] was used to construct CyaA fusions. In addition, two plasmids were constructed to contain the pUC-18 polylinker, at the same reading frame like pMMB-cyaA-C, in order to generate C-terminal fusions. For the over expression of effectors in yeast the pUC-18 polylinker was cloned into pGREG523 [99], between the EcoRI and HincII restriction sites to generate pRam (this vector was used to construct 13× myc fusions under the yeast GAL1 promoter). For the effectors localization experiments the 13× myc tag was amplified by PCR from pRam using the primers Myc-F-NdeI and Myc-R-yeast, and the PCR product was digested with NdeI and EcoRI. The pUC-18 polylinker was digested from pMMB-cyaA-C with EcoRI and BamHI and the resulting inserts were cloned in a 3-way ligation into pMMB207-NdeI [98], digested with NdeI and BamHI, to generate pMMB-13× myc (this vector was used to construct 13× myc fusions under the bacterial Ptac promoter).
The L. pneumophila genes examined were amplified by PCR using a pair of primers containing suitable restriction sites (Table S2). The PCR products were subsequently digested with the relevant enzymes, and cloned into pUC-18. The plasmids inserts were sequenced to verify that no mutations were introduced during the PCR. The genes were then digested with the same enzymes and cloned into the suitable plasmids described above. Lpg1888 was also cloned into pGREG536 that contain the same reading frame as the above mentioned vectors to generate pGREG536-1888 (generating a 7xHA fusion under the yeast GAL1 promoter).
The pah1 gene was amplified by PCR with its native promoter using the Pah1-for and Pah1-rev primers. The PCR product was cloned into pUC-18, sequenced, and then digested out from pUC-18 using XbaI and PvuII. C-terminal 3xHA tag was amplified by PCR with the primers HA-for and HA-rev using the pYM1 plasmid [100] as template, followed by cloning into pUC-18, sequencing and digest with XbaI and SalI. Both inserts were then cloned into pGREG505 digested with Ecl136 and SalI, in a 3-way ligation, to generate pGREG505-Pah1-3xHA. The genes dgk1, spo7, nem1, pho80 and pho85 were amplified by PCR using the DGK1-SpeI and DGK1-SalI primers for dgk1, the SPO7-SpeI and SPO7-SalI primers for spo7, the Nem1-SpeI and Nem1-SalI primers for nem1, the Pho80-SpeI and Pho80-SalI primers for pho80 and the Pho85-SalI-for and Pho85-SalI-rev primers for pho85 (Table S2). The PCR products were cloned into pUC-18, sequenced, and then digested out from pUC-18 using SpeI and SalI (for pho85 only SalI was used), followed by cloning into different vectors from the pGREG series [99] digested with the same enzymes; pGREG506 for dgk1 to generate pGREG506-Dgk1, pGREG505 or pGREG506 for nem1 to generate pGREG505-Nem1 and pGREG506-Nem1, respectively, pGREG503 or pGREG505 for spo7 to generate pGREG503-Spo7 and pGREG505-Spo7, respectively, pGREG504 or GREG505 for pho80 to generate pGREG504-Pho80 and pGREG505-Pho80, respectively and pGREG506 for pho85 to generate pGREG506-Pho85.
The plasmid pmCherryC1-hMPV [101], that contains an EcoRI site at the same reading frame like in pUC-18, was used in order to construct C-terminal mCherry fusions under the viral pCMV promoter, using the same restriction enzymes as was mentioned above for both lpg2552 and lpg1888, to generate the plasmids listed in the Table S1.
Fragments of 1 kb from the upstream and the downstream regions of the lpg2552 and lpg1888 genes were amplified by PCR using genomic L. pneumophila DNA as a template and pairs of primers containing suitable restriction sites (Table S2). The resulting fragments were digested with the appropriate enzymes and cloned into pUC-18 to generate pRV-lpg2552-UP and pRV-lpg2552-DW, respectively, for lpg2552, and pRV-lpg1888-UP and pRV-lpg1888-DW, respectively, for lpg1888, and sequenced. These two pairs of plasmids were then digested with the respective restriction enzymes and cloned into pUC-18 together with the Km resistance cassette digested with SalI to generate pRV-lpg2552-KM, for lpg2552, or together with the Gm resistance cassette digested with EcoRV to generate pRV-lpg1888-GM, for lpg1888. The two fragments containing the upstream region, the downstream region and the Km/Gm cassette between them were digested with PvuII or SmaI, respectively, and cloned into pLAW344 digested with EcoRV to generate pRV-lpg2552::KM-del and pRV-lpg1888::GM-del, for lpg2552 and lpg1888, respectively. These two plasmids were used for allelic exchange as previously described [102]. For the construction of the double lpg2552::Km/lpg1888::Gm deletion mutant, pRV-lpg1888::GM-del was used to generate the lpg1888::Gm deletion in the lpg2552::Km deletion mutant (RV-L6-45). These strains were examined for intracellular growth in A. castellanii as previously described [103].
In order to construct yeast deletion mutants in the genes pah1 and nem1, a KanMX resistance cassette was amplified by PCR from pM4754 [104] using primer containing the first and last 50 bp of each gene; Pah1-kanMX-for and Pah1-kanMX-rev for pah1, and NEM1-kanMX-for and NEM1-kanMX-rev for nem1. The PCR products were then ethanol precipitated, transformed into wild-type yeast using standard lithium acetate protocol [105], spotted on YPD plates (20 gr glucose, 10 gr yeast-extract, 20 gr peptone in 1 L of distilled H2O) that contained 200 µg/ml G418 and incubated for 2–3 days at 30°C, followed by replica plating on similar plates and incubation for additional 2–3 days at 30°C. Single colonies were then isolated on similar plates and the deletions were verified by PCR.
In order to mutate specific amino acids in the active sites of lpg1888, Pah1 and Dgk1 the PCR overlap-extension approach was used [106], in a similar way as described before [98]. For the construction of site specific mutants in the putative PLD active sites of lpg1888, the primers lpg1888-K165R-F and lpg1888-K165R-R were used to generate 1888-K165R and the primers lpg1888-K376R-F and lpg1888-K376R-R were used to generate 1888-K376R. For the construction of site specific mutant in the Pap1 active site of Pah1, the primers Pah1-D398E-for and Pah1-D398E-rev were used to generate Pah1-D398E. For the construction of site specific mutants in the diacylglycerol kinase active sites of Dgk1, the primers Dgk-R76A-F and Dgk-R76A-R were used to generate Dgk1-R76A and the primers Dgk-D177A-F and Dgk-D177A-R were used to generate Dgk1-D177A.
For all protein fusions described above, the formation of a fusion protein with a proper size was validated by Western blot analysis using the anti CyaA antibody 3D1 (Santa Cruz Biotechnology, Inc.) in the case of the CyaA fusions, using the anti myc antibody 9E10 (Santa Cruz Biotechnology, Inc.) in the case of the 13× myc fusions or using the anti HA antibody F-7 (Santa Cruz Biotechnology, Inc.) in the case of the HA tag fusions. In all cases the primary antibody was diluted 1∶500 and goat anti-mouse IgG conjugated to horseradish peroxidase (Jackson Immunoresearch Laboratories, Inc.) diluted 1∶10,000 was used as the secondary antibody.
Differentiated HL-60-derived human macrophages plated in 24-wells tissue culture dishes at a concentration of 2.5×106 cells/well were used for the assay. Bacteria were grown on CYE (ACES-buffered charcoal yeast extract) plates containing chloramphenicol for 48 h. The bacteria were scraped off the plates and suspended in AYE (ACES-buffered yeast extract) medium, the optical density at 600 nm (OD600) was adjusted to 0.1 in AYE containing chloramphenicol, and the resulting cultures were grown on a roller drum for 17 to 18 h until an OD600 of about 3 (stationary phase) was reached. The bacteria were then diluted in fresh AYE medium to obtain an OD600 of 0.2 and grown for 2 h. IPTG was added to final concentration of 1 mM, and the cultures were grown for additional 2 h. Cells were infected with bacteria harboring the appropriate plasmids at a multiplicity of infection of 4, and the plates were centrifuged at 180×g for 5 min, followed by incubation at 37°C under CO2 (5%) for 2 h. Cells were then washed twice with ice-cold PBS (1.4 M NaCl, 27 mM KCl, 100 mM Na2HPO4, 18 mM KH2PO4) and lysed with 200 µl of lysis buffer (50 mM HCl, 0.1% Triton X-100) at 4°C for 30 min. Lysed samples were boiled for 5 min and neutralized with NaOH. 110 µl of each sample was then transferred to a new tube and 220 µl of cold 95% ethanol was added. Samples were then centrifuged for 5 min at 4°C and the supernatant was transferred to a new tube and stored at −20°C until the next step was performed. The samples were dried in a speed-vac and suspended in 110 µl of sterile DDW. Samples were incubated at 42°C for 5 min, followed by 5 min incubation at room temperature. The levels of cyclic AMP (cAMP) were determined using the cAMP Biotrak enzyme immunoassay system (Amersham Biosciences) according to the manufacturer's instructions.
L. pneumophila effectors encoding genes and S. cerevisiae encoding genes were cloned under the GAL1 promoter in the pGERG yeast expression vectors series as described above. Plasmids were transformed into yeast cells using standard lithium acetate protocol [105], and transformants were selected for the appropriate prototrophy on minimal SD (synthetic defined) dropout plates (20 gr glucose, 6.7 gr yeast nitrogen base, 20 gr agar, 1.5 gr amino-acids mixture without the selective ones, in 1 L of distilled H2O). Resulting transformants were then grown over-night in liquid SD culture medium at 30°C, cell number was adjusted and a series of tenfold dilutions were made. The cultures were then spotted onto the respective SD dropout plates containing 2% glucose or galactose. Plates were incubated at 30°C or 37°C for 2–3 days and visualizes for differences in growth.
Wild-type S. cerevisiae expressing lpg2552 from the GAL1 promoter (pRam-lpg2552) was transformed with a Yep24 based, high copy number, yeast genomic library [107]. About 160,000 transformants were screened for their ability to suppress the toxicity of the lpg2552 over expression on galactose plates at 37°C for three days, and 23 suspected colonies were then isolated twice on similar plates. The suspected suppressors were then subjected to Western-blot analysis in order to confirm that lpg2552 is still intact, and only three suppressors gave a positive result, “Sup-1”, “Sup-13” and “Sup-14”. The library plasmid was recovered from each of these suppressor colonies and re-transformed into the original screening strain to verify the suppression effect. Two of these suspects- “Sup-13” and “Sup-14”, kept the suppressor phenotype at this stage. Sequencing of “Sup-14” reveled that the genomic fragment cloned in the plasmid contained the yeast HIS3 gene and therefore it was left out (HIS3 was the marker that was used to keep lpg2552 plasmid in the yeast cells). “Sup-13” was sequenced and found to contain a fragment of the yeast genome and three sub-clones were constructed from it. Digestion of “Sup-13” with PvuII and self ligation generated pSup-13-sub-clone-1. Digestion of “Sup-13” with SacI and self ligation generated pSup-13-sub-clone-2. Digestion of pSup-13-sub-clone-1 with SmaI and BstEII, followed by treatment with Klenow fragment and self ligation generated pSup-13-sub-clone-3.
S. cerevisiae containing plasmids expressing lpg2552, Spo7-Nem1 or a vector were grown on a roller drum over-night in the appropriate SD medium at 30°C. The following day the cultures were centrifuged and resuspended in SD medium containing 2% galactose and the cultures were grown on a roller drum for additional 6 h at 37°C. The cells were then harvested and subjected to SDS PAGE (0.8%) followed by Western-blot analysis using the anti HA antibody.
COS7 cells were transfected with the FuGENE (Roche) transfection reagent according to the manufacturer's instructions. Briefly, COS7 cells were grown in DMEM (Invitrogen) medium supplemented with 10% FBS. A day prior to transfection the cells were plated in 6-well plates containing 25 mm glass coverslips at a concentration of 3×105 cells per well. The next day the medium was replaced and the cells were transfected using a total of 1–2 µg DNA per well. Following 44–48 h of incubation at 37°C under CO2 (5%), the cells were used for live imaging. The intracellular distribution of the GFP-DAG sensor was classified through: (i) the visual inspection of 330 cells co-expressing Cherry-LecE and GFP-DAG and 473 cells expressing GFP-DAG alone, from three independent experiments; and (ii) the measurement of the ratio of peri-nuclear GFP-DAG to the total cell intensity of the GFP signal; in cells in which a peri-nuclear GFP-DAG signal could be identified. For this quantification, the entire cell volume was imaged, images were projected into two dimensions by summing the pixel intensities of each plane, and GFP signals were identified through intensity-based segmentation. Signal intensities were calculated with Slidebook. Two independent experiments, comprising 40 cells per condition, were performed.
Infected cells were visualized by confocal microscopy. Coverslips were inserted into a 24-wells tissue culture dishes and incubated for 1 h with 10% Poly-L-Lysine (Sigma) diluted in PBS (1.5 M NaCl, 78 mM Na2HPO4, 18.5 mM NaH2PO4·H2O), followed by three washes with PBS. U937 cells were then differentiated into human-like macrophages by addition of 10% normal human serum and 10 ng/ml of phorbol 12-myristate 13-acetate (TPA) (Sigma) at concentration of 0.5×106 cells per well, and incubated at 37°C under CO2 (5%) for 48 h. Bacteria were grown as described above for the CyaA translocation assay. The cells were washed twice with RPMI supplemented with 2 mM glutamine and infected with the wild-type strain (JR32) expressing either the 13×myc-tagged lpg2552 or lpg1888 at multiplicity of infection of 5. Plates were then centrifuged at 180×g for 5 min, incubated at 37°C under CO2 (5%) for 1 h, washed 3 times with PBS++ (PBS containing 1 mM CaCl2 and 0.125 mM MgCl2). The cells were fixed with ice-cold methanol for 5 min, washed twice with PBS and perforated with ice-cold acetone for 2 min. Coverslips were blocked for 10 min with PBS containing 10% BSA and stained with monoclonal chicken anti myc antibody (Millipore) diluted 1∶20 and mouse anti L. pneumophila antibody (Santa Cruz Biotechnology, Inc.) diluted 1∶100 in PBS containing 10% BSA for 1 h, followed by two 5 min washes in PBS containing 10% BSA. Coverslips were then stained with DAPI (Sigma) and with the secondary antibodies Alexa488 goat anti chicken (Invitrogen Inc) and Cy3 donkey anti mouse (Jackson Immunoresearch Laboratories Inc) diluted 1∶400 in PBS containing 10% BSA, followed by two 5 min washes in PBS containing 10% BSA. Coverslips were then mounted on glass slides using mounting solution (Golden Bridge).
Images were acquired using a motorized spinning-disc confocal microscope (Yokogawa CSU-22, Zeiss Axiovert 200 M). The confocal illumination was with 40 mW 473 nm and 10 mW 561 nm solid state lasers. Images were acquired with a 63× oil immersion objective (Plan Apochromat, NA 1.4) For the effectors localization after infection, a Cool Snap HQ-CCD camera (Photometrics) was employed, with a typical exposure times of ∼1 s, images were acquired with 1×1 binning, yielding a pixel size of 0.065 µm. For presentation, fluorescence intensity values were corrected for the contribution of non-specific binding of the secondary/labeled antibody. For the PA and DAG sensors analysis, an Evolve EMCCD camera (Photometrics) was employed, typical exposures of 20–100 ms, 1×1 binning yielding a pixel size of 0.25 µm. Three dimensional image stacks were acquired by sequential acquisition of views recorded every 70–300 ms along the z-axis by varying the position of a piezo electrically controlled stage (step size of 0.4 µm). All images were analyzed with SlideBook software (version 5.0; Intelligent Imaging Innovations).
|
10.1371/journal.pgen.1003111 | ATX1-Generated H3K4me3 Is Required for Efficient Elongation of Transcription, Not Initiation, at ATX1-Regulated Genes | Tri-methylated H3 lysine 4 (H3K4me3) is associated with transcriptionally active genes, but its function in the transcription process is still unclear. Point mutations in the catalytic domain of ATX1 (ARABIDOPSIS TRITHORAX1), a H3K4 methyltransferase, and RNAi knockdowns of subunits of the AtCOMPASS–like (Arabidopsis Complex Proteins Associated with Set) were used to address this question. We demonstrate that both ATX1 and AtCOMPASS–like are required for high level accumulation of TBP (TATA-binding protein) and Pol II at promoters and that this requirement is independent of the catalytic histone modifying activity. However, the catalytic function is critically required for transcription as H3K4me3 levels determine the efficiency of transcription elongation. The roles of H3K4me3, ATX1, and AtCOMPASS–like may be of a general relevance for transcription of Trithorax-activated eukaryotic genes.
| We provide a definitive answer to the question regarding the role of histone H3 lysine 4 tri-methylation marks in the transcription of two ATX1-regulated genes. Despite the proven correlation between the gene transcriptional activity and the level of H3K4me3 modification on the nucleosomes, whether H3K4me3 contributes to, or simply “registers,” active transcription has remained unclear. Another broader-relevance question is whether histone-modifying proteins are required for recruitment of the general transcription machinery, thus playing roles beyond their catalytic activity. Using a combination of gene deletion and specific point mutation analyses, we untangle overlapping effects and reveal that H3K4me3 is not required for TBP/Pol II recruitment to promoters but is critical as an activating mark for transcription elongation. The existing hitherto ambiguity about the role of H3K4me3 as an activating mark has been largely due to the unknown duality of the ATX1/AtCOMPASS functions: facilitating PIC assembly and producing H3K4me3 as an activating mark for transcription elongation.
| The H3K4me3 mark is generally associated with transcriptionally active genes [1]–[3]. Its genome-wide distribution in yeast, animal, and plant genomes displays remarkably conserved, predominantly gene-associated, patterns with a strong bias towards the 5′-ends of transcribed genes [4]–[7]. Despite the demonstrated ability of chromatin remodeling/modifying and mRNA processing proteins to bind the H3K4me3 modification, the actual contribution of H3K4me3 to transcription is still unclear [8], [9].
The eukaryotic histone methyltransferases responsible for the H3K4me3 mark have diverged both evolutionarily and functionally into two families [10]. The TRITHORAX (TRX) family including Drosophila trithorax (Trx), mammalian MLL1-4, and Arabidopsis ATX 1-2 segregate into a phylogenetic subgroup that is distinct from the SET family containing yeast Set1 and its orthologs in other species [11]. SET family members operate more globally across the genome while the TRX family members are more gene-specific. In yeast, Set1is the sole methyltransferase establishing the genome-wide mono-, di-, and tri-methyl H3K4 marks, while MLL1 tri-methylates less than 5% of human genes [12], [13]. Like MLL1, ATX1 tri-methylates H3K4 at specific genes, but is not responsible for overall nucleosome modifications in Arabidopsis [14].
The mechanism of ATX1-dependent gene regulation in Arabidopsis involves features that are both similar and different from yeast Set1 and mammalian MLL models. A distinguishing feature of ATX1-dependent gene regulation is that ATX1 has dual roles upstream and downstream of the transcription start sites (TSS) of regulated genes [15]. At promoters ATX1 is found in a complex with TBP and Pol II affecting the formation/stability of the transcription preinitiation complex (PIC). The second role is within the transcribed region where ATX1 establishes a peak of H3K4me3 modified nucleosomes about 300 bp downstream of the TSS. ATX1's recruitment and ability to tri-methylate nucleosomes in this region requires the activated form of Pol II (phosphorylated at its carboxyl terminal domain (CTD) repeat at serine 5 (Ser5P) [15]. ATX1 binds directly to Ser5P Pol II, an interaction different from the Paf (Polymerase associated factor)-mediated binding of Set1/COMPASS to Ser5P Pol II in yeast [16].
Both SET and TRX family proteins (including the human, Drosophila, and Arabidopsis counterparts) operate within specific complexes, called COMPASS or COMPASS-LIKE, respectively. Both types of complexes share three conserved subunits, WDR5- ASH2L- RbBp5 that are critical for methyltransferase activity of the respective SET or TRX catalytic subunit [17]–[20]. The structural organization and the mechanism by which these three subunits stimulate the enzyme activity and H3K4me3 accumulation have been actively pursued and a significant amount of data for the biochemical and molecular mechanisms is available [21]–[23].
Although it has been well established that knockdown or deletion of a COMPASS or COMPASS-like subunit results in reduced mRNA and H3K4me3 levels of specific genes [17]–[20], how the specific stages of transcription are affected by this deficiency is less understood. Recently, the Drosophila Set1 (dSet1) was shown to be required for efficient release of Pol II into transcription elongation from the heat shock 70 (hsp70) gene [24]. However, the roles of the TRX (MLL or ATX1) type COMPASS-like complexes in the transcription process have not been fully elucidated. Due to the structural and functional differences between the SET and the TRX family members [11], as well as differences in the protein composition and the interaction between the respective subunits of the COMPASS or COMPASS-like complexes [9], it is expected that the complexes supporting the activity of Set1 (including the human and Drosophila homologs) and those for TRX (MLL/ATX1) have functionally diversified as well.
Here, we study the roles of the Arabidopsis ATX1/COMPASS-like during the specific stages of transcription of two ATX1-regulated genes, WRKY70 (AT3G56400, encoding a member of the WRKY family of transcription factors) and LTP7 (AT2G15050, encoding a lipid transfer protein from an antimicrobial peptide family) [25]. ATX1 establishes the H3K4me3 marks at the 5′-end nucleosomes of these genes and is required for their optimal expression in leaves under regular homeostatic conditions [14]. Earlier, it has been reported that ATX1/AtCOMPASS–like affects transcript levels from the developmentally regulated FLC gene [20], [26]. However, how specific stages of transcription are affected has not been elucidated.
To distinguish the discrete transcription stages dependent on H3K4me3 levels from effects caused by the structural disruption of ATX1/AtCOMPASS–like, we used a combination of RNAi-mediated knockdowns of AtCOMPASS–like subunits and specific point mutations to inactivate the catalytic domain of ATX1. We demonstrate that ATX1, AtCOMPASS–like, and H3K4me3 have distinct effects on PIC formation and the transition to transcription elongation. ATX1 and AtCOMPASS–like are required for efficient PIC formation. In contrast to the MLL1-regulated gene model [27], the ATX1-generated H3K4me3 mark is not required for TBP recruitment during transcription initiation, but is critical for activating transcription elongation.
Two Arabidopsis proteins, AtWDR5a and AtWDR5b, are related to WDR5 but only AtWDR5a can form a complex with the other AtCOMPASS–like subunits [20]. We refer to AtWDR5a as AtWDR5 from here on. To analyze the function of the three core AtCOMPASS–like subunits, we generated plants expressing AtWDR5-RNAi, AtASH2-RNAi or AtRbBp5-RNAi constructs. Knockdown lines produced less transcripts from the respective subunit genes, confirming efficient knockdown of their target mRNAs (Figure S1A). The AtWDR5-RNAi, AtASH2-RNAi and AtRbBp5-RNAi knockdown lines displayed early flowering phenotypes similar to the atx1 phenotype, supporting their function in a shared complex (Figure S1B; [20]). Lower expression of any of the AtCOMPASS–like subunits in the respective AtWDR5, AtASH2, or AtRbBp5 RNAi knockdown lines resulted in significantly reduced H3K4me3 levels at the 5′-ends of the WRKY70 and LTP7 genes known to be direct targets of ATX1 (Figure 1A). The WRKY70 and LTP7 genes also produced significantly reduced transcript levels in these knockdown lines (Figure 1B). We conclude that each of the three core AtCOMPASS–like subunits (AtWDR5, AtASH2, and AtRbBp5) must be present for the wild type H3K4me3 and transcript levels from the ATX1-regulated WRKY70 and LTP7 genes.
The presence and the distribution patterns of AtWDR5 at the two ATX1-regulated genes were determined by ChIP analysis with antiWDR5 antibodies. AtWDR5 was found at the promoters and at the transcription start sites (TSS) regions of WRKY70 and LTP7 (Figure 2A, 2B). AtWDR5 accumulation peaked at the 5′-ends, then gradually tapered off downstream, in a profile similar to that of ATX1 (Figure 2C middle row) and H3K4me3 (Figure 2C, bottom row). The overlapping distribution patterns of ATX1 and AtWDR5 are consistent with a function in a shared complex. In addition, ATX1 and AtWDR5 interact directly in the yeast two-hybrid (Y2-H) binding system (Figure S2C; also shown in [20]) and a TAP-tagged AtWDR5 fusion protein used as bait in a pull-down assay successfully recovered ATX1 from total cellular protein extracts (for more details see Methods and Figure S2A).
The ATX1-AtWDR5 interaction was mapped further by Y2-H analysis and in pull down assays (Figure S2C, S2D; see Text S1). Detailed analysis of retained fragments (Figure S2A) by mass spectroscopy (Figure S3A–S3C) confirmed the ATX1 domain necessary and sufficient to bind to AtWDR5 was located immediately upstream of the SET domain (Figure S2B–S2D) and was similar (Figure S3D) to the Win (WDR5-interacting) peptide of MLL1 [21], [22]. This result is important as it indicates that ATX1 in Arabidopsis interacts with AtWDR5 through a conserved domain similar to the MLL1-WDR5 interaction in mammalian cells [28], [29].
Next, we determined whether ATX1 recruits AtCOMPASS–like (via AtWDR5) to the target genes or whether the presence of AtWDR5 was needed for recruiting ATX1. The AtWDR5 levels at the WRKY70 and LTP7 genes in atx1 mutant and wild type backgrounds, determined by ChIP-PCR with antiWDR5 antibodies, were strongly diminished in atx1 relative to the wild type background (Figure 2B). The results indicated ATX1 was required for wild type level occupancy of AtWDR5 at these genes. Thereby, AtWDR5 occupancy is dependent on ATX1 presence at the ATX1-regulated genes and, most likely, ATX1 helps recruit AtCOMPASS–like via binding to AtWDR5.
To determine whether ATX1 presence at the ATX1-regulated genes can occur independently of AtCOMPASS–like, we analyzed ATX1 levels by ChIP-PCR with antiATX1 antibodies in plants depleted for AtWDR5. First, we confirmed that in the AtWDR5-RNAi knockdown lines (Figure S1A) the amounts of AtWDR5 protein at the target gene loci was strongly reduced (Figure 2C, top row). We found that the amounts of ATX1 bound at the target genes in these AtWDR5-depleted lines were similar to their levels in the wild type (Figure 2C, middle row). We conclude that recruitment of AtWDR5 requires ATX1 but the converse is not true: ATX1 occupancy does not depend on AtWDR5.
Interestingly, the amount of ATX1 at the 5′-end regions of the target genes was slightly lower in AtASH2-RNAi lines (Figure S4), possibly suggesting an indirect effect of AtASH2, as AtASH2 does not directly interact with ATX1 (Figure S2C). Likewise, MLL1 does not bind Ash2L directly, but Ash2L is required for maintaining the integrity of the complex at the HOX loci [30].
To determine whether the reduced transcript levels at the two target genes in the RNAi knockdown lines (Figure 1B) were due to defects in transcription or resulted from post-transcriptional events, we measured their rates of transcription. Nuclear run-on assays indicated that in the AtWDR5-RNAi and AtASH2-RNAi knockdown lines the WRKY70 and LTP7 genes were transcribed at much lower rates than in wild type (Figure 3). The reduction in the transcription rates is large enough to indicate that a reduced rate of transcription is the primary defect causing lower WRKY70 and LTP7 transcript levels in the AtWDR5-RNAi and AtASH2-RNAi knockdown lines.
To gain insights into the role of AtCOMPASS–like in specific stages of transcription, we examined a possible role at the promoters by measuring TBP accumulation in the RNAi knockdown lines. Decreased AtWDR5 or AtASH2 mRNA levels correlated with ∼50% decrease in TBP levels (Figure 4A) suggesting an involvement of AtCOMPASS–like in TBP/PIC assembly.
Reduced TBP levels were likely to be associated with reduced Pol II recruitment. Therefore, we examined the occupancy of total Pol II at the analyzed genes. Total Pol II levels in AtWDR5-RNAi or AtASH2-RNAi knockdown lines were measured by ChIP with antibodies that do not discriminate between the non-phosphorylated and phosphorylated forms of Pol II (Figure 4B). Total Pol II accumulation at the 5′-ends of the three genes ranged from 64%–81% of wild type levels (Table 1).
Next, we measured the amount of Pol II phosphorylated at serine 5 of the CTD repeat (Ser5P Pol II) as this modification marks the transition of Pol II from the promoter (promoter clearance) to the sites of transcription initiation [31]–[33]. Ser5P Pol II levels near the TSSs of the genes in the AtWDR5-RNAi or AtASH2-RNAi lines were 75%–85% of wild type levels (Figure 4C, Table 1) indicating that Pol II accumulation at the TSSs was affected less strongly than the TBP/Pol II levels at the promoters. The strong reductions in the genes' transcript levels and transcription rates in the RNAi lines (Figure 4B), and the presence of relatively high total Pol II and/or Ser5P Pol II at the promoters and the 5′-ends of the genes (Figure 4B, 4C) suggested that disruption of AtCOMPASS–like also affected transcription downstream of these stages of transcription.
The amounts of total Pol II towards the genes' 3′-ends were measured to determine if transcription elongation was impaired. The total Pol II levels at the 3′-ends of the WRKY70 and LTP7 genes were strongly decreased in the RNAi lines (48%–55% of wild type levels, Figure 4B; Table 1). The differences between lower Pol II amounts at the 3(-ends and higher amounts at the 5(-ends were significant: (p-value(0.05 for WRKY70 and LTP7). Less Pol II at the genes' 3(-ends suggested impaired transcription elongation.
The distribution of Ser2P Pol II, which marks the transition of Pol II to the elongation phase [17], [32], [34], [35], was analyzed next. In the wild type background the Ser2P Pol II distribution increased towards the 3(-ends of the genes (Figure 5). In the AtWDR5-RNAi or AtASH2-RNAi knockdown lines, however, the Ser2P Pol II occupancy was considerably reduced (Figure 5). Importantly, the differences between the lower amounts of Pol II Ser2P at the 3′ ends and the higher amounts of Pol II Ser5P at the 5′ ends were significant (p-values<0.01, Table 1). We conclude that disruption of AtCOMPASS–like affected the transition from transcription initiation to transcription elongation.
Collectively, the results indicate that AtCOMPASS–like plays a role at the promoters (lower TBP and Pol II levels in RNAi lines), has a lesser effect on the Ser5P Pol II levels during the promoter clearance and transcription initiation, but is critically required for productive transcription elongation at the ATX-regulated genes.
The use of an ATX1 T-DNA insertion mutant (atx1) or RNAi lines for depleting individual subunits of ATX1/AtCOMPASS–like decreases the amounts of both the intact complex and of H3K4me3, making it impossible to elucidate which of these alterations were affecting transcription. To distinguish the effects caused by changes in the structure of the ATX1/AtCOMPASS–like complex from effects caused by diminished H3K4me3 levels, we constructed an ATX1 mutant transgene, ATX1-set, containing point mutations expected to inactivate its catalytic methyltransferase domain while maintaining its structural integrity. This catalytically inactive ATX1-set gene contains 5 tyrosine to alanine mutations at positions that are evolutionarily conserved in ATX1 and MLL1 (see Methods). One of the conserved tyrosines (ATX1 Y1015) is known to be essential for the methyltransferase activity of the SET domain of human SET7/9 [36]. Expression of the HA-tagged ATX1-set protein in transgenic atx1::ATX1-set lines was verified by immunoblot analysis (Figure S5A). The failure of ATX1-set to rescue the early-flowering phenotype of atx1 (Figure S5B) supports a deficiency in ATX1 function in the atx1::ATX1-set lines.
Analyses of the transcriptional responses of ATX1-regulated genes in atx1 plants expressing the ATX1-set transgene (atx1::ATX1-set) indicated the WRKY 70 and LTP7 transcripts were not restored to their wild type levels (Figure 6A). The reduced WRKY 70 and LTP7 transcription was due to the mutations in the catalytic domain and not the HA tag fusion protein structure as complementation of atx1 with a HA-tagged version of wild type ATX1 restored the expression of these genes (Figure S5C). Complemented atx1 plants also displayed the wild type flowering phenotype (Figure S5B) supporting the conclusion that the effects observed in the atx1::ATX1-set lines were caused by the deficient catalytic activity of ATX1-set.
One possible mechanism for the lower WRKY 70 and LTP7 transcription in the atx1::ATX1-set background could be the inability of ATX1-set to be recruited to its targets. ChIP assays with antiHA antibodies indicated the ATX1-set protein was located at the WRKY 70 and LTP7 genes (Figure 6B). However, despite the ATX1-set recruitment and accumulation at its targets, H3K4me3 levels were significantly lower than in wild type and comparable to the levels in atx1 mutants (Figure 6C). We conclude that, although ATX1-set was present at the 5′-ends of its gene targets, the ATX1-set histone modifying activity was strongly decreased or absent. The next question, then, was whether ATX1-set could still recruit AtWDR5.
ChIP assays with antiWDR5 antibodies indicated that both the amounts and the distribution patterns of AtWDR5 were similar to those in wild type and significantly higher than in an atx1 background (Figure 7A). The results indicated the ATX1-set protein maintained its structural integrity and ability to interact with the AtCOMPASS–like complex at its target genes. These results justify the use of the atx1::ATX1-set plants to assess how diminished H3K4me3 levels affected transcription without apparent changes in the other functions of the ATX1-set/AtCOMPASS–like complex.
The diminished transcript and H3K4me3 levels at the WRKY70 and LTP7 genes in the atx1::ATX1-set background (Figure 6C) indicate this histone modification is a major factor responsible for the decreased transcript levels from these genes (Figure 6A). Next, we analyzed the effects of the diminished amounts of H3K4me3 in atx1::ATX1-set plants upon the levels of the elongating Ser2P Pol II. We found that the levels of Ser2P Pol II were low and comparable to the levels in the atx1 background (Figure 7B), indicating overall transcription was diminished.
Whether the diminished H3K4me3 levels affected TBP recruitment to the promoters was analyzed in atx1::ATX1-set plants. TBP accumulated at the promoters to levels similar to wild type and much higher than in the atx1 background (Figure 7C). Accumulation of Pol II in its initiation-activated form (Ser5P) at the WRKY70 and LTP7 promoters was nearly at wild type levels in atx1::ATX1-set and again at higher levels than in atx1 (Figure 7D). The most important consequences of this result are that the absence of ATX1-generated H3K4me3 marks did not markedly interfere with the assembly of the basal transcriptional machinery and, that the primary defect in transcription was in the attenuated levels of Pol II Ser2P levels at the genes 3′ ends (Figure 7B).
Summarily, despite ‘normal’ recruitment of TBP, ATX1 and AtCOMPASS–like to the 5′-ends of the genes, the rates of transcription elongation were diminished when H3K4me3 levels were low, providing compelling evidence that H3K4me3 is an activating mark for elongation at ATX1/AtCOMPASS–like regulated genes.
A large body of published work has demonstrated that expressed genes have higher levels of tri-methylated H3K4 residues on their nucleosomes than non-expressed genes [12], [14], [37]–[39]. Although deficiencies in H3K4me3 via knockdown of COMPASS subunits result in reduced levels of mRNA production [17]–[20], the mechanisms by which H3K4me3 affects transcription are still emerging [8], [9]. Revealing a causative link between H3K4me3 and transcription has been particularly challenging, as histone methyltransferases are multidomain proteins that function within large protein complexes. Consequently, interpretation of results based on knockdown mutations may be misleading as the effects could result from disruption of the multiple functions of the protein and complexes involved. For example, mutating ATX1 affects the assembly or stability of TBP/PIC as well as H3K4 methylation at downstream nucleosomes [15]. These observations underscore the need for specific mutations that affect only one function while maintaining the structural integrity of the protein of interest.
In addition to reported similarities of AtCOMPASS–like [20] with the extensively studied COMPASS/COMPASS-like complexes in animals and yeast, we establish a role for AtCOMPASS–like in transcription that has not been reported for yeast, fly or mammalian complexes. We demonstrate that AtCOMPASS–like, as shown earlier for ATX1 [15], has dual roles in transcription initiation and H3K4 tri-methylation. Specifically, AtCOMPASS–like is recruited to promoters by ATX1 (Figure 2B) and plays a role in TBP/PIC assembly and/or stability. Reduced amounts of AtWDR5 or AtASH2 caused ∼50% decreases in TBP levels at the promoters (Figure 4A), attenuated transcription, and reduced Pol II levels on the transcribed genes (Table 1). These results link the AtCOMPASS–like complex with the basal transcriptional machinery. In these studies, as in all studies reported for other systems, we have used RNAi to deplete subunits of AtCOMPASS–like. As a consequence, it was not possible to separate the effects of structurally disrupting AtCOMPASS–like from the effects resulting from low H3K4me3 levels. Here, we successfully uncouple the ATX1/AtCOMPASS–like structural contributions from changes in H3K4me3 levels through analysis of atx1::ATX1-set mutant plants. The point mutations in this ATX1-set mutant protein greatly diminish (or eliminate) methyltransferase activity in vivo as ATX1 target genes had H3K4me3 levels that were identical to those in the atx1 background. The apparent structural integrity of the ATX1-set mutant supported, first, by its ability to be correctly recruited to its target genes and second, by its ability to efficiently recruit AtWDR5 (Figure 7A) and to recruit/stabilize TBP/Pol II to promoters (Figure 7C, 7D) despite an apparent lack of catalytic methyltransferase activity in vivo, allowed us to clearly separate the dual roles of ATX1/AtCOMPASS–like in transcription.
The ATX1-regulated WRKY70 and LTP genes in wild type, atx1, and atx1::ATX1-set mutant backgrounds displayed clear differences in their transcriptional behavior. The strongly reduced WRKY70 and LTP transcript production in the absence of ATX1 in atx1 mutants was consistent with lower TBP and Pol II occupancy at the promoters [15]. In contrast, in the atx1::ATX1-set mutant, the TBP/Pol II (PIC) levels at the promoters were similar to the wild type (Figure 7C, 7D). This result demonstrates that PIC levels at promoters depend on the structural integrity of ATX1 but not on its H3K4 methyltransferase activity. The results from this study, together with the finding of ATX1 in a protein complex with TBP and ATX1's ability to bind directly to the non-phosphorylated form of Pol II [15], define a novel role for ATX1/AtCOMPASS–like as a transcriptional co-activator separate and largely independent of its histone modifying activity. This model differs from the binding of the TAF3 subunit of TFIID to H3K4me3 at MLL1-regulated genes [27]. Additionally, the Arabidopsis genome lacks a TAF3 subunit [40], making anchoring of TFIID to H3K4me3 nucleosome an unlikely mechanism for ATX1-regulated genes.
It is important to note also that, hitherto, a role of histone modifying proteins in PIC formation, that is independent of their histone modification activity, has been found only for yeast histone acetyltransferases [41]–[43]. Our results provide the first demonstration of a histone methyltransferase as an essential component of the general transcription machinery independent of its methyltransferase activity.
Furthermore, as knockdown of the AtWDR5 or AtASH2 subunits reduced TBP occupancy by ∼50% at the promoters, it was surprising that the Ser5P Pol II levels at the 5′-ends of the genes was only slightly decreased (Figure 4C; Table 1). As Ser5P Pol II is a biochemical marker for transcription initiation and early elongation [44]–[46], the results indicated that although required for normal TBP levels, ATX1/AtCOMPASS–like had a lesser effect on Pol II levels after promoter clearance. A possible reason is that a rate-limiting step in transcription on these templates is downstream of these early stages.
As the ATX1-dependent TBP levels at the promoters were similar to wild type, the attenuated transcript production in atx1::ATX1-set mutant lines indicated H3K4me3 was required for efficient transcriptional processes taking place after PIC formation. Together, the results showing relatively high TBP and Pol II levels at the genes' 5′ ends (Figure 4A–4C), reduced rates of transcription (Figure 3), and reduced amounts of Pol II and its elongating Ser2P form at the genes' 3′ ends (Figure 5, Table 1), indicate that transcription elongation is diminished in H3K4me3 deficient genes.
Our results are consistent with a model (Figure 8) in which lower transcription and lower Ser2P Pol II amounts at the genes' 3′ ends are due to slow release of Pol II from a promoter proximal pause site into productive elongation. Diminished release would account for the accumulation of Ser5P Pol II at the genes' 5′ ends, relative to their 3′ ends, in agreement with the model suggested for the Drosophila hsp70 gene when depleted of the dSet1 protein [24]. We suggest that H3K4me3 generated by either Set/COMPASS or TRX/COMPASS-like complexes plays similar roles in activating the transition to transcription elongation. However, for the Set/COMPASS it remains to be established whether it affects the basal transcriptional machinery similarly to the role found here for ATX1/AtCOMPASS–like.
Regulation of elongation is emerging as a critical mechanism for regulating transcription in developmentally regulated and heat-shock induced animal genes, where the limiting step is the release of paused/stalled Pol II into elongation [47]–[50]. It is important to emphasize a principle difference between animal genes regulated by paused/stalled Pol II and the ATX1-regulated genes reported here: these animal genes carry pre-accumulated Ser5P Pol II at their 5′-ends before entering active transcription and they require stimulation by additional factors to release them into productive elongation[33], [47], [49]. In contrast, the ATX1-regulated genes studied here are actively transcribed in non-stressed differentiated tissues and do not have paused/stalled Pol II at their 5′-ends. However, the genes experience accumulation of Ser5P Pol II downstream of TSS (as a form of pausing/stalling) when H3K4me3 levels are depleted. It is tempting to speculate that regulation of H3K4me3 levels is a more general mechanism controlling elongation not limited to inducible genes or genes with pre-stalled Pol II.
Summarily, we conclude that although the presence of ATX1/AtCOMPASS–like is required for assembly of the basal transcription machinery (transcription initiation) at the promoter, the H3K4me3 mark generated by ATX1/AtCOMPASS–like is not required for transcription initiation, but is an activating mark for transcription elongation. The mechanisms by which H3K4me3 affects transition to productive transcription elongation remain to be established. H3K4me3 may be responsible for the generation of a chromatin structure at the 5′-end to ensure optimal Pol II release into productive elongation and/or recruitment of pre-mRNA processing and elongation factors to the 5′ regions of genes [18], [51]–[53].
Lastly, our finding that the ATX1-Win domain is a functional counterpart of the Win domain in MLL1 (Figure S2, Figure S3) [21], [22], [54] suggests that ATX1 integrates into the AtCOMPASS–like exclusively through the Win-mediated binding to AtWDR5. This result underscores the relatedness of the ATX1/AtCOMPASS–like with the human MLL1/COMPASS-like [28]. Therefore, our results may have a broader relevance for the TRX-regulated genes in eukaryotes.
Arabidopsis plants (WS ecotype) were grown for 14–21 d in potting soil in growth rooms at 22°C with a 12-h light photoperiod. Descriptions of all the cloning vectors and primers used in this study, as well as plasmid construction and generation of AtWDR5a-RNAi, AtASH-RNAi, and AtRbBp5-RNAi transgenic lines, are provided in ST 1. Genes used in this study have the following IDs: ATX1 (At2g31650), WDR5a (At3g49660), AtASH2 (At1g51450), AtRbBp5 (At3g21060), TBP(At3g13445), WRKY70 (AT3G56400), LTP7 (AT2G15050).
The AH190 strain was transformed with one of the following bait constructs: pGBKT-AtWDR5, pGBKT-AtASH2, or pGBKT-AtRbBp5, then transformed with one of the prey constructs: pGADT7-ATX1, pGADT7-ATX1N, pGADT7-ATX1C, pGADT7-ATX1DH, pGADT7-ATX1win, pGADT7-ATX1SET, pGADT7-AtAsh2, or pGADT7-AtRbBp5. Yeast were scored for protein interactions by their ability to grow on SD medium lacking Trp, Leu, His, and Ade.
For GST-bead pull down assays, GST beads were incubated with 2 µg of each GST fusion protein, washed, and then incubated with 3 µg of a His-fusion protein overnight at 4°C. Mock controls used extracts prepared from E. coli containing the His-Tag or GST vectors. The beads were washed five times (1×PBS buffer, pH 7.4, containing 140 mM NaCl, 1 mM PMSF and 0.1% TritonX-100), and the remaining proteins eluted from the washed beads in SDS-loading buffer, separated on a 12% PAGE/SDS gel, and analyzed by anti-GST (G018, Applied Biological Materials, Richmond, BC, Canada, lot: 5019) or anti-His antibody (05-949, Millipore, Lot:1487531).
Analyses were performed as described earlier [55].
The ChIP assay was performed using a modified method [56]. Briefly, 3 g of leaves were fixed with 1% formaldehyde for 10 min and quenched in 0.125 M glycine. The treated leaves were ground in a mortar and pestle in liquid nitrogen, the resulting powder was solubilized in extraction buffer and filtered through miracloth. After sonication and centrifugation, the supernatant was pre-cleared with protein A magnetic beads (Invitrogen, Carlsbad, CA), and immunoprecipitated with one of the following antibodies recognizing: Pol II (ab817, Abcam, Cambridge, MA, Lot: 669648); the Ser2P form of Pol II CTD (ab5095, Abcam, Cambridge, MA, Lot: 703307); the Ser5P form of Pol II CTD (ab5131, Abcam, Cambridge, MA, Lot: 806890); trimethyl-H3K4 (ab8580, Abcam, Cambridge, MA, Lot: 598382); ATX1 (rabbit sera, GenScript, SC1031); AtWDR5 (ab75439, Abcam, Cambridge, MA, Lot:872536);); TBP (ab52887, Abcam, Lot:347607), or control IgG serum added for overnight incubation at 4°C. The antibody-protein complexes were isolated by binding to protein A or protein G beads. The washed beads were heated at 65°C for 8 h with proteinase K to reverse the formaldehyde cross-linking and digest proteins. The sample was then extracted with phenol/chloroform, the DNA precipitated in ethanol, and then re-suspended in water. Purified DNA was analyzed by real-time PCR with gene-specific primers. In all ChIP experiments DNA has been fragmented to 100–500 bp with the majority fragments having a length of 200–300 bp.
The ATX1-setm mutant contains 5 tyrosine to alanine substitutions: Y927A, Y945A, Y954A, Y1013A, and Y1015A. The ATX1-set is a synthetic gene expressed from the MAS promoter and encoding the wild type ATX1 protein with a N-terminal HA fusion, except for the mutations noted above.
A Q-TOF Ultima tandem mass spectrometer (Waters) with electrospray ionization was used to analyze the eluting peptides. The stained bands were excised and subjected to LC/MS as described [57]. Gel pieces were digested by trypsin (no. V5111, Promega, Madison, WI) and digested peptides were extracted in 5% formic acid/50% acetonitrile and separated using C18 reversed phase LC column (75 micron×15 cm, Pepmap 300, 5 micron particle size) (Dionex, Sunnyvale, CA). A Q-TOF Ultima tandem mass spectrometer (Waters) with electrospray ionization was used to analyze the eluting peptides. The system was user-controlled employing MassLynx software (v 4.1, Waters) in data-dependant acquisition mode with the following parameters: 0.9-sec survey scan (380–1900 Da) followed by up to three 1.4-sec MS/MS acquisitions (60–1900 Da). The instrument was operated at a mass resolution of 8000. The instrument was calibrated using the fragment ion masses of doubly protonated Glu-fibrinopeptide. The peak lists of MS/MS data were generated using Distiller (Matrix Science, London, UK) using charge state recognition and de-isotoping with the other default parameters for Q-TOF data. Data base searches of the acquired MS/MS spectra were performed using Mascot (Matrix Science, v1.9.0, London, UK). The NCBI non-redundant database (2010130-10386837 sequences 3543419944 residues) was used restricted to Arabidopsis thaliana. Search parameters used were: no restrictions on protein molecular weight or pI, enzymatic specificity was set to trypsin with up to 3 missed cleavage sites, carbamidomethylation of C was selected as a fixed modification. Mass accuracy settings were 0.15 daltons for peptide mass.
Total RNA isolation and reverse transcription with oligo(dT) (18418-012; Invitrogen, Carlsbad, CA) were performed as described previously [15], [58]. Transcript levels were measured with gene-specific primers by real-time PCR analysis with a cyclerIQ real-time PCR instrument (Bio-Rad, Hercules, CA) and SYBR Green mixture (Bio-Rad, Hercules, CA). The relative amount of specific gene transcripts was quantitated with the 2−ΔΔCt calculation according to the manufacturer's software (Bio-Rad, Hercules, CA), where ΔΔCt is the difference in the threshold cycles and the reference housekeeping gene; ACT7 was used as an internal control for ChIP experiments and immunoprecipitated DNA was expressed as a percent of input DNA.
Primers used for the various cloning and analytical procedures are in Table S1.
|
10.1371/journal.pmed.1002614 | High levels of sewage contamination released from urban areas after storm events: A quantitative survey with sewage specific bacterial indicators | Past studies have demonstrated an association between waterborne disease and heavy precipitation, and climate change is predicted to increase the frequency of these types of intense storm events in some parts of the United States. In this study, we examined the linkage between rainfall and sewage contamination of urban waterways and quantified the amount of sewage released from a major urban area under different hydrologic conditions to identify conditions that increase human risk of exposure to sewage.
Rain events and low-flow periods were intensively sampled to quantify loads of sewage based on two genetic markers for human-associated indicator bacteria (human Bacteroides and Lachnospiraceae). Samples were collected at a Lake Michigan estuary and at three river locations immediately upstream. Concentrations of indicators were analyzed using quantitative polymerase chain reaction (qPCR), and loads were calculated from streamflow data collected at each location. Human-associated indicators were found during periods of low flow, and loads increased one to two orders of magnitude during rain events from stormwater discharges contaminated with sewage. Combined sewer overflow (CSO) events increased concentrations and loads of human-associated indicators an order of magnitude greater than heavy rainfall events without CSO influence. Human-associated indicator yields (load per km2 of land per day) were related to the degree of urbanization in each watershed. Contamination in surface waters were at levels above the acceptable risk for recreational use. Further, evidence of sewage exfiltration from pipes threatens drinking water distribution systems and source water. While this study clearly demonstrates widespread sewage contamination released from urban areas, a limitation of this study is understanding human exposure and illness rates, which are dependent on multiple factors, and gaps in our knowledge of the ultimate health outcomes.
With the prediction of more intense rain events in certain regions due to climate change, sewer overflows and contamination from failing sewer infrastructure may increase, resulting in increases in waterborne pathogen burdens in waterways. These findings quantify hazards in exposure pathways from rain events and illustrate the additional stress that climate change may have on urban water systems. This information could be used to prioritize efforts to invest in failing sewer infrastructure and create appropriate goals to address the health concerns posed by sewage contamination from urban areas.
| Waterborne illness has been linked to extreme rainfall, and climate change is expected to bring about more intense storms in some parts of the US.
Fecal pollution is widespread in the environment after rainfall, but high-risk sources such as sewage are difficult to distinguish from animal sources without specialized testing.
Most water assessments are based on measuring concentrations of general indicators in grab samples; here, we set out to quantify the total amount of sewage released from a city after rainfall using automated, high-frequency sampling over multiple days.
Understanding sewage contamination in response to rainfall dynamics and loads is important for understanding risk and how this might change with climate change.
We found widespread evidence of sewage contamination following rainfall at levels high enough to exceed acceptable risk (0.03) for illness if exposed to during recreation in rivers or swimming at nearby beaches.
Rainfall that exceeded 2 inches (50 mm) in 24 hours in spring was associated with much larger amounts of sewage contamination in the rivers and estuary compared with other rain events.
Under conditions of very heavy rain, combined sewer overflows occurred and released 10 times more sewage compared with when there were heavy rain events with no sewage overflows.
Sewage contamination levels were directly linked with the amount of urbanization and density of impervious surfaces in watersheds.
Human risk could be reduced by discouraging contact with river water or nearby beaches after rainfall.
Leaking sewage from sanitary sewer pipes can threaten surface waters and drinking water distribution pipes when they are under low pressure, such as during water main breaks.
Quantification of sewage releases on a city-wide scale allows cities to evaluate the integrity of their sanitation systems and gauge serious infrastructure problems under increasing pressures from climate change.
| Waterborne illness is predicted to increase as climate change alters rainfall patterns [1–4]. In particular, an increased frequency in extreme rain events is predicted for the Northeast, Pacific Northwest, and Great Lakes regions, which can increase exposure to pathogens [5]. Heavy rainfall has been linked with increased waterborne disease outbreaks [4,6]. Most notably, the waterborne outbreaks of Escherichia coli 0157:H7 and Campylobacter jejuni in Walkerton, Ontario [7,8] and Cryptosporidium in Milwaukee, Wisconsin [9] were preceded by extreme rainfall events, although these outbreaks also involved failures in drinking water treatment, monitoring, and human error [5,7,9]. The most common waterborne disease is gastrointestinal (GI) illness, and endemic occurrence in the community is difficult to quantify because most waterborne cases are sporadic and often not recognized as associated with water exposures [10]. However, studies estimate there are 11 to 19 million cases of GI illness from contaminated drinking water [11–13] and an estimated 90 million cases from exposure to recreational waters [14] each year.
Waterborne pathogens are carried in fecal pollution from animals and humans [10]. Humans can be infected following exposure (often through ingestion) to contaminated drinking or recreational water. Fecal pollution has been found to be widespread in the environment following rainfall events and/or snowmelt [15,16]; however, the source of contamination is difficult to discern using standard fecal bacteria indicators. Human fecal contamination, i.e., untreated sewage, has the highest potential to cause disease because humans are the reservoirs for many human pathogens [17], although agricultural runoff can also carry zoonotic pathogens [10]. In urban areas, untreated sewage released from failing sewer infrastructure can leach into soil and migrate into groundwater [18,19] and into drinking water distribution systems under conditions of reduced pressure [20]. Stormwater systems have been found to be frequently contaminated by sanitary sewage as a result of infiltration of leaking sewage or illicit cross-connections, resulting in untreated sewage discharging directly into rivers and streams [21,22]. Furthermore, under extreme precipitation events, sewer systems can become inundated with rainwater and cause sewer overflows [23,24]. Combined sewer systems are particularly vulnerable to overflows, as they collect runoff from impervious surfaces and convey sanitary sewage and stormwater to wastewater treatment plants. The US Environmental Protection Agency (EPA) estimates 850 billion gallons of untreated sewage is discharged annually into US waterways by combined sewer overflows (CSOs) and up to 10 billion gallons from separated sewer overflows (SSOs) [25]. Leaking septic systems may also contaminate groundwater or surface waters in suburban and rural areas [26,27].
E. coli, enterococci, and fecal coliforms are all commonly used as indicators of fecal pollution because they are present in the GI tract of humans and most warm-blooded animals and are easily grown in a laboratory [28]. These standard indicators, however, are not specific to the source of fecal contamination, which is important information needed to more accurately estimate risk to human health and assess sources of contamination [29,30]. Genetic markers for human-associated indicator bacteria, such as human Bacteroides (HB) and human Lachnospiraceae (Lachno2), can be used as proxies for human sewage. These indicators are highly correlated in sewage; thus, using them in tandem increases the reliability of tracking sewage in an urban environment where nonhuman fecal sources are also present [31,32].
Our study site in Milwaukee, Wisconsin is typical of highly urbanized areas, and by using human-associated indicators, our lab has documented frequent sewage contamination in rivers and nearshore Lake Michigan [16,31,32]. In this study, we aimed to assess the amount of sewage released from an urban area following rain events and evaluate which watershed, each with different land-use characteristics, was the largest contributor to sewage contamination. The oldest parts of the city have a combined sewer system, which are common in cities in the Northeast, Pacific Northwest, and Great Lakes regions [33]. In Milwaukee, this system overflows 1–3 times per year under conditions of extreme rainfall. We used automated, high-frequency sampling over several days to (1) quantify sewage loads discharged into Lake Michigan through the Milwaukee estuary and the three rivers upstream during low-flow periods and rain events; (2) compare the concentrations and loads produced during rain events to CSO events in relation to potential health risk; (3) investigate relationships between the degree of urbanization in watersheds and the fluxes of sewage they are contributing; and (4) establish quantitative benchmarks of sewer infrastructure integrity that can be used to monitor improvement or further deterioration.
This study was conducted in metropolitan Milwaukee, Wisconsin, at the Milwaukee estuary and at the lower reaches of the three major rivers forming the estuary—the Milwaukee (MKE), Menomonee (MN), and Kinnickinnic (KK) Rivers. The MKE River drains the largest area and has mainly rural and agricultural land uses in the headwaters and a dense urban area near the mouth. The MN River drains a much smaller area with mainly urban and residential land uses. The KK River drains the smallest area, with nearly all urban and industrial land uses and over half of the watershed covered by impervious surfaces. For more details about the sampling locations, see S1 Text and S1 Table. The sampling and data analysis plan is provided as S2 Text.
Sampling was conducted at four sites—one in each of the three rivers and one in the Milwaukee estuary. In April through September 2014 and 2015, samples were collected across the hydrograph using automated Teledyne ISCO 3700 full-size, portable, sequential samplers housed within US Geological Survey (USGS) and Milwaukee Metropolitan Sewerage District (MMSD) monitoring stations (Fig 1, S2 Table). Samples were collected during storm events with a variety of characteristics, and routine samples were collected during periods of dry weather three to four times per sampling season. Over 2,000 samples were collected during a variety of hydrologic events. During sampling periods, a 250-mL sample was collected by the automated sampler every 15 minutes into 1-L bottles and composited in the field. For rain- and CSO-event sampling, the samplers were ideally activated a minimum of two hours prior to expected rainfall and samples were collected continuously for at least 24 hours following the rainfall event. Two 1-L sample bottles were composited in the field, resulting in two-hour composite samples with eight subsamples, for rain and CSO sampling. For dry-weather sampling, the samplers were activated after at least 48 hours of dry weather. Four 1-L bottles were composited in the field, resulting in four-hour composite samples with 16 subsamples, for dry-weather sampling. Samples were retrieved daily and processed in the laboratory within six hours of collection. Sample bottles were cleaned in the field by vigorously rinsing three times with deionized water. One field bottle blank was collected per sampling event to verify that no significant contamination was caused by residual bacteria. Field bottle blanks were collected by cleaning the sample bottles according to standard procedure, pouring deionized water into the sample bottle, and transporting and processing the blank along with the environmental samples. The full method can be found at dx.doi.org/10.17504/protocols.io.prrdm56. All field blanks were nondetections, except for one blank collected during a CSO event, which had HB and Lachno2 concentrations of 550 and 699 copy numbers (CN)/100 mL, respectively. The field procedure for rinsing bottles three times before resetting the sampler after a 24-hour sampling may not have been adequate for the high concentrations in samples during a CSO. However, this one instance of residual contamination would not affect calculations, as it is a very small fraction (<0.2%) of the concentrations detected in the samples.
All samples collected by the automated samplers were analyzed by culture for E. coli, enterococci, and total fecal coliforms using standard methods [35–37]. A volume of 200 mL or 400 mL, depending on expected concentrations, of each sample was filtered onto a 0.22-μm-pore–sized mixed cellulose esters filter (47-mm diameter; Millipore, Billerica, MA) and stored at −80°C prior to conducting DNA extraction using the MPBIO FastDNA SPIN Kit for Soil (MP Biomedicals, Santa Anna, CA). Quantitative polymerase chain reaction (qPCR) was conducted using an Applied Biosystems StepOne Plus Real-Time PCR System Thermal Cycling Block (Applied Biosystems; Foster City, CA) with Taqman hydrolysis probe chemistry. Samples were analyzed by qPCR for the HB, Lachno2, and ruminant-specific assays using previously published methods [32,38]. Only samples from the MKE River and Milwaukee estuary were analyzed for the ruminant-specific indicator, because these sites were expected to have upstream agricultural influences. For additional details about qPCR analysis, assay slope, intercept, efficiency, and limit of quantification, see S3 Table. Overall, due to time and financial constraints, a representative subset of 11 rain events that spanned a range of rainfall conditions, four low-flow periods, and two CSO events were analyzed by qPCR. In total, 755 of the 2,048 samples were analyzed by qPCR for human- and ruminant-specific indicators.
For statistical analysis, results that had detectable, but not quantifiable, concentrations below the limit of quantification were assigned a value equal to the limit of quantification (225 CN/100 mL for samples in which 200 mL were filtered; 112.5 CN/100 mL for samples in which 400 mL were filtered). Results that were nondetections (<15 CN/100 mL for samples in which 200 mL were filtered; <7.5 for samples in which 400 mL were filtered) were assigned a value of zero CN/100 mL. The Spearman’s rank correlation (rho) was used to determine correlations between quantities of genetic markers for human-associated indicator bacteria and streamflow, rainfall, or standard fecal indicators. Correlations between the two human markers were assessed using Pearson’s correlation on log10-transformed data. Differences in the ratio of these markers binned by spring/summer-fall or rain/low flow were assessed by the Wilcoxon rank–sum test. All tests were considered significant at p ≤ 0.05. The R suite of packages [39] was used for all statistical analyses. The stats package in R was used for Spearman’s rank correlations, Pearson’s correlations, and Wilcoxon rank–sum tests.
Hydrologic and CSO events were defined by visually inspecting the MKE, MN, and KK River hydrographs to identify the urban runoff portion of each event. The beginning of each event was defined as the beginning of the rising limb of the hydrograph. The end of each event was designated as the approximate inflection point of the falling limb of the hydrograph, which was defined as the point where the falling limb begins to change concavity, indicating that most of the flow can be attributed to baseflow rather than runoff [40].
Maximum 24-hour mean concentrations were computed for each event in the MKE River. Streamflow was retrieved from USGS continuous monitoring stations on each river (S2 Table). Instantaneous loads of HB and Lachno2 were determined by multiplying streamflow by concentration. Event loads for each genetic marker were computed by integrating the product of flow and concentration over the duration of the hydrograph [41]. Individual concentrations were multiplied by the associated flow volume to compute incremental loadings. Results that had detectable, but not quantifiable, concentrations below the limit of quantification were assigned a value equal to the limit of quantification (225 CN/100 mL for samples in which 200 mL were filtered; 112.5 CN/100 mL for samples in which 400 mL were filtered). Results that were nondetections (<15 CN/100 mL for samples in which 200 mL were filtered; <7.5 for samples in which 400 mL were filtered) were assigned a value of 15 or 7.5 CN/100 mL, depending on the volume of sample filtered. The associated flow volume was estimated by summing the volumes halfway between the samples collected before and after the current sample. For the first sample of each event, volume was summed for the time period between the first and second sample with the first sample as the centroid. For the last sample of each event, volume was summed for the time period between the last and penultimate sample with the last sample as the centroid. Incremental loads from each individual sample were summed for a final event load. Daily fluxes were calculated by dividing the event load by the duration of the event in days. This allowed for a comparison between low-flow periods and storm events that were sampled over different time periods.
Watershed summaries for each sampling location were performed in a geographic information system (GIS). Watershed boundaries were defined for each site: upstream portions of basins were composed of existing linework from the Southeastern Wisconsin Regional Planning Commission [42], and downstream portions were composed of linework that was manually delineated while referencing seamless, online USGS topographic maps available through ESRI. The lower MKE River watershed was used to represent the urban-influenced area of the Milwaukee estuary and MKE River watersheds. Average one-hour rainfall accumulation for each watershed-defined area was determined using radar-indicated rainfall models retrieved from the National Weather Service North-Central River Forecast Center [43] or using MMSD rain gage data (see S1 Text for more information).
The geometric mean concentrations of HB and Lachno2 markers, previously determined in 98 untreated sewage influent samples from Jones Island and South Shore wastewater treatment plants in Milwaukee, Wisconsin collected from 2009 to 2011 [44], were used to estimate the equivalent amount of untreated sewage released from the Milwaukee area following rain events. The Lachno2 marker displayed lower variability in untreated sewage and had a geomean concentration of 5.94 × 107 CN/100 mL, which equates to 2.25 × 109 CN/gallon. Comparisons were expressed as gallons to parallel wastewater treatment plant reporting units on volumes treated and volumes of overflows.
Concentrations and event loads of genetic markers for human-associated fecal indicator bacteria in the Milwaukee estuary displayed seasonal patterns, as well as relationships with rainfall and river streamflow to the estuary. In 2014 and 2015, events were sampled from early spring to late summer, with total rainfall amounts ranging from 7.4 mm during an event on August 21 and 22, 2014, to 86.9 mm during a CSO event on June 17, 2014. Events sampled in the spring of each year generally had greater total rainfall depths and mean event streamflow, with higher genetic marker concentrations measured (Table 1). Maximum 24-hour mean concentrations of HB and Lachno2 were up to 15 and 6 times greater during CSOs compared to the largest rain event, respectively.
The lowest concentrations of genetic markers were found during low-flow periods. Concentrations of genetic markers showed a consistent pattern of increased concentrations with increased flow across the hydrograph measured at the estuary (Fig 2, S2 Fig). The three rivers that collectively drain to the estuary mirrored this pattern (S3 Fig, S4 Fig, S5 Fig). The full set of hydrographs and corresponding host-associated indicators for all events at the estuary and three rivers are shown in S2 Fig, S3 Fig, S4 Fig, and S5 Fig. Of all samples collected in the Milwaukee estuary during a variety of weather conditions in 2014 and 2015 that were analyzed by qPCR (n = 188, Milwaukee estuary site only), concentrations of the two human indicators were significantly correlated to event streamflow volume (HB rho = 0.69, Lachno2 rho = 0.74, p < 0.05) and maximum river streamflow (HB rho = 0.71, Lachno2 rho = 0.75, p < 0.05) measured during sample collection.
A ruminant-specific genetic marker was also analyzed in Milwaukee estuary samples because of the rural and agricultural land uses in the headwaters of the MKE River. The ruminant signal was generally either absent or present at low levels throughout the duration of a rain event but then was detected at greater levels several days following rainfall (Fig 2B, S2 Fig and S5 Fig). General indicators E. coli and enterococci were frequently elevated when either human or ruminant markers increased, illustrating the lack of specificity of these indicators (S2 Fig and S5 Fig).
HB and Lachno2 concentrations were highly correlated among all Milwaukee estuary samples (r = 0.99; p < 0.05; n = 188), with Lachno2 concentrations on average 1.9 times higher than HB concentrations. Season and/or temperature appeared to influence the ecology of these indicators differently. The ratio between Lachno2 and HB peak instantaneous and maximum 24-hour mean concentrations were significantly higher (p < 0.01) in samples collected in the spring than those collected during the summer and fall. Concentrations of Lachno2 in samples collected in the spring ranged from two to five times higher than HB concentrations, whereas samples collected in the summer and during low flow had ratios of Lachno2 to HB that ranged from 1 to 1.5.
Daily flux of HB and Lachno2 were calculated for rain events with no CSOs and low-flow periods to examine the amount of unrecognized sewage inputs as a result of rainfall (Fig 3). Rain event fluxes and total rainfall depth were significantly correlated for both human indicators (HB rho = 0.90, Lachno2 rho = 0.84, p < 0.05). Rain event daily fluxes ranged from 7.4 × 1010 CN/day of HB and 7.9 × 1010 CN/day of Lachno2 released in the estuary during an event in September 2014 (Event 7) to 6.6 × 1013 CN/day of HB and 3.1 × 1014 CN/day of Lachno2 during a storm event in April 2014 (Event 1). This equates to a 1,000-fold difference in amount of sewage released from a very light rain (9.2 mm) in fall versus a heavy rain (58.3 mm) in spring.
We measured sewage loading from the three rivers that discharge to the estuary to examine the association between sewage releases and the different size and land use of each watershed. In general, the MKE River, which has the highest flow and drainage area, had higher human genetic marker event loads and daily fluxes during rain events compared to the other rivers, but on average, this difference was modest (Fig 4).
The flux (CN per day) from each watershed following rain was driven by size, as we found that during the study period in 2014 and 2015, MKE River streamflow was on average two times greater than MN River streamflow and about five times greater than KK River streamflow, and the differences in flux of human markers from the three watersheds was proportional to this difference in flow. However, when fluxes were normalized by drainage area to calculate yield per day, mean daily yields (CN/km2 of entire watershed per day) of both human genetic markers in the KK River were approximately three times greater than those in the MN River and approximately 11 times greater than those in the MKE River (Fig 4).
We also compared sewage loading from different watersheds based on the amount of urbanization and imperviousness (increased runoff) by calculating the yield as a daily flux per urban land cover (CN/km2 of urban area per day). During rain events, urban yields of both genetic markers were more similar across the three watersheds, with average urban yields of genetic markers in the KK River only two times greater than the MN and MKE Rivers (Fig 4). Across the range of different rainfall amounts, the sewage signal from the KK River was more consistent than the MN or MKE Rivers.
During low-flow periods, mean daily fluxes of HB and Lachno2 were 30 to 40 times greater in the KK River than the MN and MKE Rivers. The KK River also has a much larger urban yield than the other two watersheds, suggesting that more sewage was released per unit urban area particularly under low-flow conditions. (Fig 4).
The KK, MN, and MKE Rivers and the estuary consistently exceeded water quality standards for E. coli, enterococci, and fecal coliforms (S1 Fig), particularly under rainfall conditions. There was overall a poor relationship between general and human indicators, with a rho = 0.29 for E. coli versus HB. These weak correlations likely reflect the nonspecific nature of these general fecal indicators. We explored the distribution of HB values in respect to E. coli above or below the water quality advisory limit of 235 colony-forming units (CFU)/100 mL (Fig 5). Very few values fell within the 90% confidence interval, illustrating the poor overall relationship. We were primarily interested in the number of values that fell in quadrant I, where E. coli values were below 235 CFU/100 mL and were outside the 90% confidence interval of what would be predicted for HB; these samples represent the greatest concern for protecting human health because E. coli would not indicate sewage contamination was present.
Two CSO events were sampled on June 18 and 19, 2014, and April 9 and 10, 2015. Based on volumes reported by MMSD, approximately 341.2 million gallons (MG) of untreated sewage mixed with stormwater was released during the 2014 CSO, and 681.1 MG was released during the 2015 CSO. The majority of sewage released during the CSOs occurred at discharge points from the combined sewer system along the KK, MN, and MKE Rivers. The loads of HB and Lachno2 captured at the automated sampling station at each river were determined, and assuming the markers were primarily from CSOs, these loads were proportional to the volume of release in each river (S4 Table). This illustrates the ability of the two human-associated markers to quantify sewage releases reliably.
During CSOs, large volumes of stormwater are mixed with untreated sewage and discharged into urban waterways, making it difficult to estimate what portion of the discharged water is raw sewage and what portion is stormwater. Concentrations of genetic markers for human-associated fecal indicator bacteria in untreated sewage influent samples at the Jones Island and South Shore wastewater treatment plants were used to estimate how many gallons of untreated sewage were discharged into the rivers and the estuary. Lachno2 marker was used because it was the most consistent of the two human makers in treatment plant influent. The geometric mean concentration of Lachno2 in sewage influent samples collected by from 2009–2011 (n = 98) was 2.25 × 109 CN/gallon (i.e., 5.94 × 108 CN/L). Loads of the Lachno2 indicator were converted to gallons of “untreated sewage equivalents” in waterways (Table 2).
To understand climate influences on waterborne disease threats, drivers of contamination and exposure pathways need to be better characterized [5,10]. A key component is to understand mechanisms of sewage contamination of waterways and identify “high-risk” rainfall conditions, such as extreme events. Moving past simply measuring concentrations of standard fecal indicator bacteria in water to using highly specific indicators indicative of human fecal sources will allow us to quantitatively assess urban infrastructure vulnerabilities and provide better estimates of potential risk due to waterborne pathogens.
This study showed the utility of using two human genetic markers, HB and Lachno2, in tandem to reliably track sewage contamination. The steady-state concentrations in untreated sewage reflected the overall contribution of a human population of approximately 1 million people in the service area. Levels of human-associated fecal indicators have been related to pathogen concentrations in untreated sewage to estimate potential human health risk [44–46]. These assays, as well as other genetic marker assays that are used for microbial source tracking, are generally specific to humans but have been found to sporadically amplify animal fecal sources [47–50]; therefore, using two human-associated indicators improved reliability. In this study, HB and Lachno2 were highly correlated in river and estuary samples, indicating there is a high probability that the fecal pollution is from a human sewage source [31]. We found differences in the ratio of the two indicators in spring rain events compared with other sample times, suggesting Lachnospiraceae (a gram-positive organism) and Bacteroides (a gram-negative organism) have different survival characteristics in the environment. Further work to understand the variables (e.g., time and temperature) that account for different ratios would be useful to create a metric for how long sewage contamination has been in the environment, which would influence the infectivity of waterborne pathogens.
There is accumulating evidence that sewage leaking from sanitary sewer infrastructure can be mobilized during rain events and contaminate stormwater systems [21,22]. In this study, we attempt to quantify this “pulse” of sewage from a metropolitan area under different rainfall conditions. The majority of our urbanized study area has separated sewer systems, with only 12% of the sewer service area comprised of combined sewers. We found increased concentrations of human-associated indicators with increased flow across all three watersheds and in the estuary, suggesting sewage sources were dependent on hydrologic influences. These results are consistent with other studies which have also found that concentrations of fecal indicator bacteria increase and decrease with changes in streamflow [51,52].
We observed two inches (50 mm) of rain in the spring resulted in dramatic increases in sewage contamination in the absence of reported overflows. This suggests there may be a critical threshold for municipal sanitary sewers in separated sewer systems, similar to the levels of rainfall that we found can trigger a CSO [53]. This is important because there is a predicted increase in frequency of these types of rain events for certain parts of the country [2,5]. Rainwater infiltration and inflow to separated sanitary sewer pipes can drastically increase the volumes of water in the sewer system [25]. Monitoring programs in urban areas could be designed to intensively sample during large rainfalls to determine what critical rainfall amounts overwhelm separated sanitary systems their city.
Combined sewer systems in the oldest parts of some US cities are legacy infrastructure from the early 1900s, and the EPA permits a certain number of discharges from these systems [25]. Under the largest storm events, CSOs introduce pathogens, particularly human viruses, into receiving waters [54,55]. In 2014, 187 communities released 22 billion gallons of untreated sewage mixed with stormwater into the Great Lakes [33], which are a drinking water source to nearly 40 million people and have more than 500 beaches along the 4,500 miles of coastline. The highest density of combined sewer systems are in the Northeast, Pacific Northwest, and Great Lakes regions, which are the same regions that are predicted to have the largest increase in extreme events due to climate change [2,5,56]. Although CSOs pose an obvious health risk over a few days per year, this is rivaled by the chronic health risk caused by lesser but more persistent contamination introduced after rainfall.
We demonstrated that across three watersheds with varying drainage areas and land use, urbanization can primarily account for sewage yields, which suggests our results could be generalized to other urban areas in the US. We found the KK River was a consistent sewage source regardless of rainfall. The KK River is the most urbanized and downstream watershed of the three, suggesting that even low amounts of rainfall effectively mobilize sewage that has escaped the sanitary sewer pipes through failing infrastructure or illicit connections, whereas in the MN and MKE River watersheds, larger rainfalls were needed to mobilize this system. These results might suggest that failing infrastructure is more problematic in the KK watershed. Differences in transport or attenuation of contamination from this small watershed near the estuary compared with the larger MN and MKE watershed may also play a role. Numerous stormwater outfalls line the concrete channel of the KK River and likely serve as a conduit for leaking sanitary sewers and sewage from illicit cross-connections to reach the river [22,32].
Standard fecal indicators can exceed standards when sewage indicators are at low levels (Fig 5, indicated by quadrant IV). The majority of samples with these characteristics were in the MN and MKE Rivers, which suggests a portion of the water quality exceedances in these rivers may be attributable to sources other than sewage. In contrast, the KK river appears to have sewage as the major source of contamination. Possible nonhuman sources include pet waste and urban wildlife (in urban areas) and animal manure and wildlife (in rural and agricultural areas). Following rainfall, a clear signal from the ruminant genetic marker, with decreases in sewage markers, indicated that agricultural runoff is a likely source of fecal pollution in the MKE River watershed late in the event. While human sewage is considered the highest risk, agricultural runoff can also carry human pathogens [57].
Epidemiology studies and evaluation of outbreaks have identified an association between rain events and GI illness, particularly in children [4,6,58,59]. Release of untreated sewage is the major pathway for introduction of waterborne pathogens into the environment, creating exposure routes through recreational and drinking water. In our study area, drinking water is drawn from Lake Michigan several kilometers from the harbor and intakes are at a depth of approximately 20 meters, so any contamination in source water is highly diluted; however, communities that draw their drinking from rivers near urban areas may have higher concentrations of sewage contamination. Furthermore, drinking water treatment is designed to remove pathogens to levels safe for consumption, but the Milwaukee Cryptosporidium outbreak of 1993 illustrates the consequences of failures in this protective barrier [9].
Drinking water distribution systems may be more the more likely route by which humans are exposed to pathogens from sewage. Release of untreated sewage through stormwater systems is an indicator of sewage exfiltration from failing sanitary sewer pipes [21,60]. This leaking sewage can infiltrate drinking water distribution pipes under conditions of low water pressure or when there is a water main break [61]. Sewage can also contaminate groundwater that is used as a drinking water source, which is of high concern [18,19,59,62], particularly if it is untreated [63]. Drinking water systems, along with wastewater and sewer conveyance infrastructure, are ranked as a D− and D, respectively, by the American Association of Civil Engineers [64], and can be expected to deteriorate further over time without significant investments.
Exposure to pathogens through recreational water is not trivial, which is illustrated by a study that estimates there are 90 million cases of illness related to recreational contact with contaminated water per year [14]. Urban beaches are often located near river discharge that can impact those sites, especially following sewer overflows [55,65]. Kayaking and rowing in urban rivers is also becoming more popular, but these waterways often do not have monitoring or water quality advisory systems in place. After rainfall, concentrations of the human-associated indicator HB ranging from 4,200 to 7,800 CN/100 mL have been estimated to be equivalent to a 0.03 risk of illness following typical recreational exposure [44,57,46]. Concentrations and loads during CSO events were approximately 10-fold higher than rain events with no sewage overflows, and studies have noted that a 1:30 dilution of CSO water still presents a serious health risk [54]. Delivery of diluted CSO-contaminated water to nearby beaches is of high concern since culturable indicators may be short lived and contamination may go unrecognized using standard beach monitoring methods [65].
Linking pathogen levels in the environment to human health outcomes is complex and not feasible to do within a single study framework. Pathogens are generally at low levels and intermediately present in the environment, making them difficult to quantify in exposure pathways. In this study, source-specific indicators of sewage helped fill this gap and can infer environmental concentrations of waterborne pathogens [30,44]. These indicators can be used to assess recreational waters and surface waters; however, documenting sewage intrusion into drinking water distribution systems is not practical because contamination occurs infrequently and sporadically in small segments of the system [20]. Large epidemiology studies have linked viruses in tap water with illness in communities served by untreated groundwater, but measurements were taken over a fixed 12-week period and did not focus on rainfall events [63]. Multiple studies have linked rainfall occurrence with illness, which in and of itself is challenging. Most health outcomes are based on either outbreak data [4] or hospital or clinic visits [58,59], but the major health effects are likely sporadic occurrences of GI illness, which usually go unreported [1,5]. Quantitative microbial risk assessment can bridge some of these gaps and provide insights into intermediate risk factors through determining levels of sewage contamination in exposure pathways. Better characterization of actual exposure rates and adverse outcomes needs to be linked with pathogen burdens in the environment to fully understand health outcomes due to rain events.
GI illness has been shown to increase in the community following rainfall [4,6,58,59]. We demonstrated that sewage contamination, which carries many GI pathogens, is widespread in urban waterways following rainfall and 10-fold higher following CSOs. Human exposure could be reduced by limiting contact with recreational waters and following boil water advisories when they are issued due to water main breaks or other breaches in drinking water systems. Furthermore, vulnerable populations such as those that are immunocompromised should be cautious about exposure to surface waters after rainfall. Sewage contamination was related to the degree of urbanization in the watershed, illustrating the widespread nature of urban sewer infrastructure problems. Urban sewer infrastructure is currently under stress during rainfall due to deterioration of pipes and legacy combined sewer systems and may be more vulnerable in the future with changing rainfall patterns under climate change conditions. Future investments in repairing these systems and public health messages that are informative about potential exposure could reduce the endemic waterborne illness burden due to sewage contamination.
|
10.1371/journal.pcbi.1003894 | Some Work and Some Play: Microscopic and Macroscopic Approaches to Labor and Leisure | Given the option, humans and other animals elect to distribute their time between work and leisure, rather than choosing all of one and none of the other. Traditional accounts of partial allocation have characterised behavior on a macroscopic timescale, reporting and studying the mean times spent in work or leisure. However, averaging over the more microscopic processes that govern choices is known to pose tricky theoretical problems, and also eschews any possibility of direct contact with the neural computations involved. We develop a microscopic framework, formalized as a semi-Markov decision process with possibly stochastic choices, in which subjects approximately maximise their expected returns by making momentary commitments to one or other activity. We show macroscopic utilities that arise from microscopic ones, and demonstrate how facets such as imperfect substitutability can arise in a more straightforward microscopic manner.
| Dividing limited time between work and leisure when both are attractive is a common everyday decision. Rather than doing one exclusively, humans and other animals distribute their time between both. Traditional explanations of this phenomenon have studied the macroscopic average times spent in both. By contrast, we develop a microscopic framework in which we can model the real-time decisions that underpin these averages. In the framework, subjects' choices are approximately optimal, according to a natural, microscopic, utility function. We show that the assumptions of previous theories are not necessary for partial allocation to be optimal, and show possibilities and limits to the integration of macroscopic and microscopic views. Our approach opens new vistas onto the real-time processes underlying cost-benefit decision-making.
| When suitably free, humans and other animals divide their limited time between work, i.e., performing employer-defined tasks remunerated by rewards such as money or food, and leisure, i.e., activities pursued for themselves that appear to confer intrinsic benefit. The division of time provides insights into these quantities and their interaction, and has been addressed by both microeconomics and behavioral psychology.
Microeconomic labor supply theorists [1] have adopted a normative perspective, formulating what a rational agent should do. Accounts from behavioral psychology have been descriptive, detailing how subjects allocate their time, for example, proportionally to the relative payoffs from work and leisure [2]–[8]. Common to these approaches is the coarse, macroscopic timescale at which behavior is characterised, focusing on average times spent in work and leisure. By contrast, microscopic analyses characterise the fine temporal topography of work and leisure choices, and so offer a foundation for examining, rather than averaging away, rich psychological and neural processes. Tying microscopic and macroscopic choices together is known to be difficult in general [9], because the former involves a much more elaborate state space than the latter.
Here, we build an approximately optimal stochastic control theoretic model of decision-making at a microscopic level. We show how averaging over the microscopic choices yields a characterizable superset of traditional macroscopic theories, and casts the assumptions necessary for the latter to capture partial allocation in a different light. We make the novel prediction that partial allocation requires neither stochastic choices (as generally assumed by accounts from behavioral psychology) nor the marginal utility of leisure to depend on the amount of work performed. We use a simplification of a particularly stark labor task as a paradigm example to show how macroscopic and microscopic theories of the partial allocation of time between work and leisure can be tied. We therefore do not attempt to model actual data from this task; a qualitative account is available in [10].
We consider a Cumulative Handling Time task [11], [12] in which subjects must accumulate work up to a total time-period called the price (see Table 1 for a list of symbols and their meanings) to gain a reward. The price and the objective strength of the reward are defined by the experimenter. Note that the price is an experimenter determined time-period, hence we shall use “long” and “short” to denote its duration. Subjects are free to distribute leisure bouts in between work bouts (Fig.S1A). The CHT controls both the (average) minimum inter-reward interval and the amount of work required to earn a reward. This makes the CHT a generalisation of common schedules of reinforcement such as Fixed Ratio, or Variable Interval, which control one but not the other.
Reward and leisure are both assumed to enjoy a subjective worth. We call these microscopic utilities to distinguish them from the macroscopic utilities used by traditional theories. The microscopic utility of the former is called the (subjective) reward intensity (, in arbitrary units); the ratio of this to the price is called the payoff (or in economic nomenclature, wage rate) . For simplicity, we consider the objective price, recognising that its subjective value may differ. We explore different functional forms for the presumed microscopic utility of leisure.
This paradigm was originally developed in the context of rats pressing down an unweighted lever to gain non-satiating, brain stimulation reward (BSR), or alternatively choosing leisure in the form of resting, grooming, exploring, etc. However, as noted above, we do not model data, but rather consider an abstracted version of the task in order to concentrate on the relationship between microscopic and macroscopic descriptions.
The key macroscopic statistic is the Time Allocation (): the proportion of trial time that the subject spends working [2]. Fig.S1B shows example TAs for a typical subject. As expected, the TA increases with reward intensity and decreases with price. A microscopic analysis, as shown by ethograms in (Fig.S1C), considers the detailed temporal topography of choice, recording when and for how long each act of work or leisure occurred. Note that at intermediate payoffs, when partial allocation is most noticeable, subjects consume almost all leisure immediately after getting a reward, and then work continuously for each entire price [13].
Labor supply theory and generalized matching average over the temporal topography shown in Fig.S1C). By contrast, we follow [10], [18], [19] in formulating a so-called micro Semi-Markov Decision Process (SMDP) [20], [21] (Fig. 3A) with actions, states, and utilities, for which policies (i.e., the stochastic choices of actions at states) are quantified by the average reward per unit time accrued over the long run. We formulated the general normative, microscopic theoretical framework in [10]. Here we delineate a simplified model pertinent to the partial allocation problem.
By integrating the microscopic choices from our model, we can compare it with macroscopic descriptions such as the mountain model. We saw that linear generates partial allocation with stochasticity. It therefore generates smooth (non-step function) macroscopic time allocation curves as a function of both reward intensity and price. Consequently, 3-dimensional relationships can be derived that are qualitatively similar to those specified by the mountain model (when expressed in terms of subjective reward intensity, compare Fig. 6A with Fig. 5).
However, when is non-linear, more complicated structures arise. If the price is increased while holding the reward intensity fixed, the reward rate (Eq. (2)) decreases hyperbolically and eventually asymptotes (Fig.7A). Consequently, unlike the mean, the mode of the gamma-like distribution does not substantially increase with the price (see Figs.3C and 7B). Since the mode determines the duration of the majority of leisure bouts, these do not increase substantially. If the subject continues to work for the entire price duration (Fig.7C), then, surprisingly from the macroscopic perspective of the generalized matching model, the total work time and thus the TA will increase, rather than decrease with the price (Figs.6B and 7A, lower panel). This prediction is readily amenable to experimental test.
Since for linear , leisure durations are governed by substantially changing means and not modes, TAs are in general smaller than for strictly concave , implying that higher payoffs are necessary to capture the entire TA range.
We studied the problem of partial time allocation – when reward intensities and prices are not extreme, both animals and humans divide their time between work and leisure. Traditional theories such as the microeconomic theory of labor supply, or accounts from behavioral psychology based on the generalised matching law, have characterised behavior at a macroscopic level, studying average times spent in work or leisure. While labor supply approaches have studied choices within periods of time, these have been limited to maximising utility within these time windows [32]–and thus, still average times within these windows. We proposed a normative, microscopic approach using the reinforcement learning framework of Semi-Markov Decision Processes. Although we applied it to the labor-leisure tradeoff, this is actually a more general theoretical framework for temporally relevant decision-making. By integrating the microscopic choices of our model over time, we were able to account for the nature of macroscopic partial allocation.
We showed how assumptions about microscopic and macroscopic quantities relate. In labor supply theory, the marginal utility of leisure may (although not necessarily) depend on the amount of work (or rewards) consumed, and (unlike in the behavioral data) choices are classically deterministic. We considered a stochastic policy of the same form as emerges for standard random utility models, but directed at microscopic, rather than macroscopic, choices. Macroscopic random utility theory considers stochasticity to be due to unobservable noise, which is added to the representation of utility. The subject chooses the combination of cumulative work and leisure times that maximizes this net utility (including the noise term). If the noise is assumed to be Gumbel distributed (i.e. drawn from an extreme value distribution of type I), then the probability of choosing the optimal combination is a softmax. The softmax function that we employ is over microscopic durations, and arises from an (equivalently arbitrary) assumption that subjects have a taste for entropic policies. Randomness is thus directly built into the fabric of our model, rather than being an afterthought. It generates partial allocation even when the marginal microscopic utility of leisure is independent of work.
Previous exercises attempting to link macroscopic static and dynamic frameworks have not been generally successful [9]. Optimal choice in a dynamic context generally depends on the microscopic state, whose evolution is invisible at a macroscopic level. This allows the macroscopic average choice obtained after integrating out such states (i.e., the average choice under the stationary distribution) to appear counterintuitive, possibly even violating rationality constraints. In our case, the key feature of the microscopic state is implicit in the non-memorylessness of the policies allowed in an SMDP – e.g., that the hazard function governing the probability a leisure bout will end a certain time after it begun is not independent of time.
An example of the problems comes from observing that time allocation to working under conventional macroscopic labor supply accounts generally increases with reward and decreases with price. Something similar is true of the macroscopic, mountain-like, consequence of generalized matching. We showed in our framework that, although this can be true, it is nevertheless the case that for certain non-linearities, the time allocated to working can increase rather than decrease as the price increases, yielding complicated 3-dimensional relationships and non-monotonic contours that elude the mountain model. We thus derived a transparent link between microscopic and macroscopic frameworks. Whereas animals have been previously shown consistently to work more when work-requirements are greater (one idea is that this arises from sunk costs [33], [34]), the apparent anomaly discussed here only occurs at longer prices and is due to the form of the microscopic utility of leisure. This is an obvious candidate for empirical investigation [35].
Non-linear benefit of leisure functions can also lead to partial allocation for deterministic choices. This applies even for functions that differ from those common in labor supply theory in virtue of satisfying independence between the microscopic utilities of working and engaging in leisure. Of course, the marginal microscopic utility of leisure might depend on work or rewards – for instance due to fatigue or satiation. However, carefully eliminating such dependencies (by, e.g., allowing subjects sufficient rest inbetween trials, and using non-satiating rewards like BSR) may provide an avenue to quantify aspects of the microscopic utility of leisure empirically. This should help reveal why and how subjects partially allocate their time. It would then be natural to extend the study to considerations of effort, fatigue and cognitive computational costs [36]–[40] (e.g. from holding down weighted levers or performing cognitively demanding tasks) and the effects of manipulating motivational state [12], [41], [42]. It is by taking advantage of the greater precision available from the detailed topography of work and leisure that we may hope to gain insight into these most important details. Although previous work has described aspects of this topography [37], [43], our precise control theoretic formalization could offer enrichment.
The utilities considered in macroscopic labor supply theory are ordinal, whereas the microscopic utilities used in our framework are cardinal and, by analogy with quantities investigated in discrete choice paradigms [22]–[24], open for direct neural investigation. One of the key goals of our work is to provide a formal framework within which this can happen.
Finally, our work provides a foundation for studying critical psychological processes and neural computations at an appropriate timescale. Real-time or quasi-real-time recording methods in routine use in neuroscience such as electrophysiology, large-scale imaging, or fast-scan cyclic voltammetry allow us to correlate the activity of neural populations or concentrations of neuromodulators with the execution of behaviors. Likewise, fast causal manipulations via such methods as optogenetics allow the circuits governing these behaviors to be probed in a highly selective manner. There is an evident mismatch between the microscopic timescale over which these methods operate and the macroscopic timescales over which (a) behavior has often been characterised; and (b) the quantities such as costs and benefits which underpin the pertinence of the behavior have been defined. Our normative microscopic account may therefore provide an illuminating framework within which to build explanations that span multiple levels.
See Micro-SMDP methods in Text S1.
|
10.1371/journal.pgen.1000738 | Identification of Positive Regulators of the Yeast Fps1 Glycerol Channel | The yeast Fps1 protein is an aquaglyceroporin that functions as the major facilitator of glycerol transport in response to changes in extracellular osmolarity. Although the High Osmolarity Glycerol pathway is thought to have a function in at least basal control of Fps1 activity, its mode of regulation is not understood. We describe the identification of a pair of positive regulators of the Fps1 glycerol channel, Rgc1 (Ypr115w) and Rgc2 (Ask10). An rgc1/2Δ mutant experiences cell wall stress that results from osmotic pressure associated with hyper-accumulation of glycerol. Accumulation of glycerol in the rgc1/2Δ mutant results from a defect in Fps1 activity as evidenced by suppression of the defect through Fps1 overexpression, failure to release glycerol upon hypo-osmotic shock, and resistance to arsenite, a toxic metalloid that enters the cell through Fps1. Regulation of Fps1 by Rgc1/2 appears to be indirect; however, evidence is presented supporting the view that Rgc1/2 regulate Fps1 channel activity, rather than its expression, folding, or localization. Rgc2 was phosphorylated in response to stresses that lead to regulation of Fps1. This stress-induced phosphorylation was partially dependent on the Hog1 MAPK. Hog1 was also required for basal phosphorylation of Rgc2, suggesting a mechanism by which Hog1 may regulate Fps1 indirectly.
| When challenged by changes in extracellular osmolarity, many fungal species regulate their intracellular glycerol concentration to modulate their internal osmotic pressure. Maintenance of osmotic homeostasis prevents either cellular collapse under hyper-osmotic stress or cell rupture under hypo-osmotic stress. In baker's yeast, the Fps1 glycerol channel functions as the main vent for glycerol. Proper regulation of Fps1 is critical to the maintenance of osmotic homeostasis. In this study, we identify a pair of proteins (Rgc1 and Rgc2) that function as positive regulators of Fps1 activity. Their absence results in hyper-accumulation of glycerol and consequent cell lysis due to impaired Fps1 channel activity. Additionally, we found that these glycerol channel regulators function between the Hog1 (High Osmolarity Glycerol response) signaling kinase and Fps1, defining a signaling pathway for control of glycerol efflux. Because members of the Rgc1/2 family are found among pathogenic fungal species, but not in humans, they represent potentially attractive targets for antifungal drug development.
| Under conditions of high osmolarity stress, many fungal species, including Saccharomyces cerevisiae, maintain osmotic equilibrium by producing and retaining high concentrations of glycerol as a compatible solute [1],[2]. Intracellular glycerol concentration is regulated in S. cerevisiae in part by the plasma membrane aquaglyceroporin, Fps1 [3]–[5]. Increased external osmolarity induces Fps1 closure, whereas decreased osmolarity causes channel opening, both within seconds of the change in external osmolarity [5]. This channel is required for survival of a hypo-osmotic shock when yeast cells have to rapidly export glycerol to prevent bursting [3],[5], and is required for controlling turgor pressure during fusion of mating yeast cells [6]. The pathway responsible for regulation of Fps1 in response to changes in osmolarity has not been delineated, but appears to involve the Hog1 (High Osmolarity Glycerol response) MAP kinase [5],[7],[8]. Hog1 is activated in response to hyper-osmotic stress to mediate the biosynthesis of glycerol and perhaps its retention as well through inhibition of Fps1 channel activity. Although a hog1Δ mutant displays an elevated rate of glycerol uptake in the absence of osmotic stress, it is not impaired for Fps1 closure in response to hyper-osmotic stress [5], suggesting that Hog1 regulates the basal activity of Fps1.
Fps1 is regulated not only in response to changes in external osmolarity, but also by exposure to acetic acid [9], and in response to trivalent metalloids (e.g. arsenite and antimonite) [10],[11]. Both acetic acid and metalloids enter the cell through Fps1 and induce Hog1 activation. Fps1 is down-regulated by acetic acid treatment through ubiquitin-mediated endocytosis, which is triggered by its phosphorylation by Hog1 on Thr231 and Ser537 [9]. By contrast, metalloids down-regulate both the expression of FPS1 and its channel activity [10].
We describe the identification of a pair of paralogous S. cerevisiae proteins, Ask10 and Ypr115w that are positive regulators of glycerol efflux through Fps1. The ASK10 and YPR115w genes encode members of a family of pleckstrin homology (PH) domain proteins in yeast that includes Slm1 and Slm2 [12]. The Ask10 protein shares 41% sequence identity with its paralog Ypr115w. Although PH domains are known to bind phosphatidylinositides [13], the PH domains of Ask10 and Ypr115w are interrupted by long insertions, prompting the suggestion that they bind different ligands [12], or even serve as protein-binding domains [14].
The ASK10 gene was suggested previously to play a role in cell wall biogenesis through its identification in a genetic screen for activators of the Skn7 transcriptional regulator (Activator of SKN7) [15], which has been reported to influence cell wall assembly and cell wall stress signaling [16]–[19]. Additionally, Ask10 has been reported to be a component of the Srb/Mediator complex of RNA polymerase II [20], which is required for repression of several stress responsive genes [21],[22]. In this context, Ask10 was implicated in oxidative stress-induced destruction of the Srb11 C-type cyclin [20].
There are no reports on the function of YPR115w, or on the consequences of mutations in both Ask10 and Ypr115w. In this study, we describe the behavior of an ask10Δ ypr115wΔ mutant, finding that it displays a cell lysis defect that results from hyper-accumulation of glycerol. We find further that a defect in the function of the Fps1 glycerol channel is responsible for the ask10Δ ypr115wΔ phenotype. For this reason, we have given the name RGC1 (for Regulator of the Glycerol Channel) to YPR115w and suggest RGC2 as an alternate name for ASK10. Because the fungal kingdom is replete with members of this family of proteins, but they are not represented in animal cells, Rgc1/2 orthologs represent potentially attractive antifungal drug targets.
We constructed a double deletion mutant of RGC1 and RGC2 to test its susceptibility to cell wall stress. The double rgc1/2Δ mutant, but not the single mutants, displayed a temperature-sensitive growth defect (37°C; Figure 1A) accompanied by cell lysis, as judged by the presence of non-refractile “ghosts.” This result is in contrast to that reported by Cohen et al. [20], who found that deletion of ASK10 (RGC2) alone conferred a temperature-sensitive phenotype in the same strain background. The growth defect of the rgc1/2Δ mutant was suppressed by inclusion of sorbitol in the medium for osmotic support (Figure 1A), indicating that cell lysis is the cause of the terminal mutant phenotype. To determine if the PH domain of Rgc2 was important for its function, we tested two C-terminal truncation mutants of RGC2 for their ability to complement the rgc1/2Δ mutant cell lysis defect. The rgc2 (1–720) allele, which is missing the C-terminal 426 residues, but retains the PH domain, complemented the double mutant when over-expressed (Figure 1B). By contrast, the rgc2 (1–420) allele, which lacks the PH domain, failed to complement the double mutant. Neither allele complemented the mutant when expressed at low level from the chromosome (data not shown). This reveals that the C-terminal 426 residues are not critical to the function of Rgc2, and suggests that the PH domain contributes to its function.
Mutants that display osmotic-remedial cell lysis are typically compromised for cell wall biogenesis. To test this, we measured the rate of cell lysis of the rgc1/2Δ mutant by digestion of the cell wall with zymolyase, a wall degrading enzyme. Surprisingly, this mutant did not lyse more rapidly than the wild-type strain, but displayed slower than normal lysis kinetics (Figure 2A), suggesting that it produces a fortified cell wall. The single rgc1Δ and rgc2Δ mutants were slightly more resistant to zymolyase than was wild-type. A mutant that produces a fortified cell wall, but is nevertheless susceptible to cell lysis upon imposition of a cell wall stress may be interpreted to be under constitutive cell wall stress. We tested this by measuring the transcriptional output of the cell wall integrity (CWI) pathway. The rgc1/2Δ mutant was strongly activated for transcription of a PRM5-lacZ reporter (Figure 2B), which is under the control of the Mpk1 MAP kinase and the Rlm1 transcription factor [23]. This mutant also displayed constitutively active Mpk1, as judged by the phosphorylation state of this MAP kinase (Figure 2C). These results confirm that the rgc1/2Δ mutant experiences severe cell wall stress, to which it responds by fortifying the cell wall, and also explains its lysis defect in response to additional cell wall stress at high temperature. In further support of this conclusion, we found that the rgc1/2Δ mutant is sensitized to growth inhibition by caspofungin (Figure 2D), an antifungal drug that interferes with cell wall biosynthesis by inhibiting β-glucan synthase activity [24]. Caspofungin treatment prevents the fortification of the cell wall that is essential to the survival of this mutant.
To understand the cause of the cell wall stress in the rgc1/2Δ mutant, we conducted a dosage suppressor screen for high-copy number plasmids that conferred growth at 37°C. A single class of strong suppressor was identified as the FPS1 gene (Figure 3A). FPS1 encodes an aquaglyceroporin that is the major facilitator of glycerol uptake and efflux in yeast [3],[5]. This plasma membrane channel protein also mediates uptake of toxic metalloids, such as arsenite and antimonite [10],[11]. One interpretation of the suppression result is that the rgc1/2Δ mutant experiences abnormally high turgor pressure from accumulation of glycerol, which yeast cells use as a compatible solute for osmoregulation. Measurement of intracellular glycerol concentration confirmed that the rgc1/2Δ mutant has a 5.9-fold higher glycerol level than wild-type cells under normal growth conditions, a value that is approximately half that of an fps1Δ mutant and approximately equal to that of wild-type cells exposed to hyper-osmotic shock (Figure 3B). To determine if excess intracellular glycerol is responsible for the phenotypic defects of this mutant, we blocked glycerol biosynthesis at the first committed and rate limiting step, glycerol-3-phosphate dehydrogenase (GPD) [25]–[27]. GPD is encoded by the paralogous genes GPD1 and GPD2. Deletion of either GPD1 or GPD2 alone did not suppress the lysis defect of the rgc1/2Δ mutant, but blocking glycerol biosynthesis completely by deletion of both GPD1 and GPD2 allowed growth at 37°C (Figure 3C), confirming that glycerol accumulation is responsible for the cell lysis defect. This also provides an explanation for the fortified cell wall of the rgc1/2Δ mutant as a response to the stress of abnormally high turgor pressure. Consistent with this interpretation, the gpd1/2Δ mutations relieved the cell wall stress signaling observed in the rgc1/2Δ mutant (Figure 2B and 2C). Finally, the gpd1/2Δ mutations relieved the caspofungin sensitivity of the rgc1/2Δ mutant (Figure 2D).
We considered two possible explanations for the hyper-accumulation of glycerol in the rgc1/2Δ mutant – the mutant either produces excess glycerol, or it is impaired for glycerol efflux through Fps1. These hypotheses generate different predictions for the impact of the rgc1/2Δ mutations on the phenotype of an fps1Δ mutant. If the rgc1/2Δ mutant produces excess glycerol, this should exacerbate the lysis defect of an fps1Δ mutant, which is blocked for glycerol efflux. By contrast, if the rgc1/2Δ mutant is blocked for glycerol efflux through Fps1, loss of the glycerol channel should not result in an additive defect. Figure 3D shows that an fps1Δ mutant displays a temperature-sensitive growth defect that is slightly more severe than that of the rgc1/2Δ mutant, with a semi-permissive growth temperature of 34.5°C. The fps1Δ mutant growth defect at elevated temperature is also the result of cell lysis (data not shown). Significantly, the rgc1/2Δ fps1Δ triple mutant behaves identically to the fps1Δ mutant (Figure 3D), supporting the hypothesis that the rgc1/2Δ mutant is impaired for glycerol efflux through Fps1. To test this directly, we measured export of glycerol from cells exposed to a hypo-osmotic shock, a condition that would induce glycerol efflux through Fps1. We found that glycerol was released from wild-type cells, but not from the rgc1/2Δ mutant or the fps1Δ mutant (Figure 3E), supporting the conclusion that Rgc1 and Rgc2 regulate glycerol efflux through Fps1. The varied initial content of 14C-labeled glycerol among these mutants is a consequence of differential glycerol loading, reflecting the importance of Fps1 for glycerol influx as well as efflux. Finally, strong overexpression of RGC2 failed to suppress the temperature-sensitivity of the fps1Δ mutant (data not shown), thus establishing an epistatic relationship that places RGC1 and RGC2 above FPS1 in a common pathway for glycerol efflux.
To explore the mechanism by which Rgc1/2 regulate Fps1, we first examined the Fps1 protein level in an rgc1/2Δ mutant. Despite the observed defect in glycerol efflux of the rgc1/2Δ mutant, this mutant maintains strongly elevated Fps1 protein levels as compared to wild-type (increased approximately 10-fold; Figure 4A), suggesting that it attempts to compensate for impaired Fps1 function by increasing the number of channel proteins. The increase in Fps1 protein is a consequence of elevated glycerol concentration resulting from the rgc1/2Δ mutation, because the Fps1 level was reduced in an rgc1/2Δ gpd1/2Δ mutant (Figure 4A), which is blocked for glycerol production. The increased steady-state level of Fps1 in the rgc1/2Δ mutant is not the result of transcriptional induction, because FPS1 was expressed under the control of a heterologous promoter (MET25). This conclusion was supported by the finding that expression from an FPS1-lacZ reporter was not altered in the rgc1/2Δ mutant (Figure 4B). An even greater increase in Fps1 protein level in the rgc1/2Δ compared to wild-type (approximately 20-fold) was observed when FPS1 was expressed from its native promoter on a multi-copy plasmid (Figure 4C). Evidently, ectopic overexpression of FPS1 suppresses the rgc1/2Δ lysis defect by assisting the cell in its efforts to enhance glycerol efflux through an impaired channel. Under these conditions the cells retain more than 20-fold higher levels of Fps1 protein than wild-type cells (the comparison in Figure 4C was to wild-type cells also expressing FPS1 from a multi-copy plasmid). Therefore, we conclude that Fps1 channels in the rgc1/2Δ mutant retain less than 5% of normal activity.
To determine the cause of the increased steady-state level of Fps1 in the rgc1/2Δ mutant, we conducted a test of Fps1 stability. Fps1 levels were followed in cells in which FPS1 transcription was shut down with the simultaneous inhibition of protein synthesis. We found that Fps1 was stabilized in the rgc1/2Δ mutant relative to wild-type cells (Figure S1). Therefore, we conclude that increased intracellular glycerol in the rgc1/2Δ mutant, which is caused by a deficiency in Fps1 function, induces an increase in the level of weakly functional Fps1 through protein stabilization.
We conclude further that, because the rgc1/2Δ mutant does not display diminished Fps1 levels, Rgc1/2 positively regulate Fps1 function by a mechanism other than increased protein level. Fps1 migrates as a doublet as a consequence of phosphorylation [11], although the responsible protein kinase has not been identified. It is interesting to note that the more slowly migrating band (the phosphorylated form) predominates in the rgc1/2Δ mutant (Figure 4A).
Both Rgc1 and Rgc2 have been reported to reside in the cytoplasm [28]. If these proteins function as activators of the Fps1 glycerol channel, they might be expected to interact with Fps1 at the plasma membrane. We examined the intracellular localization of Rgc2-GFP2 in response to hypo-osmotic shock, conditions in which the Fps1 channel must be opened to allow glycerol efflux. Figure 5A shows that under unstressed conditions, Rgc2-GFP2 displays diffuse cytoplasmic localization, but very rapidly aggregates into punctate spots that appear near the cell surface in response to hypo-osmotic shock. These spots dissipate over a period of approximately 45 seconds (Figure 5B). Fps1 has been reported to reside in punctate spots at the plasma membrane [5]. Therefore, we asked if Rgc2-GFP2 co-localizes with Fps1-tdTomato in response to hypo-osmotic shock. Figure 5C shows that these spots do not co-localize. Other efforts to detect physical interaction between Rgc2 and Fps1 (e.g. co-precipitation and two-hybrid analyses; data not shown) failed to provide such evidence. Additionally, the number, location, and intensity of Fps1 punctate spots do not appear to be altered in an rgc1/2Δ mutant (Figure S2). This last result is difficult to understand considering that the mutant retains much more Fps1. It is possible that the fluorescent protein is preferentially cleaved from the stabilized Fps1 and digested in the vacuole. Nevertheless, the Fps1 we can detect in the rgc1/2Δ mutant appears to reside in the same location as in wild-type cells. These results, taken in the aggregate, suggest that regulation of Fps1 by Rgc1/2 is at the level of channel activity, rather than channel expression or localization.
Fps1 is a multi-pass plasma membrane protein with cytoplasmic N-terminal and C-terminal extensions that are inhibitory to channel function [5],[29],[30]. Truncation of these extensions results in constitutively open forms of the Fps1 channel. To explore the dependence of open channel character of Fps1 mutants on Rgc1/2 function, we tested their ability to allow xylitol uptake. A gpd1/2Δ mutant is very sensitive to high external osmolarity, because it cannot produce glycerol to re-establish osmotic balance. However, open channel fps1 mutants suppress this defect when hyper-osmotic shock is induced by 1M xylitol, which enters the cell only through unregulated Fps1 to restore osmotic balance [30]. Although a gpd1/2Δ mutant expressing wild-type FPS1 grew very poorly in the presence of xylitol, two Fps1 open channel mutants, one with an N-terminal truncation (fps1-Δ1, produces Fps1Δ12–231) [5], the other with a C-terminal truncation (fps1-C 1 produces Fps1Δ534–650) [30], conferred growth on xylitol to a gpd1/2Δ mutant even in the absence of RGC1/2 (Figure 6). This result indicates that the open channel mutants of Fps1 obviate the requirement for Rgc1/2 for Fps1 function, and support the conclusion that Fps1 is properly folded and localized independently of Rgc1/2 function.
The toxic metalloids arsenite and antimonite enter yeast cells through the Fps1 channel [10],[11]. An fps1Δ mutant is therefore resistant to toxicity of these metalloids. As a further test of the role of Rgc1 and Rgc2 in the regulation of Fps1, we examined the sensitivity of mutants in these genes to arsenite. Wild-type cells were sensitive to growth inhibition by 5 mM arsenite, but both the rgc1Δ and rgc2Δ mutants were resistant to this treatment (Figure 7A). Moreover, the rgc1/2Δ double mutant was resistant to 10 mM arsenite, consistent with the additive nature of Rgc1 and Rgc2 function. These results further support the conclusion that Rgc1/2 function is required to open Fps1. Thorsen et al. [11] demonstrated that the Hog1 MAP kinase is activated in response to arsenite treatment and that Hog1 is required for control of basal Fps1 channel activity. A hog1Δ mutant was shown to display increased arsenite uptake and hyper-sensitivity to arsenite toxicity, both phenotypes being blocked by an fps1Δ mutation. Therefore, to place Hog1 within the Rgc1/2 – Fps1 pathway, we tested an rgc1/2Δ hog1Δ triple mutant for arsenite sensitivity. Like the rgc1/2Δ mutant, the rgc1/2Δ hog1Δ mutant was resistant to arsenite toxicity (Figure 7B). Suppression of the hog1Δ arsenite hyper-sensitivity defect by the rgc1/2Δ mutations indicated that Fps1 is closed in the triple mutant. These results suggest that Hog1 promotes Fps1 closure by inhibiting the action of Rgc1/2. The order of function of these pathway components was supported by the observation that the cell lysis defect of the rgc1/2Δ mutant was not suppressed by the hog1Δ mutation (data not shown).
Because epistasis analysis revealed that Hog1 acts upstream of Rgc1 and Rgc2 to oppose their function, we asked if Rgc2 becomes phosphorylated in response to stresses that lead to the opening or closing of the Fps1 channel. Cells expressing C-terminally His-tagged Rgc2 were subjected to hypo-osmotic shock, hyper-osmotic shock (with sorbitol), or arsenite treatment. Rgc2 displayed mobility shifts on SDS-PAGE in response to all three of these stresses (Figure 8A), presumably reflecting post-translational modifications. The treatments that result in Fps1 closure (arsenite and hyper-osmotic shock) induced the greatest shifts, but hypo-osmotic shock, which induces Fps1 opening, also caused a detectable band-shift. In fact, multiple bands were detectable even in Rgc2 from unstressed cells. To determine if these mobility shifts were dependent upon Hog1, we examined Rgc2 mobility in a hog1Δ mutant. The absence of Hog1 did not prevent the stress-induced Rgc2 band-shifts, but in all cases reduced the extent of shift (Figure 8A). Rgc2 from unstressed cells also displayed increased mobility in a hog1Δ mutant (Figure 8B), suggesting that Rgc2 sustains a basal level of Hog1-dependent phosphorylation.
This experiment also revealed the existence of additional modifications in response to these stresses that are Hog1-independent. To determine if these additional modifications were phosphorylations, we subjected Rgc2 isolated from stressed cells to protein phosphatase treatment. For all three stresses, phosphatase treatment collapsed the Rgc2 band to the same level as phosphatase treated, unstressed Rgc2 (Figure 8C). We conclude that although basal phosphorylation of Rgc2 is Hog1-dependent, other protein kinases are responsible for the hyper-phosphorylation observed in response to Fps1-regulating stresses.
It has been demonstrated that in the absence of Hog1, hyper-osmotic stress activates the Fus3 and Kss1 MAP kinases through inappropriate cross-talk [31]. Therefore, to determine if the Rgc2 band-shift observed in response to high osmolarity in the absence of Hog1 was due to such cross-talk, we tested a hog1Δ ste11Δ double mutant, which is blocked for activation of Fus3 and Kss1. The mobility shift observed for Rgc2 in this mutant was indistinguishable from that of the hog1Δ mutant (Figure S3), indicating that these MAP kinases are not responsible for the hyper-osmotic stress-induced phosphorylation.
Glycerol serves as a compatible solute in S. cerevisiae and other yeasts, allowing cells to respond quickly to changes in external osmolarity. A key component in the control of cytoplasmic glycerol concentration is the Fps1 glycerol channel. Although Fps1 is known to close under conditions of hyper-osmotic stress, and open in response to hypo-osmotic shock [3],[5], the mechanism by which Fps1 function is modulated is not understood. In this study, we describe a regulatory pathway for the control of this glycerol channel.
We identified a pair of paralogous genes, RGC1 (Regulator of the Glycerol Channel; YPR115w) and RGC2 (ASK10), that function as positive regulators of Fps1. The studies described reveal that loss of function of both RGC1/2 results in cell wall stress that is caused by excess turgor pressure associated with elevated intracellular glycerol concentration. The increase in glycerol is the consequence of impaired Fps1 function. We found that the increased turgor pressure experienced by the rgc1/2Δ mutant provokes the cell to activate the CWI signaling pathway and to fortify the cell wall. Nevertheless, imposing additional cell wall stress on this mutant induced cell lysis, a defect that was suppressed by blocking glycerol synthesis. In this regard, it is interesting to note that blocking the function of the glycerol channel activators also sensitized cells to caspofungin, an antifungal agent that acts by inhibiting cell wall biosynthesis [24]. Evidently, preventing the cells from responding to their internally imposed cell wall stress is lethal. Therefore, Rgc1/2 might be suitable antifungal targets for combination therapy with caspofungin.
The mechanism by which Rgc1/2 regulate Fps1 remains unclear. Although there is some evidence that Rgc2 (Ask10) can act as a transcriptional regulator (see below), we did not find that Rgc1/2 control Fps1 transcription. We were not able to detect direct interaction between Rgc2 and the Fps1 channel. However, the findings that Fps1 localizes to the plasma membrane in the presence or absence of Rgc1/2 and that constitutive mutants of Fps1 retain their open channel character independently of Rgc1/2 suggests that these proteins regulate Fps1 through its activity, rather than at an earlier step, such as protein folding, or proper localization. Rgc1/2 control of Fps1 folding or localization would be expected to impact the function of open channel mutants as well as the wild-type.
Fps1 is unusual in its possession of extensions at both its cytoplasmic N-terminus and C-terminus that play a role in regulating Fps1 channel activity [29],[30]. These extensions have been suggested to function as flaps that restrict the flow of glycerol through the channel. However, the mechanism by which they respond to changes in extracellular osmolarity remains largely unknown.
The HOG pathway is activated in response to hyper-osmotic stress [8]. Hog1, the stress-activated MAP kinase at the base of this pathway plays a poorly-defined role in the regulation of Fps1. A hog1Δ mutant exhibits a glycerol uptake rate that is approximately 3-fold-higher than that of wild-type cells [5],[11]. However, this mutant is not impaired for Fps1 closure in response to hyper-osmotic stress [5], suggesting that Hog1 regulates the basal activity of Fps1 (i.e. in the absence of osmostress), but not the osmotic stress-induced closure. Basal inhibition of Fps1 by Hog1 may result from phosphorylation at Thr231, which resides within the N-terminal extension, because Hog1 can phosphorylate this site in vitro [11], and mutation of Thr231 to Ala results in constitutive Fps1 activity [11],[29].
In addition to glycerol, the toxic metalloid arsenite enters the cell through the Fps1 glycerol channel [10]. Loss of Fps1 function confers resistance to arsenite, whereas loss of Hog1 function results in an increase in the rate of arsenite uptake through Fps1 and consequent hyper-sensitivity to the metalloid [11]. We found that null mutations in RGC1/2 also conferred resistance to arsenite, consistent with the conclusion that Rgc1 and Rgc2 are important for Fps1 channel activity. The rgc1/2Δ mutations suppressed the arsenite hyper-sensitivity of a hog1Δ mutation. In fact, loss of RGC1/2 function was completely epistatic to the hog1Δ mutation with regard to arsenite sensitivity, suggesting that Hog1 exerts its negative effect on Fps1 channel function by inhibiting Rgc1 and Rgc2.
We found that Rgc2 undergoes phosphorylation-induced band-shifts in response to various Fps1-regulatory stresses (i.e. hypo- and hyper-osmotic shock, and arsenite stress). These phosphorylations were partially dependent on Hog1, as intermediate shifts were observed in a hog1Δ mutant. Rgc2 also appears to undergo basal phosphorylation that is Hog1-dependent. The PhosphoPep database (part of the Saccharomyces Genome Database) [32] identifies 5 phosphorylation sites on Rgc1 and 10 in Rgc2 from unstressed cells. However, only one of these sites in Rgc2 (Thr808), and none in Rgc1 reside at potential Hog1 phosphorylation motifs (S/TP), suggesting that the observed Hog1-basal phosphorylation of Rgc2 is largely, or entirely indirect.
It is also possible that Hog1 inhibits basal Fps1 activity both directly, through phosphorylation of Thr231, and indirectly through phosphorylation of Rgc1/2. In any case, it is clear that other protein kinases contribute to the regulation of Rgc2 (and probably Rgc1), and consequently Fps1, in response to various stresses. These results establish a regulatory pathway from Hog1 to Rgc1/2 to Fps1, in which Rgc1 and Rgc2 are positive regulators of Fps1 channel activity and Hog1 inhibits Fps1 through inhibition of Rgc1/2. Although the interaction between these proteins and Hog1 may be direct, the phosphorylation sites on Rgc1 and Rgc2 remain to be identified.
It is possible that Rgc1/2 are multifunctional proteins. Overexpression of Ask10 was reported to enhance growth of a strain in which histidine production was under the control of (lexAop)-HIS3 reporter driven by a LexA-Skn7 fusion [15]. However, ASK10 overexpression failed to drive a similarly regulated (lexAop)-lacZ reporter. This was in contrast to the behavior of MID2, another gene identified in this screen that activated both reporters [18], raising the possibility that Ask10 does not activate Skn7-mediated transcription.
A second report, by Cohen et al. [20], suggested that Ask10 participates in the oxidative stress-induced destruction of Srb11, a C-type cyclin that is part of the Mediator complex of RNA polymerase II. These investigators identified Ask10 in a two-hybrid screen for Srb11-interacting proteins. They further demonstrated that, like Srb11 and its cyclin-dependent kinase (Srb10), Ask10 is a component of the RNA polymerase II holoenzyme. We do not know how the function of Rgc1/2 as regulators of Fps1 might relate to their reported roles in stress-activated transcription.
Rgc1 and Rgc2 are large proteins (120kD and 127kDa, respectively), and our immunoblot analysis of Rgc2 suggests that its regulation in response to different stresses that regulate Fps1 is complex. The unstressed and stressed forms of Rgc2 all migrate as several distinct bands. We have shown that these bands represent a variety of phosphorylated states of Rgc2. Although identities of many of the phosphorylation sites are not known, numerous Rgc1 and Rgc2 phosphorylation sites have been identified in response to DNA damage stress. Albuquerque et al. [33] identified 17 phosphorylation sites in Rgc1 and 20 in Rgc2 in response to treatment with the DNA alkylating agent, MMS. Additionally, as noted above, numerous basal phosphorylation sites in Rgc1 and Rgc2 are reported the PhosphoPep database [32]. Only a few of these sites overlap with those found in MMS-treated cells.
Finally, Cohen et al. [20] found that Rgc2 (Ask10) is phosphorylated in response to oxidative stress induced by hydrogen peroxide. These authors reported that the redundant MAPK kinases of the Cell Wall Integrity (CWI) signaling pathway (Mkk1 and Mkk2) were responsible for this modification. Oddly, however, none of the four MAP kinases in yeast were found to be involved [20]. We revisited this result, finding that none of the kinases within the CWI MAPK cascade (including Mkk1/2) were required for the oxidative stress-induced phosphorylation of Rgc2 (Figure S4).
Rgc1/2 may function to integrate multiple stress signals, only some of which are known to control Fps1 channel activity. The regulation of Rgc1/2 by phosphorylation in response to different stresses appears to be complex. Moreover, these proteins may have additional functions that have yet to be identified.
The S. cerevisiae strains used in this study are listed in Table 1. Yeast cultures were grown in YPD (1% Bacto yeast extract, 2% Bacto Peptone, 2% glucose) or SD (0.67% Yeast nitrogen base, 2% glucose) supplemented with the appropriate nutrients to select for plasmids. Yeast strains bearing multiple deleted genes were constructed by genetic crosses, followed by PCR-based detection of the deleted alleles. Diploid strains were used for most experiments, because the cell lysis phenotypes were more pronounced in diploids than in haploids, and also because diploids have a reduced tendency to acquire suppressor mutations.
Three different genomic clones of FPS1 were isolated from a high-copy genomic library in pRS202 (gift of P. Hieter) as suppressors of the temperature-sensitivity of a rgc1/2Δ mutant. The screen was conducted in the rgc1/2Δ mutant (DL3209) by plating transformations directly at 37°C. Plasmids were isolated from colonies arising after 3 days. A total of approximately 10,000 transformants were subjected to selection (as judged by low-temperature plating). This was calculated, based on an average insert size of 6 kb, to be approximately 5 genome-equivalents. Deletion analysis of one of these plasmids (p2165) confirmed that FPS1 was responsible for the suppression activity.
Two reporter plasmids for different transcriptional outputs were used in this study. One reporter, PRM5 (−994 to +1)-lacZ (p1366) responds to the cell wall stress transcription factor, Rlm1 [23]. The other, FPS1 (−933 to −57)-CYC1-lacZ (p2213), was constructed by PCR amplification of the 5′ non-coding region of FPS1 using primers with Xho1 (upstream primer) and Sph1 (downstream primer) sites for cloning into the Xho1 and Sph1 sites of pLG178 (p904) [34]. This placed the regulatory sequences for FPS1 upstream of the basal CYC1 promoter linked to lacZ.
The entire FPS1 gene was amplified by PCR from genomic yeast DNA (EG123 strain background) using a pair of primers 650 bp 5′ to the start codon and 500 bp 3′ to the stop codon. The primers were designed with a Not1 site (5′ primer) and a Sal1 site (3′ primer) for subcloning into pRS316 [35] to produce pRS316-FPS1 (p2833). Open channel mutant fps1-Δ1 in a multi-copy vector (YEp195-fps1-Δ1; p2496) was the gift of M. Mollapour. Open-channel mutant fps1-C1 (YEp181-fps1-C1-myc; p2829) was the gift of S. Hohmann.
The FPS1 gene, fused with a C-terminal Flag epitope, was expressed under the control of the MET25 promoter. The FPS1 coding sequence amplified from genomic DNA (EG123) with an XbaI site immediately 5′ to the initiation codon and a HindIII site immediately 3′ to the final codon and inserted into pRS426-MET25P-FLAG (p2186) so as to fuse the C-terminus with the Flag coding sequence, yielding MET25P-FPS1-FLAG (p2492). The YEpmyc181-FPS1 plasmid (p2184) was the gift of S. Hohmann).
The FPS1 gene was tagged at its C-terminus with tdTomato (red fluorescence) [36] and expressed under the control of its own promoter in two steps. First, the tdTomato coding sequence was subcloned from pRSET-B [tdTomato] (gift of R. Tsien) into pRS316 at the BamHI and EcoRI sites, yielding p2487. Next, the FPS1 gene was amplified (omitting the endogenous stop codon) from genomic DNA (EG123) and inserted into p2487 using NotI and SpeI sites designed into the primers. This fused the FPS1 reading frame with tdTomato, yielding pRS316-FPS1-tdTomato (p2489).
The RGC2 gene was tagged at its C-terminus with 6xHis and expressed under the control of the MET25 promoter. The RGC2 coding sequence was amplified by PCR from genomic yeast DNA using primers that included XbaI and XhoI sites and cloned behind the MET25 promoter in pUT36 (p2415) [37] to yield pUT36-MET25P-RGC2-HIS6 (p2501). His-tagged C-terminal truncations of Rgc2 were also expressed under the control the MET25 promoter. The first 1260 base pairs (amino acids 1–420) or 2160 base pairs (amino acids 1–720) of RGC2 were amplified from genomic DNA (wild-type strain EG123) by PCR using a forward primer that contained an XbaI site immediately 5′ to the start codon and reverse primers that introduced a 6xHis tag followed by a stop codon and an XhoI site. The two regions were inserted into pUT36, resulting in pUT36-MET25P-rgc2(1–420)-His6 (p2808) and pUT36-MET25P-rgc2(1–720)-His6 (p2809).
The RGC2 coding sequence was tagged at its C-terminus with two tandem copies of GFP and expressed under the control of the MET25 promoter in three steps. In the first step, the RGC2 promoter and coding sequence (omitting the endogenous stop codon) was amplified by PCR and inserted into the Not1 and Sma1 sites of pRS315[GPF] (p1164) [38] to yield pRS315-RGC2-GFP (p2478). In the second step, RGC2-GFP was amplified by PCR from p2478 and inserted in the same way into pRS315[GFP], to yield RGC2-GFP2 (p2479). In the final step, the RGC2-GFP2 coding sequence only was amplified by PCR and inserted into pRS414-MET25P (p976) using Spe1 and EcoRV sites designed into the primers. This yielded pRS414-MET25P-RGC2-GFP2 (p2481).
The RGC1 gene with 800 bp of upstream sequence was amplified by PCR from genomic EG123 DNA and using a forward primer that introduced a NotI site and a reverse primer that introduced a SalI site and cloned into centromeric vector pRS313 [35], yielding pRS313-RGC1 (p2627). FPS1 and RGC1/2 constructs were validated by DNA sequence analysis and all were tested for functionality of these proteins by complementation of the cell lysis defects associated with an fps1Δ mutant or an rgc1/2Δ mutant, respectively.
Zymolyase sensitivity was carried out as described previously [39]. Promoter-lacZ expression experiments for determination of cell wall stress were carried out as described previously [40], with methods for β-galactosidase assays described in Zhao et al. [41]. Intracellular glycerol concentrations were measured in whole cells grown in YPD and centrifuged briefly to remove the culture supernatant. Enzymatic assays for glycerol were carried out using a kit from Boehringer Mannheim and normalized to A600 of the initial culture. Efflux measurements of 14C-glycerol were carried out as described by Tamas et al. [5]. Briefly, cells from log-phase cultures (30 ml) grown in YPD were washed in ice-cold MES buffer (10 mM MES, pH 6.0), resuspended in 1 ml ice-cold labeling buffer solution (10 mM MES buffer + 300 mM [14C]glycerol) and incubated for 1 hour at 30°C to load cells with labeled glycerol. Cells were then diluted 10-fold in ice-cold MES buffer to induce hypo-osmotic shock. Aliquots of cells were filtered onto Whatman GFB 25 mm discs at various time points, and washed with MES buffer. Radioactivity of dried filters was measured by a scintillation counter.
For detection of total Mpk1 and activated Mpk1, protein samples (20 µg) were separated by SDS-PAGE (7.5% gels) followed by immunoblot analysis. Total Mpk1 was detected with rabbit polyclonal antibodies from Santa Cruz Biotechnologies. Activated Mpk1 was detected with rabbit polyclonal α-phospho-p44/p42 MAPK (Thr202/Tyr204) antibodies (New England Biolabs). Both primary antibodies were used at a dilution of 1∶2000. Secondary donkey anti-rabbit antibodies (GE Healthcare) were used at a dilution of 1∶5000.
For detection of Fps1-Flag, protein samples (4 µg) were separated by SDS-PAGE (7.5% gels) followed by immunoblot analysis using mouse monoclonal α-FLAG antibody (M2; Sigma) at a dilution of 1∶10,000. For detection of Fps1-Myc, protein samples (25 µg) were separated by SDS-PAGE (7.5% gels) followed by immunoblot analysis using mouse monoclonal α-Myc antibody (9E10; BabCo) at a dilution of 1∶10,000. For detection of Rgc2-His6, protein samples (16 µg) were separated by SDS-PAGE (7.5% gels) followed by immunoblot analysis using mouse monoclonal α-tetra-HIS antibody (Qiagen) at a dilution of 1∶5000. Secondary antibodies (goat anti-mouse; Amersham) were used at a dilution of 1∶5000. For protein phosphatase treatment of Rgc2-His6, Nickel NTA agarose (Qiagen) was used to precipitate Rgc2-His6 from protein extracts (100 µg) prior to treatment with calf intestinal phosphatase (CIP; Promega) with, or without phosphatase inhibitor (10 mM Na3VO4) for 1 hour at 37°C. Precipitates were processed for immunoblot detection of Rgc2-His6.
Diploid cells transformed with plasmids that express Rgc2-GFP2 with out without Fps1-tdTomato were grown in selective medium and visualized by fluorescence microscopy using a Zeiss Axioplan II with a 100x objective and fitted with a GFP and RFP filter. For hypo-osmotic shock experiments, log-phase cultures (1 ml) were centrifuged briefly to pellet cells, which were resuspended in 0.5 ml distilled water for 20 seconds to impose hypo-osmotic shock, followed by the addition of 0.5 ml 20 mM NaN3, 20 mM NaF, 20 mM Tris buffer to block further membrane transport [42] and set on ice for 20 seconds. Samples were centrifuged briefly to concentrate cells and mounted for microscopy. The membrane transport inhibitors were omitted from the time-course experiment.
|
10.1371/journal.pntd.0001976 | Non-Invasive In Vivo Study of the Trypanosoma vivax Infectious Process Consolidates the Brain Commitment in Late Infections | Trypanosoma vivax, one of the leading parasites responsible for Animal African Trypanosomosis (Nagana), is generally cyclically transmitted by Glossina spp. but in areas devoid of the tsetse flies in Africa or in Latin American countries is mechanically transmitted across vertebrate hosts by other haematophagous insects, including tabanids. We followed on from our recent studies on the maintenance of this parasite in vivo and in vitro, and its genetic manipulation, by constructing a West African IL1392 T. vivax strain that stably expresses firefly luciferase and is fully virulent for immunocompetent mice. We report here on a study where murine infection with this strain was monitored in vivo using a non-invasive method. Study findings fully support the use of this strain in the assessment of parasite dynamics in vivo since a strong correlation was found between whole body light emission measured over the course of the infection and parasitemia determined microscopically. In addition, parasitemia and survival rates were very similar for mice infected by the intraperitoneal and sub-cutaneous routes, except for a longer prepatent period following sub-cutaneous inoculation with the parasite. Our results clearly show that when administered by the subcutaneous route, the parasite is retained few days in the skin close to the inoculation site where it multiplies before passing into the bloodstream. Ex vivo bioluminescence analyses of organs isolated from infected mice corroborated our previous histopathological observations with parasite infiltration into spleen, liver and lungs. Finally, our study reinforces previous observations on the presence of the parasite in the central nervous system and consequently the brain commitment in the very late phases of the experimental infection.
| Very little work has been performed on Trypanosoma vivax for decades, but the recent development of murine infection models and axenic cultures has enabled the genetic manipulation of this parasite and has opened the door to a more in-depth understanding of its biology and the infectious process that leads to animal trypanosomosis. We report herein the characterization of a luciferase-expressing strain that can be used to follow parasite dynamics in vivo in real time using a non-invasive method. Regardless of the inoculation parasite route and some minor differences concerning the length of the prepatent period of infection following the subcutaneous injection of the parasites, we highlight the general commitment of the organs triggered by the infection and particularly the presence of the parasite in the brain at late phases of disease. The study presented herein provides new insights into the interaction between T. vivax and its mammalian host and assesses new tools for in vivo drug screening.
| Animal African trypanosomosis (AAT) is a major protozoan disease due to trypanosomes. The disease which is endemic in Africa is mainly caused by Trypanosoma vivax, T. congolense and T. b. brucei. T. vivax accounts for up to half of all AAT prevalence in West Africa where it is considered to be the major pathogen that together with T. congolense causes 3 million cattle deaths annually [1]–[3]. Furthermore, T. vivax but also T. equiperdum and T. evansi trigger different pathologies (Nagana, Dourine and Surra, respectively) and are species that have spread to South America. Globalization and livestock trade between countries, coupled with the lack of rapid diagnostic tools and the transport of infected animals to non-endemic areas have a huge impact on agriculture and, consequently, on the economy of breeding and endurance.
One of the specificities of T. vivax compared to other animal trypanosomes (i.e. T. brucei spp and T. congolense) is its ability to be transmitted not only by Glossina spp. (tsetse) flies but also by other biting flies of the Tabanidae and Muscidae families that can mechanically transmit the parasite among mammalian hosts [4], [5]. It is noteworthy that Glossina spp. are the only vectors in which T. vivax is able to multiply and pursue its differentiation into metacyclic forms. In contrast, T. vivax is unable to grow or multiply in other insects that can only mechanically transmit the parasite. Regardless of the natural type of transmission (cyclical or mechanical), T. vivax is inoculated in the subcutaneous tissue and the infective forms join the bloodstream via the lymphatic system. After one or more parasitemia peaks, the animals generally show neurological disorders in late phases of infection and perish [6], [7]. Ruminants and equines infected with T. vivax show a range of tissue damage and the diversity of the pathognomonic signs and the severity of the disease frequently correlate with the degree to which the host shows resistance (“tolerance”) or susceptibility to the parasite. Few studies have been conducted to compare the infective process following a bite by tsetse or tabanids, or experimental infections by intraperitoneal or subcutaneous inoculation routes [8].
In efforts to overcome the problems encountered when studying T. vivax infection and pathology in the field, we recently developed murine models that deliver sustained and reproducible infections which successfully mimic the parasitological, histological and pathological features of the infection and closely resemble those observed in cattle trypanosomosis [9], [10]. For instance, histopathological examinations performed throughout the infective process showed many necrotic foci in lymphoid and non-lymphoid organs with extravasated blood cells and trypanosomes in hemorrhagic spots. Most importantly, the infection resulted in multifocal lesions in the central nervous system along with vasogenic edema and damaged blood vessels characteristic of the late-stage ischemic necrosis caused by the wild-type strain. Although the presence of trypanosomes in the meningeal blood vessels at these late stages was suggestive of blood-brain barrier crossing and invasion of cells and parasites into the brain parenchyma [9], our knowledge of the invasive characteristics of T. vivax, its tissue tropism, the temporal course of its invasion and the crucial question of the permeabilization of the blood brain barrier is still incomplete.
In order to address some of these questions and supplement the conventional anatomic pathology examinations conducted during studies of the infectious process, we took full advantage of the latest advances made in T. vivax genetic manipulation [11] and engineered a parasite strain that stably expresses firefly luciferase. Here we report on the in vitro and in vivo characterization of the T. vivax luciferase strain and the validation of real-time biophotonic detection systems employed to study the propagation of this parasite in vivo. We determined method limits of detection and linearity ranges to better correlate mouse parasitemia with luminescence measured in vivo, and analyzed the course of the infection and parasite tissue distribution over time. Finally, we compared infection dynamics and organ commitment after subcutaneous and intraperitoneal inoculations with the parasite. Our results confirmed the usefulness of real-time biophotonic analysis in the study and monitoring of the T. vivax infectious process in vivo. Irrespective of the causes that have conducted each mouse to perish during early phases of infection, such as anemia, hyperparasitemia or organ failure, our data provide important evidence that at long-term trypanosomes attain the central nervous system of all the animals which have showed a extended survival just some days before death.
All mice were housed in our animal care facility in compliance with European animal welfare regulations. Institut Pasteur is a member of Committee #1 of the Comité Régional d'Ethique pour l'Expérimentation Animale (CREEA), Ile de France. Animal housing conditions and the procedures used in the work described herein were approved by the “Direction des Transports et de la Protection du Public, Sous-Direction de la Protection Sanitaire et de l'Environnement, Police Sanitaire des Animaux” under number B 75-15-28, in accordance with the Ethics Charter of animal experimentation that includes appropriate procedures to minimize pain and animal suffering. PM is authorized to perform experiments on vertebrate animals (license #75-846 issued by the Paris Department of Veterinary Services, DDSV) and is responsible for all the experiments conducted personally or under her supervision as governed by the laws and regulations relating to the protection of animals.
Trypanosoma (Dutonella) vivax IL 1392 was originally derived from the Zaria Y486 Nigerian isolate [12]. These parasites have recently been characterized and are maintained in the laboratory by continuous passages in mice, as previously described in detail [9]. Seven to ten week-old male Swiss Outbred mice (CD-1, RJOrl:SWISS) (Janvier, France) were used in all experiments. They were injected intraperitoneally or sub-cutaneously with bloodstream forms of T. vivax (102 parasites/mouse). Parasitemia was determined as previously described [9]. All animal work was conducted in accordance with relevant national and international guidelines (see above).
A luciferase assay kit (Roche Molecular Biochemicals; Mannhein, Germany) was used to monitor luciferase expression. Serial dilutions of parasite suspensions were washed in PBS and pellets were suspended in 150 µl of cell lysis buffer. The lysates were then transferred into white, 96-well microplates (Dynex Technologies, Chantilly, France). Light emission was initiated by adding the luciferin-containing reagent, in accordance with manufacturer instructions. The plates were immediately transferred to the luminometer (Berthold XS3 LB960; Thoiry, France) and light emission was measured for 0.1 s. Luminescence was expressed in Relative Light Units (RLU).
Mice were inoculated intraperitoneally with luciferin (D-Luciferin potassium salt, Xenogen, California), the luciferase substrate, at a dose of 150 mg/kg before any bioluminescence measurements were made. They were anaesthetized in a 2.5% isoflurane atmosphere (Aerane, Baxter SA, Maurepas, France) for 5 minutes and kept in the imaging chamber for analysis. Emitted photons were acquired for 1 minute by a charge couple device (CCD) camera (IVIS Imaging System Lumina, Caliper, Villepinte, France) set in high resolution (medium binning) mode. The analysis was then performed after defining a region of interest (ROI). The same ROI was used for all animals and all time points. Total photons emitted from the image of each mouse were quantified using Living Image software (Xenogen Corporation, Almeda, California), and results were expressed as number of photons/sec/ROI.
The methods used to engineer this T. vivax strain that stably expresses firefly luciferase (TvLrDNA-luc) have been described elsewhere [11]. The infective forms of these recombinant parasites maintain their infectivity in immunocompetent mouse strains and show the same parasitemia profiles over time and result in similar levels of mortality as wild type (WT) T. vivax. Luciferase expression levels in non infective epimastigote axenic forms (EPI) and in infective bloodstream trypomastigote (BSF) forms of TvLrDNA-luc were compared by measuring the luciferase activity of equivalent numbers of parasites purified from axenic cultures or mouse blood, respectively. Serial dilutions of EPI and BSF were washed, lysed and the extract supernatants assayed in parallel for in vitro luciferase activity by measuring relative light emission (RLU) initiated by adding luciferin substrate. Figure 1A illustrates the linearity of the RLU results over more than 3 logs for both EPI and BSF TvLrDNA-luc extract supernatants. Since no bioluminescence was detected in the WT EPI and BSF parasites included in the assays (<50 RLU), these results indicate that this light emission is specific to bioluminescent (luciferase-expressing) parasites.
Limits of detection were about 300 EPI and 8000 BSF using the bioluminescence assay. EPI clearly gave 7 to 10 fold the luciferase activity of purified BSF. To check that BSF serial passages in vivo do not result in any loss of the construction carrying the luciferase and the resistance marker genes, TvLrDNA-luc parasite strain was maintained in vivo without drug pressure for 3, 7 or 15 sequential passages and compared for light emission. Figure 1B shows that light emission was comparable whatever the number of passages in vivo. The results obtained confirmed that the differences observed between EPI and BSF were not due to any in vivo loss of the construction. These differences therefore may reflect dissimilar stabilities of the enzyme at different temperatures or, like for other trypanosomatids [13], may stem from the distinct morphometries of EPI and BSF which are compatible with their size, nucleic acids and protein contents. Altogether, plasmid integration in the ribosomal region of T.vivax was shown to be stable over at least 15 consecutive passages, corresponding to 15 weeks.
Parasite dissemination and disease progression in a T. vivax-infected mouse model previously studied in the laboratory, was followed by using sensitive, non-invasive optical imaging to track TvLrDNA-luc parasite strain in live mice. The bioluminescent signal obtained in vivo with TvLrDNA-luc parasites was validated by considering criteria such as background spontaneous signals obtained from whole-body images of mice infected with WT parasites and optimal time period between substrate administration to live animals and image capture. Firstly, a group of mice were infected with WT parasites, injected with D-Luciferin and subsequently exposed to photon detection under the IVIS Lumina Imaging System (IVIS) to determine background emissions across the entire body. This resulted in 106 ph./s being considered as the background level for further in vivo experiments using the TvLrDNA-luc parasite strain (not shown). Secondly, another group of mice were infected with TvLrDNA-luc parasite strain and analyzed once the infection had resulted in moderate parasitemia (106 parasites/mL). Total body light emission was determined under the IVIS at time points 1, 3, 5, 10, 15 and 20 minutes after D-Luciferin injection. As can be seen in Figure 2A, light detection was maximal between 5 and 10 minutes after substrate injection and did not show any major variations up to 20 minutes after the injection. Further in vivo measurements were then made 10 minutes after the D-Luciferin injection. Lastly, we checked whether or not the light emitted correlated with the T. vivax infectious process in vivo. To do this, we determined whether or not the parasitemia observed microscopically correlated with the bioluminescence measured over the whole animal. Parasites were counted under a light microscope in five microliters of blood harvested individually from the tail vein of mice infected with 102 TvLrDNA-luc parasites, and parasitemia was expressed as number of parasites per mL of blood. Mice were immediately injected with D-Luciferin and submitted to whole-body imaging. An extensive and increasing light emission was observed during the course of the infection, as shown in Figure 2B. Total bioluminescence increased in the course of the infection and in line with the parasite count obtained optically, reaching ph./s levels that were more than 1000 fold the background level with parasitemia of 108 parasites/mL. The exhaustive plotting of bioluminescence versus parasitemia depicted in Figure 2B represents 35 individual measurements obtained in a group of 15 mice and shows a close correlation between the 2 parameters, as confirmed by a high Spearman rank correlation coefficient of 0.9365. These findings validated the TvLrDNA-luc bioluminescent strain and the baseline imaging parameters necessary to analyze and monitor T. vivax infection and disease progression in vivo.
In order to gain a clearer insight into disease progression, and in particular determine whether T. vivax multiplication is confined to the vascular compartment, we compared the parasite dissemination after infection by two different routes. The conventional intraperitoneal (IP) route commonly used for mouse experimental infections was compared to the subcutaneous (SC) inoculation route that closely resembles the natural infections, cyclic or mechanically conveyed by the insects. The course of the resulting infection together with parasitemia and survival rates were therefore studied in mice infected subcutaneously or intraperitoneally with 102 TvLrDNA-luc BSF parasites. As can be seen in Figures 3A and 3B, no substantial differences were observed for parasitemia between the two groups. As expected, a straightforward correlation was found between the numbers of parasites determined optically and the signals resulting from bioluminescent parasites. The only difference found between the groups was the length of the prepatent period preceding the microscopically-detectable parasitemia (104 parasites per ml of blood). As shown in Figure 3B, mice infected by the subcutaneous route showed a longer prepatent period (9 to 10 days) and consequently a more delayed onset parasitemia than mice inoculated by the intraperitoneal route that presented detectable parasites 5 days after infection (Figure 3A). Similar data were obtained after experimental infections with T. congolense, as reported previously [8]. With the exception of this time lag, parasite multiplication was seen to follow the same kinetics with both routes of infection and parasitemias were invariably similar after day 15 of infection. Accordingly, survival rates during infection were not significantly different between mice injected IP or SC, with 30% of the mice dying by day 15 p.i., 40% between days 20 and 22 and the remaining dying by day 28 post-infection (Figure 3C).
Interestingly, while detectable light emission in IP-infected mice invariably correlated with parasite appearance in peripheral blood, we noted that light emission was already detectable in SC-infected mice in the prepatent period. In efforts to investigate this phenomenon, we infected a group of mice with 102 TvLrDNA-luc by the SC route and monitored bioluminescence every day during the prepatent period. Light emissions were seen to increase between days 8 and 9 post-infection, suggesting that the parasite load was increasing but remained below the limit of detection of the microscopy visualization technique (<104/mL). By the time parasitemia had become detectable (day 9–10), the mice showed a bright spot on the right lateral flank close to the injection site (Figure 4A) that gradually increased thereafter (Figure 4B). The mice were sacrificed on day 10 for gross anatomy and a representative mouse is shown in Figure 4C. As can be seen, the light emission is confined to the skin near the SC inoculation site. The increase in light emissions in this area and the very circumscribed foci of photons shows that the infection initially develops in situ and that parasite multiplication takes place in the skin close to the injection site before parasites reach the bloodstream.
T. vivax infection was followed in vivo by inoculating new groups of mice by the IP route with 102 TvLrDNA-luc BSF and following the infection by biphotonic analysis. Mouse parasitemias were measured individually both microscopically and by light emission, and as described here above, the two techniques gave comparable results. Groups of at least 3 mice were analyzed at each time point and bioluminescence recorded individually. Light became detectable 5 days after infection and at this point was 4 fold background levels in non infected control mice. These observations correlated with very low parasitemias (1–2×104 parasites/mL), as determined microscopically. Figure 5A shows the results obtained with one infected mouse representative of a group of 3 mice examined by time of infection as compared to a uninfected control during the study period of 4 weeks. The first foci observed were located at the muzzle and the inguinal regions. Once parasitemia increased (105–106 parasites/mL), photons were detected along the entire body, with hotspots corresponding to spleen, lungs and liver. Attempts were made to accurately define the dynamics of parasite dissemination by segmenting the images obtained for each animal into several areas corresponding to whole body (R), head (R1), thoracic region (R2), abdomen (R3), inguinal (R4) and testis areas (R5) (Figure 5B). The bioluminescence detected from day 10 for each of these defined areas was up to 10000 fold background levels (Figure 5C) for some areas. No apparent parasite infiltration/retention was seen in any particular region, supporting the notion that the development of T. vivax is confined to the vascular compartment.
In a further set of experiments we compared the distribution of T. vivax at key time points in the infectious process by ex vivo examination of the main organs affected after injection of the TvLrDNA-luc strain. Groups of 3 mice per time point were sacrificed and the light emitted by spleen, lungs, liver and brain was promptly recorded in the presence of excess of D-Luciferin. As can be seen in Figure 6A and regardless the individual variation in the level of bioluminescence observed inside the group, the spleen was affected soonest after infection and constituted one of the first sites of parasitic retention (day 10 p.i., peaking by day 15 p.i.), as shown by at least 1000 fold greater light emission than in uninfected mice. Photon emission increases were recorded for all the organs tested after day 10, with elevated levels in liver and in particular lungs (Figures 6B and 6C). At about day 20 p.i., the luminescence in the lungs of infected mice accounted for up to 15% of the total signal recorded for the entire body, compared with 1% in lungs harvested from non-infected mice. Likewise, while no surprising light emission is apparent in the hearts after 10 days of infection (Figure 6C arrows), a significant rise in photons per second (1,4×107±6×106 ph./s, not shown) is recorded for the mouse hearts to attain more than 1000 times the background levels for the organ by day 28 of infection.
Use of bioluminescent T. vivax in vivo also allowed the validation of previous data which showed that the parasite may cross the blood brain vessels and lodge into the brain parenchyma [6], [14]. Indeed the bioluminescence signal from the brains of the animals that resisted longer the hyperparasitemia peak (20–30% survival by day 20, see Figure 3C), increases substantially from day 20 (Figure 7) but it is only visible after 25 days p.i. in localized light emission foci, just some days preceding death.
T. vivax is one of the leading parasites responsible for AAT, or Nagana, that still ranks among the most neglected diseases. One of its main particularities that can explain its capacity to emerge in areas free from tsetse flies is its ability to also be transmitted mechanically by a broad spectra of haematophagous insects [5], [15]–[17]. We have previously developed experimental models of T. vivax infection using mice infected with the ILRAD1392 reference strain [9]. Immunobiological and immunophysiopathological analyses confirmed that these models are reliable and consistent with all the relevant characteristics of the animal disease [9], [10]. Furthermore, we have also developed robust methods for parasite growth and differentiation in axenic cultures, paving the way to appropriate conditions for the first genetic manipulation of T. vivax [11]. In the study reported herein we generated an ILRAD1392 strain that stably expresses firefly luciferase (TvLrDNA-luc). Then, in detailed studies using this bioluminescent parasite, we ascertained that it has the same virulence in immunocompetent mouse models and behaves in the same manner as WT parasites. We established that these TvLrDNA-luc mutant parasites can be used successfully to i) monitor in vivo the infectious processes triggered by T. vivax and ii) evaluate organ infiltration by these parasites in vivo and ex-vivo. To the best of our knowledge, the study reported herein is the first to use a systematic imaging method to study the experimental model of T. vivax infection in vivo and the first to demonstrate the presence of the parasite in the brain by simple bioluminescent signal.
The use of bioluminescent parasites is a strategy of choice for investigating and following infectious processes in vivo [18], [19]. The approach has been used successfully to analyze infections caused by Plasmodium berghei, Leishmania major, Toxoplasma gondii, Trypanosoma cruzi and T. brucei [13], [20]–[28]. Here, an imaging system is used to quantify the light emitted by transfected cells - and in particular by microorganisms constitutively expressing the luciferase reporter gene - and thus monitor the infectious process in vivo without animal sacrifice. The technique is widely used for instance to screen new active compounds with the intention of discovering novel chemotherapies [20], [22], [29]–[31]. We decided to use the TvLrDNA-luc strain to follow the dynamics of T. vivax infection in vivo from the start to the end of the infectious process. We compared the results obtained with two different routes of infection, i.e. the conventional experimental intraperitoneal (IP) route and subcutaneous inoculation that mimics natural transmission of the parasite by the vector. The only difference observed between these routes was that the subcutaneous injection resulted in a more prolonged prepatent period and in parasite multiplication in situ for some days before it reached the bloodstream. These observations are fully consistent with data previously obtained mainly with T. congolense and T. evansi where it was shown that the inoculation of metacyclics into the animal's hypodermis either by insect bite or syringe resulted in the development of a local inflammatory reaction, called inoculation chancre [32]–[35]. This was followed by parasite multiplication in the skin, as shown by conventional histopathology, and their migration into the bloodstream. In our study, we did not observe any inflammatory reaction at the injection site. This discrepancy could be due to the nature of the parasitic form injected since it has been previously reported that T. congolense bloodstream forms, at least in low numbers, are unable to induce a metacyclics-like inflammatory reaction [36].
Although it is generally recognized that African trypanosomes migrate from the skin into the blood via the lymph system, our results did not demonstrate whether or not parasite multiplication in this compartment contributes to the infection spreading or to the orientation of the immune response. However, the light emissions we measured clearly demonstrated that growing parasites were accumulating close to the inoculation site. The results we obtained with bioluminescent T. vivax are fully consistent with our previous reports [10] and confirm that the parasite spreads across the spleen and liver compartments. It is interesting to note that lung infiltration by the parasite is difficult or impossible to observe by immunohistopathology, contrasting with our present observations. But, considering that the bioluminescent reaction is dependent on the O2 concentration [37], we cannot exclude that the substantial signal in the lungs during the late phases of the infection partially results from the abundance of O2 in this organ. Our previous report [9] has revealed the presence of innumerous parasites in the ventricular cavities of the heart. The present data showed a considerable and unfailing increase of luminescence throughout the period of study which attains its maximum between days 15 and 20 of infection when 40% of the mice die. These observations are suggestive of a congestive heart failure-inducing death, consistent with that reported for infected cattle [38], [39]. Noteworthy, the progressive increase of bioluminescence observed in testis (R5) is suggestive that parasites can pass through Sertoli cell barrier during infection and thus contribute to reproductive disorders in the seminiferous tubules, as already suggested in reports with host infected with T. vivax and T.b. brucei [27], [40]–[42]
Our observations together with earlier histopathological studies [9] nevertheless suggest that the parasite reaches the brain tissues during late (encephalic ?) phases of the infection for those mice that survive longer the hyperparasitemia occurrence. The light emitted by the brain increased slightly up till day 20 post-infection (<400 fold the background) then rose substantially by day 25, reaching sufficient levels (up to 2000 fold the background) to provide a picture of parenchymal infiltration. These observations corroborate previous reports showing molecular and histopathological data on the detection of T. vivax both in the cerebrospinal fluid and the nervous tissue parenchyma of goats [14]. Correspondingly, at late stages of T.b. brucei infection in rat and mouse models, the parasite actively migrate out of the cerebral blood vessels, cross the endothelial basement membrane, the perivascular space and the parenchymal membrane to invade the brain parenchyma, with no signs of plasma protein leakage into the brain [43]. These results were indicative that parasites had penetrated the parenchyma through the blood brain barrier rather than from circumventricular organs or through the cerebrospinal fluid (for a review, see [44]). Our present data could not clarify if the bioluminescent signal results from intra or extra vascular parasites. Using PCR of CSF extracts and histopathology of the brain may give a better picture of this question (ongoing experiments).
Altogether, the results reported herein strongly support use of the TvLrDNA-luc strain for detailed in vivo studies of the infectious process triggered by T. vivax. In particular, our data showed that the TvLrDNA-luc strain is highly appropriate to ascertain the evolution of the infection and the mechanisms involved in the progression of the disease. A more in-depth comprehension of the strategies set in place by the parasite to persist inside the host could open up perspectives for the development of a new therapeutic strategy against AAT. Our data also validate the use of bioluminescent T. vivax in high throughput drug screening strategies.
|
10.1371/journal.ppat.1003005 | Exploring New Biological Functions of Amyloids: Bacteria Cell Agglutination Mediated by Host Protein Aggregation | Antimicrobial proteins and peptides (AMPs) are important effectors of the innate immune system that play a vital role in the prevention of infections. Recent advances have highlighted the similarity between AMPs and amyloid proteins. Using the Eosinophil Cationic Protein as a model, we have rationalized the structure-activity relationships between amyloid aggregation and antimicrobial activity. Our results show how protein aggregation can induce bacteria agglutination and cell death. Using confocal and total internal reflection fluorescence microscopy we have tracked the formation in situ of protein amyloid-like aggregates at the bacteria surface and on membrane models. In both cases, fibrillar aggregates able to bind to amyloid diagnostic dyes were detected. Additionally, a single point mutation (Ile13 to Ala) can suppress the protein amyloid behavior, abolishing the agglutinating activity and impairing the antimicrobial action. The mutant is also defective in triggering both leakage and lipid vesicle aggregation. We conclude that ECP aggregation at the bacterial surface is essential for its cytotoxicity. Hence, we propose here a new prospective biological function for amyloid-like aggregates with potential biological relevance.
| Microbial infections are reported among the worst human diseases and cause millions of deaths per year over the world. Antibiotics are used to treat infections and have saved more lives than any other drug in human history. However, due to extended use, many strains are becoming refractive to common antibiotics. In this light, new promising compounds, like antimicrobial proteins and peptides (AMPs) are being investigated. Some AMPs also show agglutinating activity; this is the ability to clump bacteria after treatment. This feature is particularly appealing because agglutinating peptides could be used to keep bacteria to the infection focus, helping microbe clearance by host immune cells. In this study, we propose a novel mechanism to explain agglutinating activity at a molecular level using Eosinophil Cationic Protein. We show that the agglutinating mechanism is driven by the protein amyloid-like aggregation at the bacteria cell surface. Accordingly, elimination of the amyloid behavior abolishes both the agglutinating and the antimicrobial activities. This study provides a new concept on how Nature could exploit amyloid-like aggregates to fight bacterial infections. Moreover, these results could also add new insights in understanding the relation between infection and inflammation with dementia and amyloid-related diseases like Alzheimer.
| Antimicrobial proteins and peptides (AMPs) represent a wide family that contributes to the host defense system with multiple pathogen killing strategies [1]–[3]. Their fast and multitarget mechanism of action reduces the emergence of bacteria resistance and represents a valuable alternative for common antibiotics [4], [5].
The mechanism of action of AMPs has been systematically investigated, suggesting that AMPs bind to bacteria cell membranes and disrupt cell homeostasis. However, more investigations are needed to completely understand how different structures determine the function of AMPs [6]–[12]. Membrane damage is a multifaceted mechanism that can involve different peptide assemblies and ultimately promotes membrane permeabilization when achieving a critical concentration [13], [14]. Several authors have highlighted the striking resemblance of membrane disrupting mechanisms with those observed for amyloid peptides and proteins [15]–[17]. In both cases, membrane composition (e.g. cholesterol content) and biophysical properties (e.g. membrane fluidity and curvature) were found critical for the peptide action [13], [15], [18]–[26]. Furthermore, we have recently suggested that antimicrobial activity could have arisen through cationization of amyloid-prone regions [27]. In this light, some AMPs have been described to form amyloid structures in vitro [28], [29] and some amyloid peptides have also been considered as putative AMPs [30], [31]. In fact, we have proposed that inherent AMP aggregation properties can modulate antimicrobial activity [32].
Interestingly, some antimicrobial proteins and peptides have been found to agglutinate bacteria cells. In this sense, bacteria agglutination has been ascribed to unspecific adhesion through hydrophobic interactions, as observed for synthetic peptides derived from the parotid secretory protein [33]. Comparative analysis on those peptides highlighted the contributions of both hydrophobic and cationic residues in the agglutination activity [33]. These results suggest that some AMPs could exploit their intrinsic aggregation properties, by triggering bacteria agglutination as part of its mechanism of action as observed for a wealth source of AMPs in saliva, which provides a first barrier to bacteria adherence in the oral cavity [34]. Agglutinating activity has been reported crucial for the antimicrobial function of Eosinophil Cationic Protein (ECP) [35], a small cationic protein specifically secreted by eosinophil granules during inflammation processes with diverse antipathogen activities [36]–[38]. ECP displays high antimicrobial action, with a specific bacteria agglutination activity reported for Gram-negative bacteria, at a concentration range close to the minimal inhibitory concentration, a behavior that may represent an effective bactericidal mechanism in vivo [39].
In order to characterize the relation between AMPs, bacteria agglutination and amyloid aggregation, we have used ECP as a model of study. We present here a detailed characterization of protein-mediated bacteria agglutination and prove the contribution of an aggregation prone domain to the protein antimicrobial action. Complementary studies on model membranes provide a further understanding of the membrane damage process promoted by protein aggregation.
ECP was previously reported to aggregate in vivo on both bacterial and eukaryotic cell surface without detectable internalization [39], [40]. Though these findings were essential to explain the antimicrobial and cytotoxic properties of ECP, the real nature of the aggregation process remained unknown. Besides, the protein has a high affinity towards lipopolysaccharides (LPS) [41] and agglutinates all tested Gram-negative strains [42]. On the other hand, ECP has been reported to form amyloid-like aggregates in vitro at specific conditions due to a hydrophobic patch located at the N-terminus. Remarkably, protein amyloid-like aggregation was efficiently abolished by mutating Ile 13 to Ala [28]. The screening of the protein primary structure [43]–[45] and the design of derived peptides [42], [46] also allocated the antimicrobial region at the N-terminus. As the antimicrobial and amyloid active segments of the protein colocalize [28], [35], [42], [46], it is tempting to hypothesize that bacteria agglutination by ECP could be directly dependent on an amyloid-like aggregation process. This hypothesis raises some exciting questions: (i) Is cell agglutination required for antimicrobial activity? (ii) Is cell agglutination mediated by protein aggregation at the bacteria surface? (iii) Are aggregates formed on the surface of bacteria of amyloid nature?
To address the first question we compared the antimicrobial action of wild type ECP (wtECP) with the I13A mutant, previously described to be unable to form aggregates in vitro [28]. The antimicrobial assays reveal that, while wtECP has an average minimal inhibitory concentration (MIC) value around 0.5–1 µM, the I13A mutant is unable to kill bacteria even at 5 µM concentration (Table 1). To further correlate ECP antimicrobial and agglutination activities we studied bacteria cell cultures by confocal microscopy using the SYTO9/Propidium iodide nucleic acid fluorescent labels that allow registering both cell agglutination and viability over time. Interestingly, wtECP can agglutinate Gram-negative bacteria before a viability decrease is observed (Figure 1A), however no cell agglutination takes place when bacteria are incubated with the I13A variant, even after 4 hours (Supporting Information Figure S1). These results are also supported by minimal agglutination concentrations (MAC) close to the MIC values (Table 1) and by FACS experiments showing that wtECP but not I13A mutant is able to agglutinate E. coli cells (Figure 1B). Thus, ECP antimicrobial activity on Gram-negative strains is strongly affected when abolishing the agglutination behavior (Ile13 to Ala mutation).
To further analyze the protein agglutination mechanism, we tested the wtECP and I13A mutant action on a simpler biophysical system such as phospholipid membranes where liposome agglutination is registered as a function of protein concentration. In contrast to wtECP, I13A mutant completely looses the ability to agglutinate membranes (Figure 2A). In particular, when following wtECP agglutinating activity as a function of ionic strength, we observe that liposome agglutination is enhanced at high NaCl concentration (Supporting Information, Figure S2). These results suggest that vesicle agglutination is promoted by hydrophobic interactions. Even more, leakage activity in model membranes is also lost for I13A mutant (Figure 2B), meaning that protein aggregation on the membrane surface is important not only for agglutination but also for later membrane permeabilization. These results are entirely consistent with those described above for bacteria cell cultures where the Ile to Ala mutation not only abolishes the cell agglutinating activity of ECP but also its bactericidal action.
Next, to address the question whether cell agglutination is consistently driven by protein aggregation at the bacteria surface, we incubated bacteria cultures with ECP and visualized the samples using confocal microscopy. Our results show that wtECP binds to the bacteria surface and a strong protein signal is registered at the aggregation zones (Figure 3A). On the contrary, though cell interaction is maintained for the I13A mutant, agglutination is observed neither in bacteria cell cultures nor in model membranes (Figures 3A and 3B). As expected, for model membranes we show that only wtECP is able to promote agglutination (Figure 3B). Therefore, we conclude that protein aggregation on the cellular surface is required for bacteria agglutination, which turns to be essential for the antimicrobial action. Agglutination is also observed in the presence of 20% plasma in a similar extent, suggesting that ECP agglutination is likely to take place in the physiological context (Supplementary Information Figure S3). As previously mentioned, ECP binding to bacteria is favored by interactions with the LPS outer membrane [35], [41], [47]. Consistently, we show here that LPS binding activity is lost for the I13A mutant, when compared with wtECP (Supplementary Information Figure S4).
At this point however, the nature of the protein aggregates remained unknown. Thus, having previously shown that ECP is able to form amyloid-like aggregates in vitro, we decided to test if the observed aggregates have an amyloid-like structure using the amyloid-diagnostic dyes Thioflavin-T and Congo Red. When bacteria cultures are incubated with non-labeled wtECP, stained with ThT and visualized by total internal reflection fluorescence (TIRF) microscopy, we show that wtECP amyloid-like aggregates are located also at the cell surface (Figure 4A) similarly as what we observe for Alexa labeled wtECP (Figure 3A). Consistently, no staining is observed for non-incubated cultures and for the I13A mutant (Figure 4A). Moreover, upon bacteria incubation with wtECP, a red shift in the Congo Red spectrum is observed (Supplementary Information Figure S5A), revealing that the protein amyloid-like aggregation is triggered upon incubation with bacteria cultures.
Though ECP was previously shown to form amyloid-like aggregates in vitro only at low pH after a long incubation time (1–2 weeks), amyloid-like structures observed here are detected after only 4 hours of incubation. However, it is well known that some proteins can accelerate its aggregation kinetics in the presence of membrane-like environments [48]–[50]. Our results show that wtECP is able to form fibrillar-like aggregates on model membranes with an average size of 845±150 nm (Figure 4B), comparable in size with the wtECP aggregates observed in vitro in the absence of lipid membranes (∼150 nm) [28]. In fact, when tested for ThT binding, we observe aggregates with similar size (Figure 4B). When wtECP is incubated with model membranes and tested for Congo Red binding, we obtain again a noticeable spectral shift (Supplementary Information Figure S5B). To complete these results we have also performed all the experiments detailed above using the I13A mutant and found it to be unable to form amyloid-like aggregates (Figure 4).
The results presented here for ECP reinforce the hypothesis that an amyloid-like aggregation process is taking place in the bacteria surface that drives bacteria cell agglutination, which is essential for the antimicrobial activity of the protein. In summary, after binding to the bacteria surface, a rearrangement of the protein could take place, exposing the hydrophobic N-terminal patch of the protein. Following, the aggregation process would start promoting the agglutination of the bacteria cells through the aggregation of the surface-attached protein molecules. The formation of aggregates on the bacteria surface will disrupt the lipopolysaccharide bilayer of Gram-negative cells exposing the internal cytoplasmatic membrane to the protein action, promoting the membrane disruption and eventually the bacteria killing.
Cell agglutinating activity provides a particularly appealing feature that may contribute to the clearance of bacteria at the infectious focus. In this sense, bacteria agglutination would prepare the field before host phagocytic cells enter in the scene [33]. However, despite the interest in the pharmaceutical industry to identify the structural determinants for bacteria cell agglutination, bibliography on that subject is scarce and only few agglutinating antimicrobial proteins are described in the literature. Excitingly, there may be other proteins and peptides with similar characteristics that also follow the proposed model. Hence, the agglutinating mechanism may represent a more generalized process that may derivate in amyloid deposit formation at bacterial infection focuses.
Besides, it has been reported that systematic exposure to inflammation may represent a risk factor on developing Alzheimer's disease [51], [52] and other types of dementia [53]. Some studies have also demonstrated that the release of inflammatory mediators can also cause generalized cytotoxicity. In particular, ECP has been discovered to be cytotoxic [40], [54] and neurotoxic, causing the Gordon phenomenon after injection intratechally in rabbits [55]. Therefore, our results suggest that the release of inflammatory mediators after infection (like AMPs) may either seed the aggregation processes in the brain and/or influence the membrane biophysical properties to trigger neurotoxicity and aggregation events.
Antimicrobial activity was expressed as the MIC100, defined as the lowest protein concentration that completely inhibits microbial growth. MIC of each protein was determined from two independent experiments performed in triplicate for each concentration. Bacteria were incubated at 37°C overnight in Mueller-Hinton II (MHII) broth and diluted to give approximately 5·105 CFU/mL. Bacterial suspension was incubated with proteins at various concentrations (0.1–5 µM) at 37°C for 4 h either in MHII or 10 mM sodium phosphate buffer, 100 mM NaCl, pH 7.4. Samples were plated onto Petri dishes and incubated at 37°C overnight.
For MAC determination, bacteria cells were grown at 37°C to mid-exponential phase (OD600 = 0.6), centrifuged at 5000×g for 2 min, and resuspended in 10 mM sodium phosphate buffer, 100 mM NaCl, pH 7.4, in order to give an absorbance of 0.2 at 600 nm. A 200 µL aliquot of the bacterial suspension was incubated with proteins at various (0.1–10 µM) concentrations at 25°C for 4 h. Aggregation behavior was observed by visual inspection and minimal agglutinating concentration expressed as previously described [42].
Bacteria cells were grown at 37°C to mid-exponential phase (OD600 = 0.6), centrifuged at 5000×g for 2 min, resuspended in 10 mM sodium phosphate buffer, 100 mM NaCl, pH 7.4 or the same buffer supplemented with 20% plasma to give a final OD600 = 0.2 and preincubated for 20 min. A 500 µL aliquot of the bacterial suspension was incubated with 5 µM of wtECP or I13A mutant during 4 h. After incubation, 25000 cells were subjected to FACS analysis using a FACSCalibur cytometer (BD Biosciences, New Jersey) and a dot-plot was generated by representing the low-angle forward scattering (FSC-H) in the x-axis and the side scattering (SSC-H) in the y-axis to analyze the size and complexity of the cell cultures. Results were analyzed using FlowJo (Tree Star, Ashland, OR).
Bacteria viability assays were performed as described before [39]. Briefly, bacteria were incubated in 10 mM sodium phosphate buffer, 100 mM NaCl, pH 7.4 with 5 µM of wtECP or I13A mutant and then stained using a syto 9/propidium iodide 1∶1 mixture. The viability kinetics were monitored using a Cary Eclipse Spectrofluorimeter (Varian Inc., Palo Alto, CA, USA). To calculate bacterial viability, the signal in the range 510–540 nm was integrated to obtain the syto 9 signal (live bacteria) and from 620–650 nm to obtain the propidium iodide signal (dead bacteria). Then, the percentage of live bacteria was represented as a function of time.
The ANTS/DPX liposome leakage fluorescence assay was performed as previously described [56]. Briefly, a unique population of LUVs of DOPC/DOPG (3∶2 molar ratio) lipids was obtained containing 12.5 mM ANTS, 45 mM DPX, 20 mM NaCl, and 10 mM Tris/HCl, pH 7.4. The ANTS/DPX liposome suspension was diluted to 30 µM concentration and incubated at 25°C in the presence of wtECP or I13A mutant. Leakage activity was followed by monitoring the increase of the fluorescence at 535 nm.
For liposome agglutination, 200 µM LUV liposomes were incubated in 10 mM phosphate buffer, pH 7.4, containing 5 to 100 mM NaCl, in the presence of 5 µM wtECP or I13A mutant and the scattering signal at 470 nm was collected at 90° from the beam source using a Cary Eclipse Spectrofluorimeter (Varian Inc., Palo Alto, CA, USA) [57].
Experiments were carried out in 35 cm2 plates with a glass coverslip. For phospholipid membranes, 500 µl of 200 µM LUV liposomes (prepared as described in Supplementary Information) were incubated with 5 µM wtECP or I13A mutant for 4 h in 10 mM sodium phosphate buffer, 100 mM NaCl, pH 7.4. For bacteria, 500 µl of E. coli cells (OD600 = 0.2) were incubated with 5 µM wtECP or I13A mutant for 4 h in 10 mM sodium phosphate buffer, 100 mM NaCl, pH 7.4. RNase A was used always as a negative control. Samples of both liposomes and bacteria were imaged using a laser scanning confocal microscope (Olympus FluoView 1000 equipped with a UPlansApo 60× objective in 1.4 oil immersion objective, United Kingdom). wtECP and I13A mutant labeled with Alexa Fluor 488 were excited using a 488-nm argon laser (515–540 nm emission collected) and Vibrant DiI was excited using an orange diode (588–715 nm emission collected).
To study the interaction of proteins with lipid membranes, planar supported lipid bilayers were used (Supplementary Information). When using bacteria, glass coverslips were previously treated with 0.1% poly-L-lysine to ensure that samples will adhere to the surface. 500 µl of E. coli cells (OD600 = 0.2) were incubated with 5 µM wtECP or I13A mutant for 4 h and then transferred to poly-L-lysine treated microscopy plates and incubated for 15 minutes. To remove unattached cells, plates were washed twice with 10 mM sodium phosphate, 100 mM NaCl, pH 7.4 buffer. RNase A was used always as a negative control. Images were captured using a laser scanning confocal microscope (Olympus FluoView 1000 equipped with a PlansApo 60× TIRF objective in 1.4 oil immersion objective, United Kingdom) using the same conditions as described for confocal microscopy experiments. Thioflavin T (ThT) was used to detect amyloid aggregates. In this case, samples were incubated for 4 h with unlabeled proteins as described before and then incubated with ThT at 25 µM final concentration for 15 minutes. Then, plates were washed twice with 10 mM sodium phosphate, 100 mM NaCl buffer, pH 7.4 to remove unattached cells and ThT excess.
|
10.1371/journal.pbio.1002000 | An Adaptive Threshold in Mammalian Neocortical Evolution | Expansion of the neocortex is a hallmark of human evolution. However, determining which adaptive mechanisms facilitated its expansion remains an open question. Here we show, using the gyrencephaly index (GI) and other physiological and life-history data for 102 mammalian species, that gyrencephaly is an ancestral mammalian trait. We find that variation in GI does not evolve linearly across species, but that mammals constitute two principal groups above and below a GI threshold value of 1.5, approximately equal to 109 neurons, which may be characterized by distinct constellations of physiological and life-history traits. By integrating data on neurogenic period, neuroepithelial founder pool size, cell-cycle length, progenitor-type abundances, and cortical neuron number into discrete mathematical models, we identify symmetric proliferative divisions of basal progenitors in the subventricular zone of the developing neocortex as evolutionarily necessary for generating a 14-fold increase in daily prenatal neuron production, traversal of the GI threshold, and thus establishment of two principal groups. We conclude that, despite considerable neuroanatomical differences, changes in the length of the neurogenic period alone, rather than any novel neurogenic progenitor lineage, are sufficient to explain differences in neuron number and neocortical size between species within the same principal group.
| What are the key differences in the development and evolution of the cerebral cortex that underlie the differences in its size and degree of folding across mammals? Here, we present phylogenetic evidence that the Jurassic era mammalian ancestor may have been a relatively large-brained species with a folded neocortex. We then show that variation in the degree of cortical folding (gyrencephaly index [GI]) does not evolve linearly across species, as previously assumed, but that mammals fall into two principal groups associated with distinct ecological niches: low-GI mammals (such as mice and tarsiers) and high-GI mammals (such as dolphins and humans), which are found to generate on average 14-fold more brain weight per day of gestation. This greater daily brain weight production in mammals with a highly folded neocortex requires a specific class of progenitor cell-type to adopt a special mode of cell division, which is absent in mammals with slightly folded or unfolded neocortices. Differences among mammals within the same GI group (high or low) are not due to different programming, but rather the result of differences in the length of the neurogenic period. So, the impressively large and folded human neocortex, which is three times the size of the chimpanzee neocortex, can be explained by a modest evolutionary extension of the neurogenic period with respect to its closest primate ancestors.
| Development of the mammalian, and in particular human, neocortex involves various types of neural stem and progenitor cells that reside in the germinal layers of the cortical wall [1]–[5]. An increase in the proliferative capacity of these cells underlies the evolutionary expansion of the neocortex, notably the increase in neuron number. At the onset of mammalian cortical neurogenesis, neuroepithelial cells transform into apical radial glia (aRG), which repeatedly undergo mitosis at the apical surface of the ventricular zone (VZ) and typically divide asymmetrically to self-renew and generate either a neuron, an apical intermediate progenitor cell, a basal intermediate progenitor cell (bIP), or basal radial glia (bRG) (the latter two being collectively referred to as basal progenitors [BPs]) [1]–[5].
In contrast to aRG cells, BPs delaminate from the apical surface and translocate their nucleus to the basal-most region of the VZ to form a secondary germinal layer, the subventricular zone (SVZ), where they divide symmetrically or asymmetrically. In developing mouse neocortex, bIPs typically divide symmetrically to generate two post-mitotic neurons (neurogenic bIP) [5]–[8], whereas in the macaque and human, bIPs can also frequently undergo symmetric proliferative divisions (proliferative bIP) [5],[9],[10]. Similarly to aRG cells in the VZ, bRG cells in the SVZ divide both symmetrically and asymmetrically [9],[11]–[14], which leads to the proliferation of their population and their self-renewal, respectively [5]. Importantly, the symmetric proliferative divisions of bIPs and bRG cells result in the transit-amplification of BPs [9],[10],[12],[15],[16], which in turn allows for an increase in the efficiency of subsequent neuron generation [2],[5],[17],[18].
In mammals exhibiting an abundance of BPs during cortical neurogenesis, the SVZ becomes further compartmentalized into an inner (ISVZ) and outer SVZ (OSVZ), as first described in the macaque [19] and subsequently observed in several species in which bRG cells constitute a relatively high proportion of BPs [1]–[3],[5]. Moreover, bRG cells are characterized by radial fibers, which distinguish them from bIPs. These radial fibers of bRG cells in the OSVZ of gyrencephalic mammals typically have divergent, rather than parallel, trajectories to the cortical plate, which is thought to contribute to creating the folded cortical pattern observed in these species through the tangential expansion of migrating neurons [2],[3],[5],[20]. For this reason, and based on supporting evidence obtained in the gyrencephalic human and ferret and lissencephalic mouse, an abundance of asymmetrically dividing bRG cells in the OSVZ has been thought to be necessary for establishing a relatively large and gyrencephalic neocortex [1],[9],[11],[12]. However, subsequent work in the lissencephalic marmoset (Callithrix jacchus) has shown that bRG cells may, in fact, exist in comparable abundance in the developing neocortex of both gyrencephalic and lissencephalic species [21],[22], indicating that bRG abundance alone cannot be sufficient for either establishing or increasing cortical gyrification. Rather, the mode of cell division, that is, symmetric proliferative versus asymmetric self-renewing, of bRG cells, and of BPs in general, may be a critical determinant of the extent to which gyrification occurs and the neocortex expands. Notably, despite considerable progress in the study of brain size evolution [23]–[25], the adaptive mechanism that has evolved along certain mammalian lineages to produce a large and folded neocortex is not known.
In this study, we analyzed physiological and life-history data from 102 mammalian species (Tables S1 and S2; Database S1). We show that a gyrencephalic neocortex is ancestral to all mammals and that GI, like brain size, has increased and decreased along many mammalian lineages. These changes may be reliably characterized by convergent adaptations into two distinct physiological and life-history programs, resulting in a bimodal distribution of mammalian species with regard to the gyrencephaly index (GI) and the amount of brain weight produced per gestation day. We explain the appearance of these two groups in mammalian evolution by the adaptation of differences in the lineages and modes of cell division of progenitor cells during corticogenesis. We predict that symmetric proliferative BP divisions are key to evolutionary changes in gyrification and expansion of the neocortex.
We collected GI data (Figure S1) for 102 species sampled from every mammalian order and tested multiple models for GI evolution using a species-level supertree [26]. The model that conferred the most power to explain GI values across the phylogeny while making the fewest assumptions about the data (i.e., that had the lowest Akaike Information Criterion [AIC]) diverged significantly from a null model of stochastic evolution [27] and showed a disproportionate amount of evolutionary change to have occurred recently, rather than ancestrally, in mammals (Figure S2). We identified a folded neocortex (GI = 1.36±0.16 standard error of the mean [SEM]) as an ancestral mammalian trait (Figure 1). This result held even when additional hypothetical lissencephalic species were added to the root of the phylogenetic tree (Table S3).
It is apparent from ancestral and other internal node reconstructions (Figure S3) not only that GI is very variable, but also that reductions in the rate at which GI evolves have favored branches leading to decreases in GI (e.g., strepsirrhines and eulipotyphla) and accelerations in that rate have favored branches leading to increases in GI (e.g., carnivores and caviomorphs). A simulation of the average number of total evolutionary transitions between GI values evidences more affinity for transitioning from high-to-low than low-to-high GI values: the majority of high-to-low transitions (58.3%) occurred in species with a GI<1.47; and the fewest transitions (16.7%) occurred across a threshold GI value (see below) of ∼1.5 (Figure S4). This finding indicates that, although there is an evident trend in mammalian history to become increasingly gyrencephalic, the most variability in GI evolution has been concentrated among species below a certain threshold value (GI = 1.5). We therefore present a picture of early mammalian history, contrary to most previous work, but which is gathering evidence through novel approaches [28],[29], that the Jurassic-era mammalian ancestor may, indeed, have been a relatively large-brained (>10 g) species with a folded neocortex.
The evolutionary effects of a folded neocortex on the behavior and biology of a species is not immediately clear. We therefore analyzed associations, across the phylogeny, of GI with discrete character states of 37 physiological and life-history traits (Table S2). Distinct sets of small but significant (R2≤0.23, p<0.03) associations were found for species above and below a GI value of 1.5, indicating that these two groups of species adapt to their environments differently (Figure 2A). Although species above and below GI = 1.5 tend to fall within classical definitions of slow and fast life-histories, respectively, our results argue in favor of a dichotomy rather than a continuum and, additionally, bear out ecological and behavioral associations not historically bracketed in slow or fast life-history paradigms [30]. For example, the result that narrow habitat breadth and large population group size are associated with low-GI (<1.5) species, whereas wide habitat breadth and large social group size are associated with high-GI (>1.5) species, suggests not only that an ecological distinction be made for mammals between the population size of co-habitating individuals and the number of those individuals interacting socially, but also that the number of habitat types in which a species must compete may assert a positive selection pressure on neocortical evolution. Importantly, both the low-GI and high-GI groups are sampled from across the phylogeny, testifying to the absence of a phylogenetic signal in the establishment of the two groups and a functional role for GI in the evolution of life-history programs. Hierarchical clustering analysis also supports a bimodal distribution above and below a GI value of 1.5 (Figures 2B and S5).
In order to test the bimodal distribution explicitly, we regressed GI values against neuroanatomical traits typically identified with (and studied in the field of) neocortical evolution: brain weight, neocortical volume, and neuron number. We found that each scaling relationship could be explained comparably well by either a non-linear function (Figure 3A) or two grade-shifted linear functions, with the best-fit linear models drawing significantly different slopes for high-GI and low-GI species (Figure 3B–3D). Specifically, by plotting GI as a function of cortical neuron number, we were able to determine, with two significantly different linear regressions for high- and low-GI species (T = 4.611, degree of freedom [d.f.] = 29, p = 2.8×10−4), demarcating values of 1±0.11×109 neurons and 1.56±0.06 GI (Figure 3D), thus providing a neuron number correlate for the GI threshold. The deviation of these results from previous work, which have shown strong phylogenetic signals associated with both GI [31],[32] and neuron counts [33], may be explained both by our more than 2-fold increase in sampled species and the a priori assumption of previous work that GI and neuron number evolve as a function of phylogeny. Variation in GI, therefore, has not evolved linearly across the phylogeny, but has in fact been differentially evolved in two phenotypic groups.
By identifying an evolutionary threshold in the degree of gyrencephaly, as well as a correlate in terms of neuron number, we revealed the existence of two neocortical phenotypic groups, which found support in their distinct life-history associations (i.e., the GI is bimodally distributed and supports two principal mammalian phenotypes). These groups could be further divorced by accounting for the amount of brain weight accumulated per gestation day—a confident proxy for neonate brain weight per neurogenic period (Figure S6A and 6B)—which we show to be, on average, 14-times greater in high- compared to low-GI species (Figure 4). Notably, each GI group is constituted by both altricial and precocial species, so the degree of pre- versus post-natal development is not enough to explain the discrepancy in brain weight per gestation day in each group.
Rather, to explain the discrepancy, we introduced a deterministic model of cortical neurogenesis, using series summarizing seven neurogenic lineages (Figure 5A and 5B) and based on cell-cycle length, neuroepithelial founder pool size, neurogenic period, and estimates of relative progenitor-type population sizes (Tables 1 and 2). In total, 17 species were incorporated in the model, as we were limited by the number of species for which cortical neuron number was available. These species include species from four phylogenetic orders: Primata, Scandentia, Rodentia, and Didelphimorphia. We arrived at two models that show the highest reliability for predicting cortical neuron numbers in a range of species: a mouse neurogenic program, which implicates only asymmetrically dividing aRG and bRG cells and terminally dividing IPs (Figure 5A, lineages 1–3); and a human neurogenic program, which additionally implicates BPs undergoing symmetric proliferative divisions in the SVZ (Figure 5A, lineages 4–7). Each model is defined by the proportional occurrence of each lineage in that model (Table 2).
Using the mouse neurogenic program we were able to predict neuron counts within 2% of the observed counts for mouse and rat, but underestimated neuron counts by more than 80% in high-GI species (Figure 5C; Table S4). Increased proportional occurrences of the bRG lineage 3 (Figure 5A) with increasing brain size was required to achieve estimates with <5% deviation from observed neuron counts in the other low-GI species (Table 2; Figure S7). The human neurogenic program predicted neuron counts within 5% for all six high-GI species, but overestimated neuron counts by more than 150% for the low-GI species. Estimates of proportional occurrences of the various lineages in the mouse, marmoset, rabbit, macaque, and human are supported by previous work detailing relative abundances of different progenitor cell-types during cortical neurogenesis [7],[9]–[11],[14],[22] (IK and WBH, in preparation). Evolutionary gain or loss of proliferative potential in the SVZ is an essential mechanistic determinant of neocortical expansion, such that the presence of symmetric proliferative BP divisions in high-GI species and their absence in low-GI species is sufficient and even requisite for explaining neocortical evolution (Figure S8). Notably, the lissencephalic opossum, a marsupial species with extreme altriciality, required a decreased proportional occurrence of the bIP-containing lineage 2 (Figure 5A) and an increased proportional occurrence of the direct neurogenic lineage 1 (Figure 5A) but, like all species analyzed here, could not achieve its observed neuron count without the bRG-containing lineage 3 (Figure 5A). This suggests that bRG cells are ancestral at least to the therian stem [34].
To simulate the adaptiveness of evolving increased proliferative potential in the SVZ in two lissencephalic species—mouse and marmoset—we calculated trade-offs between neuroepithelial founder pool size and neurogenic period using mouse/marmoset and human programs of cortical neurogenesis to achieve 109 neurons. We show that, in both species, evolving a lineage of BPs capable of symmetric proliferative divisions is between two and six times more cost-efficient than either expanding founder pool size or lengthening neurogenesis; and that the marmoset, by evolving such proliferative BPs, could achieve 109 neurons by increasing either its observed founder pool or neurogenic period less than 15% (Figure 6).
We further clarified the significance to neuron output of each progenitor-type with deterministic and stochastic models of temporal dynamics and progenitor cell-type variables. The proportional contributions of each lineage to overall neuron output in the mouse and human neurogenic programs were calculated using stage-structure Lefkovitch matrices. By excluding lineages one at a time, we determined the degree to which each lineage contributed to total neuron production. From these analyses, it was clear that symmetric proliferative BPs are increasingly necessary in larger brains and that any exponential increase in neuron production is statistically implausible in the absence of such BPs (Table S5).
Finally, we described the dynamics of asymmetric self-renewing versus symmetric proliferative progenitors, isolated from their observed lineage beginning at the apical (ventricular) surface, by introducing three ordinary differential equations (ODEs) modeling a self-renewing cell that generates either a differentiated cell or proliferative cell. The ODEs describe a self-renewing mother progenitor, which can generate either a post-mitotic neuron or a proliferative daughter at each division. The proliferative daughter is allowed one symmetric proliferative division followed by self-consumption. The likelihood of a neuron or proliferative daughter being generated by the mother, therefore, is interdependent. We also include the pool of mother progenitors as a linear variable. We show that neuron output of the system increases dramatically when both the initial pool of self-renewing cells and the likelihood of those initial cells to generate proliferative, rather than differentiated, cells approaches saturation (Figure S9).
The emergence of new structures, in the most general sense, is typically limited to selection on existing developmental processes; and conserved pathways may persist, over evolutionary time, even when the phenotype is transformed or unexpressed [35]–[37]. However, it is also evident that development may be adapted without affecting phenotype (e.g., [38],[39]). Therefore, in order to understand selective pressures acting on a discontinuous or convergent trait, it is necessary to investigate the underlying developmental processes generating it.
We have shown that a gyrencephalic neocortex is ancestral to mammals. This finding is concordant with evidence [29] that the mammalian ancestor was relatively large (>1 kg) and long-lived (>25-year lifespan) and, furthermore, provides considerable resolution to recent evidence for a gyrencephalic eutherian ancestor [28] by sampling nearly twice as many species and categorizing gyrencephaly as a continuous, rather than a binary, trait.
More surprisingly, we show that convergent evolution of higher orders of gyrencephaly along divergent lineages has been accompanied by two distinct constellations of physiological and life-history paradigms. Specifically, species with a GI>1.5, which is commensurate with 1 billion cortical neurons, exhibit patterns of development and life-history that are distinct from species with a GI≤1.5, irrespective of phylogeny. This implies that there is a considerable constraint on either the ability of species of a given neocortical size to exploit certain ecologies or the potential for species of a given ecology to freely adapt neocortical size. Even marine mammals, whose selection pressures are sui generis, may largely be held to the same evolutionary stereotyping as terrestrial mammals (Figure S10).
While our results countenance previous studies showing associations between physiological and life-history traits in mammals (see [40]), we identify those traits to have a bimodal distribution, rather than to vary allometrically, across species. This distribution depicts a Waddington-type landscape for neocortical expansion—albeit relevant at the species-level—wherein the GI threshold represents an adaptive peak requiring a particular adaptation in neurogenic programming within a population for traversal. Our results may explain this landscape by mechanistic differences occurring during cortical neurogenesis between species above and below the GI threshold: the necessity of symmetric proliferative BPs in high-GI species and their putative absence in low-GI species.
The human neurogenic program constructed here clearly shows that the same neurogenic lineages in the same proportions are required to generate the neocortices of Old World monkeys, apes, and humans, and may even be extended to carnivores, cetartiodactlys, and other high-GI species (Figure S10), demonstrating that neurogenic period alone may be sufficient to explain differences in neocortical size between any species in the same GI group (Figure S11). Our data are insufficient, however, to determine whether these adult differences are uniform across the neocortex or differentially represented in infra- versus supra-granular layers [20],[41].
We propose that symmetric proliferative divisions of BPs, in addition to having an abundance of bRGs in an expanded SVZ, are necessary and sufficient for the evolution of an expanded and highly folded neocortex in mammals. Recent work in the fetal macaque supports this proposal [10]. We thus conclude that an increase in the proliferative potential of BPs is an adaptive requirement for traversing the evolutionary GI threshold identified here. But because we reconstruct the eutherian ancestor to have a GI value of 1.48±0.13 (standard error of the mean [SEM]) (Table S3), which falls within the range of the observed threshold, we are left with an ambivalent evolutionary history for mammalian neocortical expansion: either (i) BPs capable of undergoing symmetric proliferative divisions are ancestral to all eutherian mammals and were selected against along multiple lineages (e.g., rodents, strepsirrhines), so that the ultimate loss of BP proliferative potential in certain taxa, and therefore the evolution of low-GI species, is the result of divergent developmental adaptations; or (ii) such symmetric proliferative BPs are not ancestral to eutherian mammals, but evolved convergently along multiple lineages, in which case the developmental process for their inclusion in neurogenic programming may be conserved, even if that process was unexpressed for long stretches of mammalian evolution.
We have revealed an important insight into mammalian evolution: a GI threshold exists in mammalian brain evolution; neocortical expansion beyond that threshold requires a specific class of progenitor cell-type (BPs) to adopt a specific mode of cell division (symmetric proliferative); and the difference in neuron output between any species on the same side of that threshold does not appear to require adaptations to the lineage or mode of cell division during neurogenesis, but may simply reflect differences in the length of the neurogenic period. Further research into the conservation of genomic regions regulating the capacity of BPs to undergo symmetric proliferative divisions (e.g., through the establishment and maintenance of a proliferative niche in the SVZ) in low- versus high-GI species may reveal whether this mechanism for neocortical expansion has evolved independently in distantly related species or is the product of a deep homology in mammalian cortical development.
We calculated GI using images of Nissl-stained coronal sections from http://brainmuseum.org. We used 10–22 sections, equally spaced along the anterior-posterior axis of the brain, for each species (Figure S1). The inner and outer contours of the left hemisphere were traced in Fiji (http://fiji.sc/wiki/index.php/Fiji). The species for which we calculated GI are indicated by an asterisk in Table S1. Additional GI values were collected from the literature (Table S1; Database S1). Several species (e.g., platypus), whose cortical folding has been described [42],[43] but not measured according to the method established by [44], could not be included in our primary reconstructions of GI evolution (Figure 1). However, these species, assumed to be lissencephalic (GI = 1.0), were included in supplemental analyses (Table S3; see Reconstructing the evolutionary history of GI).
Work in humans and baboons has shown that interindividual variation in GI is not enough to outweigh interspecific differences [45],[46].
Variation in the mode and tempo of a continuous character trait is not always best characterized by a random walk (i.e., Brownian motion). Therefore, we compared a range of evolutionary models on the phylogenetic distribution of GI to find the best fit for the data [47]–[50]. Log-likelihood scores for each model were tried against the random walk score using the cumulative distribution function of the χ2 distribution. Maximum-likelihood ancestral character states of GI and rate-shifts in the evolution of GI were then constructed using the best-fit model, with the standard error and confidence intervals calculated from root node reconstruction in PDAP using independent contrasts [51]–[53]. Although a number of putatively lissencephalic non-eutherians were unavailable for our analyses (see Calculating GI), we nonetheless reconstructed alternative ancestral GI values that included one hypothetical monotreme and three hypothetical marsupials (Table S3). The phylogeny used in this analysis was derived from a species-level supertree [26]. We appreciate that the phylogenetic hypothesis reconstructed by [54] gives notably deeper divergence dates for mammalian sub-classes; however, not enough of our sampled species were included in this reconstruction for it to be useful here.
To trace evolutionary changes in GI at individual nodes and along lineages, we used a two-rate mode that highlighted the differences in high (>1) versus low (<1) root-to-tip substitutions and then sampled rates based on posterior probabilities across the tree using a Monte Carlo Markov Chain. We assumed that transitioning between adjacent GI values had the highest likelihood of occurrence. The rate at a given node could then be compared to the rate at the subsequent node to determine if a rate transition was likely. We corroborated these results using the auteur package [55], which calculates rate-transitions at internal nodes under the assumption of an Ornstein-Uhlenbeck selection model [34] over 1 million Monte Carlo sampling iterations drawn from random samplings of posterior distributions of lineage-specific rates.
Scaling relationships were determined for GI as a function of all continuous life-history and physiological traits, including adult cortical neuron counts. For three eulipotyphla species (Sorex fumeus, Blarina brevicauda, Scalopus aquaticus), data were available for neuron counts but not GI, and therefore we extrapolated the GI of those species on the basis of gross morphology. Finally, to test whether the bimodal distribution of GI may be influenced by the topology of the mammalian phylogenetic tree, we used an expectation-maximization algorithm. Each simulated trait was given the same variance as GI (Figure S5) and the result was averaged over 104 simulated datasets. None of the simulations produced the same bimodal distribution of species observed for GI data.
We used a comprehensive phylogenetic approach to map 37 life-history and physiological character traits collected from the literature (Tables S1 and S2) onto hypotheses of phylogenetic relationships in Mammalia, in order to examine how those traits correlate, over evolutionary time, with degree of gyrencephaly. Continuous character traits were discretized using the consensus of natural distribution breaks calculated with a Jenks-Caspall algorithm [56], model-based clustering according to the Schwarz criterion [57], and hierarchical clustering [58]. Character histories were then corrected for body mass with a phylogenetic size correction [59],[60] and summarized across the phylogeny using posterior probabilities. Associations between individual states of each character trait along those phylogenetic histories were calculated in SIMMAP (v1.5) using empirical priors based on the frequency of character states for each trait [61]; the association between any two states was a measure of the frequency of occurrence (i.e., the amount of branch length across the tree) of those states on the phylogeny. While correcting for body mass is intended to normalize the data, it cannot completely remove interdependencies between character traits. Although we cannot a priori assume that any of the traits interact, exploring interactions between them deserves further investigation.
The sums, rates, and types of changes for GI and body weight were plotted as mutational maps to assess directional biases in their evolution [62]–[64]. These were used to determine the evolutionary historical patterns of GI and, as a control, body weight. By estimating the occurrence (number of times an increase/decrease happens) and timing (where in the phylogeny the change occurs) of different values for each trait, we were able to calculate how often each trait has increased and decreased in mammalian evolution. We were therefore able to evaluate the ratio of increases over decreases for each trait (Figure S4).
We estimated neuroepithelial founder pool populations for mouse and human. For the mouse, we used coronal sections of an E11.5 mouse embryo obtained from the Allen Brain Atlas [65]. We obtained 19 sections equidistantly spaced along the anterior-posterior axis of the brain. The length of the ventricular surface of the dorsal telencephalon was manually traced in Fiji [66] on each section starting from the point above the nascent hippocampus and ending in the point above the lateral ganglionic eminence. The horizontal length of the embryonic brain at E11.5 was measured with images from [67]. Using the coronal and horizontal measurements, we constructed a polygon representing the ventricular surface of the dorsal telencephalon and calculated the area of this surface in Fiji. We measured the surface area of the end-feet of neuroepithelial cells using EM images of the coronally cut apical surface of an E11.5 embryonic mouse brain (Table S6). The diameter of a single end-foot was calculated by measuring the distance between the adherens junctions. We corroborated these end-feet calculations with published immunofluorescence stainings of the apical complex (ZO1 and N-cadherin) from an en face perspective [68],[69]. The average surface area of a single end-foot was calculated by approximating the end-foot as a hexagon; and the number of founder cells was estimated by dividing the surface of the dorsal telencephalon by the surface of an individual end-foot of the neuroepithelial cell, such thatOur final mouse values were comparable to those previously published [70]. For the human, we followed the same procedure, using ten coronal sections and one horizontal section of a gestation week (GW) 9 brain [71]. End-feet were calculated using EM images of the apical surface of a human brain at GW13. The measurements are available in Table S6. Because the number of founder cells per surface area was nearly equivalent in mouse and human (4×105/mm2), we used this ratio, along with data on ventricular volume collected from the literature (Tables S1 and S2; Database S1), to estimate neuroepithelial founder cell populations for a further 15 species (Table 1). For species where no data on ventricular volume were available, values were estimated on the basis of a regression analysis against brain weight (Figure S6). Ventricular volume was then converted to surface area for each species by approximating the ventricle as a cylinder with a 4.5-to-1 height-to-diameter proportion (this ratio was estimated on the basis of observations in mouse). Ventricular volume-derived ventricular surface area estimates were corroborated with the surface areas calculated from the literature for mouse and human. Founder cell estimates were then computed on the basis of the densities derived above for mouse and human. Using this method, but alternately ignoring our mouse and human calculations to define the parameters, we were able to predict mouse and human values within 10% of our calculations, respectively.
Workers have demonstrated the occurrence of three primary lineages of neuron generation in mouse corticogenesis (Figure 5A, lineages 1–3) [1],[5],[14],[72] and a further four lineages in primate corticogenesis (Figure 5A, lineages 4–7) [9],[10]. While there is evidence for at least one additional lineage in mouse [6], and further lineages may be speculated, we limited our model to the seven that are considered to contribute most significantly to neuron output [2],[10],[73],[74]. The extent of neuron generation in each of these seven lineages was summarized in series and solved numerically (Figure 5B).
Neurogenic period was either taken from the literature (Database S1) or estimated on the basis of a regression analysis of neurogenic period as a function of gestation period (Figure S6). Neurogenic period in human was estimated using empirical observations from the literature [75]–[77]. The averaged cell-cycle length for apical and BPs from the mouse (18.5 hours) [78]–[80] was used for all non-primates; averaged cell-cycle length for cortical areas 17 and 18 from the macaque (45 hours) was used for catarrhines [10],[81]; and an intermediary cell-cycle length (30 hours in marmoset, determined by EdU labeling; Ayako Murayama and Hideyuki Okano, personal communication), was used for platyrrhines. Using an average cell-cycle length value for all progenitor-types was found to be equally valid for predicting neuron number as using different cell-cycle lengths for each progenitor-type (Figure S12). Therefore, despite its potential shortcomings, using average cell-cycle length is a valid approach and, given the scarcity of species data on the cell-cycle length of various progenitor-types at different stages of neurogenesis, the best approach available to construct neurogenic models across many species. Generous confidence intervals (75%) for cell-cycle length are used in our models (Figure 5C), in order to show the minimal explanatory power cell-cycle length provides for interspecific differences in cortical neuron number.
Diminishing numbers of neuroepithelial cells have been observed to continue to proliferate at the ventricle until E18.5 in the mouse [7]. Therefore, final neuroepithelial founder pool estimates were calculated from the aforementioned by evenly decreasing the value of n in the Sherley equation [82] from 1 at E9.5 to 0 at E18.5 in the mouse and at comparable neurogenic stages in other species.
Neuron numbers were calculated for each species from combinations of lineages. The proportional contribution of each lineage for each species was parameterized according to existing data on progenitor cell-type abundances in mouse [14], marmoset [22], rabbit (IK and WBH, in preparation), macaque [10], and human [9],[11]. Where no such data were available, proportional contributions were permutated for all lineages until a best-fit estimate, based on cortical neuron numbers taken from the literature [33],[83]–[85], was achieved (Tables 1 and 2). Each lineage was assumed to occur from the first to final day of neurogenesis, although this is only approximately accurate. Finally, because of published estimates of postnatal apoptosis in the mammalian cortex [86]–[88], we assumed neuron counts to be 1.5-fold higher at the termination of neurogenesis than in the adult brain; therefore, neuron number at the termination of neurogenesis was estimated in each species by multiplying neuron numbers collected from the literature by 1.5. This multiplication is not represented in Table 1.
Trade-offs in adapting a human neurogenic program with either an expanding neuroepithelial founder pool or lengthening neurogenic period were tested for the mouse (Mus musculus) and marmoset (Callithrix jacchus), two lissencephalic species whose cell-type proportions during neurogenesis have been documented [7],[14],[22].
To estimate the relative reproductive value and stable-stage proportions of each of the lineages in the mouse and human neurogenic programs, we constructed a stage-structured Lefkovitch matrix, using sums of the lineage series (after 100 cycles) as fecundity values and complete permutations of the proportional contributions of each lineage as mortality values. The altered growth-rates of each lineage were calculated by excluding lineages one at a time and assuming 100% survival in the remaining lineages (Table S5).
We introduced three ODEs to explore the average dynamics of asymmetric versus symmetric progenitors, such that: if a(t), b(t), and c(t) are the numbers of asymmetrically dividing cells, differentiated cells, and proliferative cells, respectively, then,where r is equal to growth-rate. If a(t) = a0, thenandUsing these ODEs, we calculated the effect on neuron output of increasing the likelihood of symmetrically dividing daughter progenitors in the lineage (Figure S9). The interdependent growth-rates in the model reflect a purely mechanistic interpretation of determining neuronal output from a finite pool of asymmetrically dividing cells. The ODEs, therefore, may not reflect differential regulation of neuronal output via direct versus indirect neurogenesis. The daughter proliferative cells are designed to carry out one round of proliferation followed by a final round of self-consumption (Figure S9).
|
10.1371/journal.pntd.0005864 | Soil iron and aluminium concentrations and feet hygiene as possible predictors of Podoconiosis occurrence in Kenya | Podoconiosis (mossy foot) is a neglected non-filarial elephantiasis considered to be caused by predisposition to cumulative contact of uncovered feet to irritative red clay soil of volcanic origins in the tropical regions. Data from structured observational studies on occurrence of Podoconiosis and related factors are not available in Kenya.
To establish the occurrence and aspects associated with Podoconiosis, a cross-sectional survey was implemented in an area located within 30 km from the foot of volcanic Mount Longonot in the Great Rift Valley in Kenya. Five villages and 385 households were selected using multistage and systematic random sampling procedures respectively during the survey. Podoconiosis was determined by triangulating (1) the clinical diagnosis, (2) molecular assaying of sputum samples to rule out Wuchereria bancrofti microfilaria and (3) determining the concentration of six elements and properties in the soil known to be associated with Podoconiosis. A structured questionnaire was used to identify possible risk factors. Univariable and multivariable Poisson regression analyses were carried out to determine factors associated with Podoconiosis. Thirteen participants were clinically positive for Podoconiosis giving an overall prevalence of 3.4%. The prevalence ranged between 0% and 18.8% across the five villages. Molecular assay for W. bancrofti test turned negative in the 13 samples. The following factors were positively associated with the Podoconiosis prevalence (P<0.1) in the univariable analyses: (i) age, (ii) gender, (iii) education level, (iv) frequency of washing legs, (v) frequency of wearing shoes, (vi) soil pH, and (vii) village. Unexpectedly, the concentration of soil minerals previously thought to be associated with Podoconiosis was found to be negatively associated with the Podoconiosis prevalence (P<0.1). In the multivariable analyses, only frequency of wearing shoes and village turned out significant (P≤0.05). By modeling the different soil mineral concentrations and pH while adjusting for the variable frequency of wearing shoes, only iron concentration was significant and in the negative dimension (P≤0.05). However, controlling for Iron, Aluminum concentrations turned significant.
This study has pointed to a hitherto unreported occurrence of Podoconiosis cases and has contributed to the baseline knowledge on the occurrence of Podoconiosis in Kenya. Consistent with many studies, wearing shoes remain an important risk factor for the occurrence of the disease. However, our findings are inconsistent with some of the hitherto postulations that associate Podoconiosis prevalence with certain minerals in the soil in other regions in Africa. These findings provide new beginnings for the cross-disciplinary research of Podoconiosis in environmental health, socio-ecology and ecological niche and geo-spatial modeling and prediction.
| Podoconiosis is a neglected disease in the tropical regions of the world considered to be caused by prolonged contact of uncovered feet to irritant particles found in red clay soil from volcanic origins. The disease presents like filarial elephantiasis. Data from observational studies from Kenya are not available. We conducted a cross-sectional household survey to establish the prevalence and aspects related with Podoconiosis at the foot of Mount Longonot in the Great Rift Valley in Kenya. Podoconiosis was determined by combining results of clinical diagnosis, ruling out filarial elephantiasis in clinically positive Podoconiosis patients using molecular techniques and determining the concentration of elements and properties in the soil known to be associated with Podoconiosis. A structured questionnaire was used to identify possible risk factors. Out of 385 study participants, thirteen were clinically positive for Podoconiosis giving an overall prevalence of 3.4%. Molecular tests for filarial elephantiasis turned negative in the 13 participants. Factors that were associated with Podoconiosis prevalence were age, gender, education level, and frequency of washing legs, frequency of wearing shoes, soil pH and village. The concentration of soil minerals previously thought to be associated with Podoconiosis was found to be negatively associated with the Podoconiosis prevalence. However, the final analyses found frequency of wearing shoes, iron and aluminium as possible predictors of Podoconiosis occurrence in the study area. This is the first structured observational study to report occurrence of Podoconiosis in Kenya. Although some of our findings are inconsistent with some previous reports about the association of Podoconiosis and certain minerals in the soil, this study offers new beginnings for the cross-disciplinary research of Podoconiosis in fields known to influence occurrence of the disease including environmental health, socio-ecology and medical geographical approaches and predictions.
| Podoconiosis is a neglected geochemical, non-filarial, non-infectious lymphodema of the lower limb [1]. In Africa, countries in which non-filarial elephantiasis have been reported include: Tanzania [2], Uganda [3], Kenya [4], Cameroon [5], Sao Tome and Principe [6], Rwanda, Burundi, Sudan, Ethiopia [7], and Equatorial Guinea [8]. Podoconiosis is known to result from an interaction between genetics and an unusual provocative response to reactive mineral particles found in clay soil, red in color, derived from volcanic origin deposits [1]. Mineral particles from the soil are thought to penetrate the skin, then they are fought by macrophages in the lymphatic system which causes inflammation and fibrosis of vessel’s lumen leading to blockage of the lymphatic drainage [9]. This results in edematous feet and legs and subsequently progresses to elephantiasis [10] and nodular skin changes [11].
The prevalence of the disease is reported to vary considerably from country to country. For instance, reports show an average burden of 1% (range: 0% to 2.1%) in Burundi [12] and 0.6% (range: 0.1% to 1.7%) in Rwanda [12]. Other reports indicate that prevalence ranged from 0.4% to 3.7% in Ethiopia [13]. Recently, higher prevalences (range: 3.3 to 7.4%) have been reported in Ethiopia [14][15][16][17]. Non-filarial elephantiasis cases were also documented in Kenya at the foot of Mt. Kenya in the year 1948 [18]. However, there are no published findings from structured observational studies in Kenya.
Numerous structured studies investigating the role of individual-level risk factors have been carried out in Ethiopia. Gender, age, marital status, feet hygiene, level of job skills/employment, education level and house floor type have been associated with risk of Podoconiosis. Being older [15], female, single, rarely washing feet, and low skilled or jobless [19][20] showed significant association with increased occurrence of Podoconiosis. On the other hand, formal education [20] and living in a house whose floor is covered [20] were related with low risk of Podoconiosis. These findings point to the opportunities of modifying certain risk factors in the prevention of the disease.
Podoconiosis has been reported to be prevalent in highlands of tropical Africa, Central America and Northwest India all characterized by certain soil types [11][21]. Altitude and rainfall are among other factors which has been associated with occurrence of Podoconiosis [20]. These geographic characterizations are associated with consistent breakdown of molten rock and their mineral components into silicate clays. These geologic properties, through the development of peripheral water gradient potentially influence permeability of the stratum corneum in the skin and raise transdermal uptake of potential toxins and colloid-sized particles of elements common in irritant clays [10]. Hence, Podoconiosis is widespread in countries which are within the Rift Valley geological complex in Africa [11][12] and other areas with volcanic soils [22]. Soils within these areas are red clay loams, slippery and stick to the skin when wet [11]. Here lies the opportunity of utilizing the African Soil Atlas by specifically predicting Podoconiosis occurrence (http://eusoils.jrc.ec.europa.eu/content/soil-map-soil-atlas-africa) in Africa.
Moreover, Podoconiosis is occupational in nature with familial inclination in addition to a deficiency in feet hygiene. Podoconiosis occurs among farmers and other occupational groups whose feet remain uncovered and exposed to clay soil originating from alkaline volcanic rock [22]. Small particles such as silica, sodium, magnesium, aluminum, iron and potassium [12][23][24] of the incriminated soil type (almost nanoparticles) are thought to pass through the skin cracks and find their way into the lymphatic system. Besides, studies have shown high hereditability of susceptibility to Podoconiosis [11][25]. Indeed, the prevalence was reported to be higher in people who rarely wore shoes, indicating possible interrelationship between Podoconiosis, genetics, occupation, environmental factors and lifestyle [12].
Podoconiosis is associated with a host of disease burdens. The quality of life is substantially reduced [26][27]. Though non-fatal, those affected will show spoiled appearance of their legs [28]. Clinically, most patients acquire repeated infections of bacterial and fungal nature in the affected leg(s) necessitating extra medical attention [15][29]. Approximately all Podoconiosis patients suffer from acute lymphadenitis five or more times a year. It has been estimated that they lose an average of one month of economic activity every year due to morbidity [15][29][30]. In Southern Ethiopia, an assessment of the economic costs of Podoconiosis indicated that, in an area with 1.7 million residents, the cost of the disease was 16 million US Dollars annually, hence leading to Ethiopian loss of 200 million US Dollars per year. A research comparing affected and unaffected people within the same level of employment showed that those with disease are half as productive as those without disease. [31]. Stigma associated with Podoconiosis is manifested in people by dropping from schools, exclusion from social community activities, diminished marriage opportunities and reduction in economic development and psychological trauma [32].
Diagnosis of Podoconiosis is not straightforward. To rule in Podoconiosis, geographical location, history, clinical findings and confirmed absence of microfilaria or its antigen on immunological card test are used. Geographically, it has been found that Podoconiosis is prevalent in populations who live at high altitudes (>1000 metres above sea level). Clinically, Podoconiosis is an ascending and commonly bilateral non-filarial elephantiasis though asymmetric [33] and rarely involves the groin. On the other hand, lymphatic filariasis is found at altitudes lower than 1000 metres above sea level. In addition, clinical changes are first noticed at the groin in lymphatic filariasis. We used these features to triangulate the diagnosis of the disease.
This paper reports a cross-sectional household survey implemented to establish the burden and factors related with Podoconiosis occurrence. This information will be useful to the health administrators and humanitarian agencies responsible for developing and implementing targeted, appropriate and effective public health intervention strategies. This is the first field-based observational survey that has acknowledged the occurrence of Podoconiosis in Kenya.
The study was carried out in Mt. Longonot region in Nakuru County with a population of 1,603,325 (year 2009 census) [34] and an area of 2,325.8 km2. Mount Longonot is located within Nakuru County, Kenya which is an agriculturally rich county. It is also a strato-volcano situated southeast of Lake Naivasha in the Great Rift Valley of Kenya in Africa. The county generally has an elevation of 2776m [35]. Nakuru County has temperatures ranging from a minimum of 12°C to a maximum of 26°C. Rainfall ranges from 1800 to 2000mm per year [36].
A cross-sectional quantitative community-based household survey was implemented in this study. The study population consisted of women, men and children aged 5 years and above from the area of study. This age was chosen because Podoconiosis incidence rises with age [37]. We included people who were residents of the study area, had lived in the study area for five or more years and also consented to take part in the study. The exclusion criteria were: those aged less than five years, not have lived in the area for five years and above and those who did not consent to the study.
The sample size was calculated using the Cochran formula [38]. Due to unavailability of the prevalence of Podoconiosis in Nakuru County, 50% prevalence and a tolerable error (level of precision) of 5% was used to determine the sample size. A sample size of 385 participants was computed using the formula below;
n=1.962p(1-p)L2
Whereby, 1.96 was the z value for the desired confidence level (95%), p was an estimate of the probable prevalence of Podoconiosis and L was the level of precision. Multi-stage random sampling was used to select villages to be included in the study. Villages within 30 km from the foot of Mt. Longonot were identified. Nine villages were identified and five of them selected randomly. Those selected included Githarani, Scheme, Ereri, Lower Kiambogo and Upper Kiambogo.
The number of study participants in each village proportionately depended on the village population and the calculated sample size and was computed as shown below:
ni=NiN*n
Where ni was the sample size for a village i, Ni was the population size in village i, N was total population in the study site, and n was the overall calculated sample size (385). Table 1 shows the computed statistics by village with N being 6068. One participant was randomly selected in each household; therefore, the number of participants was equal to the number of households to be included in the study. Systematic selection of the households was done depending on the total number of households to the sample households required from each village by dividing the number of households in each village by sample size in that village (Table 1). For instance, from any starting point, Githarani, households were selected at the interval of every three houses. In case the third house did not have residents, we selected the next one that had residents. The study participant was subsequently selected randomly from the present household members.
With the help of research assistants, one soil sample was collected from each of the five villages giving a total of five soil samples. Each unique sample consisted of cores taken from study households and pooled together within each village. Soil was dug using a 12cm shovel to a depth of 25cm since the elements occur on the surface layer of between 0-25cm. The amount dug was placed in a bucket and thoroughly mixed. Moist soil samples were air dried at the site away from dust contamination. The soil sample bags which were well labeled with the sample code were filled half full (500g) from this mixed representative sample and tightly packed [39]. The samples were kept in a secure room at room temperature and transported as a batch to SGS Laboratories in Mombasa, Kenya.
Soil analyses was done at SGS Laboratories in Kenya to provide data on the level of concentration of the following elements—aluminum, magnesium, silica, sodium, potassium and iron, and pH- known to be associated with Podoconiosis [24]. Atomic Emission spectrometry (Spectro Flame Modula ICP) instrument using Mehlich 3 –Diluted ammonium fluoride and ammonium nitrate [40] was used for analyses. Briefly, the soil samples were dried for 72 hours at 150°F and then crushed in a Dynacrush soil crusher to a size it can pass a 20-mesh screen. Two grams of the soil sample was scooped into 70ml extraction cup made of plastic in a Styrofoam rack. A pipette machine was used to add 14ml of Melich 3 extractant to the extraction cup. The extraction cups, held in a plastic rack were positioned in the Eberbach shaker. Above the tops of the extraction cups, a sheet of plastic cover was positioned and shaker lid closed and protected to hold the racks firm. For 5 minutes, a shaker power cord was connected to a power output plug on a GRA lab timer. Later, in the filter tube in the wooden filter rack, an 11 cm #1 of filter paper was put in place. The cup holder (rack) was removed when the shaker stopped and poured into filter papers. The sample was then transferred to autosampler tubes for analysis using Inductively Coupled Plasma (ICP) spectrometry.
Data was collected for one month by the lead author and three field officers. One day prior to the survey, a brief meeting was held with the field officers to give details regarding the study. The clinical officer used a clinical investigation form and physical examination of legs to diagnose Podoconiosis. Clinical characteristics and symptoms included bilateral but asymmetrical swelling of legs below the knee, itching and burning episodes of lower legs, lack of tropic ulcers and availability of sensation. Sputum samples from clinically positive participants were collected for molecular assay to detect Wuchereria bancrofti which causes filariasis. This was carried out to rule out filariasis which presents like Podoconiosis. The molecular assaying was executed at Kenya Medical Research Institute at the Centre for Biotechnology, Research and Development. The molecular assaying is reported to have a sensitivity of 97.5% and specificity of 92.4% [41]
Social-demographic information was collected using structured questionnaire and included age, sex, marital status, education level, type of house floor, frequency of wearing shoes, occupation and level of hygiene.
Sputum samples were collected in sterile containers for PCR to rule out Wuchereria bancrofti which causes filariasis. The samples were kept in well-sealed cooler boxes containing ice packs and transported to the nearby hospital for storage for two weeks before being transported to KEMRI.
Approximately 200μl of sputum (mucoid) was used to extract Wuchereria bancrofti DNA. All procedures were done on ice. DNA was extracted from sputum samples by alkaline precipitation method as described by Zhong et al. [42]. Approximately 200μl of serum sample was added into sterile eppendorf tubes. 198μl of 1% Triton and NaOH was added and the mixture vortexed. The mixture was heated at 65°C for 30 minutes in a thermo mixer. The PH was adjusted to 8 with HCl 1:4 or 1M NaOH. The mixture was spinned quickly at 4°C, 14000 RPM for 5 minutes and the supernatant transferred into clean Eppendorf tubes. It was heated at 100°C for 5 minutes in a thermo mixer and subsequently cooled quickly on ice. Every tube was added four hundred (400) μl of absolute ethanol and kept at 70°C in the freezer for overnight. For 20 minutes, they were spinned at 4°C, 14000 RPM and the supernatant was discarded by pipetting 400μl. It was then washed thrice with 70% ethanol. They were then spinned in a micro-concentrator for 1 hour. The mixture was suspended in 50μl of TE buffer and vortexed and stored at -20°C until PCR was carried out.
Ten (10) μl of the extracted DNA was used for DNA amplification to enable detection of Wuchereria bancrofti DNA for diagnosis of filariasis. This was done using a thermo cycler machine. A master mix preparation was first prepared as shown in Table 2. The master mix (Table 2) was first vortexed. 50μl of the master was put to a well labelled PCR tubes and 10μl of the sample added. 0.5μl of Taq polymerase was also added to the sample and the tubes positioned into a thermo cycler with the programme set and allowed to run for 35 cycles. The amplified DNA (template) samples, molecular marker, positive and negative controls were loaded in 2% agarose gels in the gel electrophoresis tank and allowed to run for one hour. It was then visualized by UV illuminator and the samples viewed against 200 base pairs in the molecular marker.
Data was entered into computer Statistical Package for Social Sciences (SPSS) software v.20.0 for analysis. Descriptive analyses were initially carried out. The study outcome was computed as the proportion of the sample surveyed that had clinical Podoconiosis. As this proportion was very small (see results below), the outcome was assumed to represent a count of the number of cases in the group. To relate the count of the cases to predictors (socio-demographic and soil mineral concentrations), a Poison regression model was assumed in the form: lnE(Y)η=β0+β1X where the term on the left of the equation was the log of the expected value of counts of disease which was modelled as a linear combination of the predictors (on the right of the equal sign). The model related the log of the expected value of counts of disease and a linear combination of one predictor in univariable analysis at a significance level of P≤0.1. The model was subsequently extended to control for other predictors by including all significant variables at the univariable step in multivariable analyses at a significance level of P<0.05 as follows:
E(Y)η=β0+β1X1+β2X2+……βkXk
Where k was the number of predictors.
Multivariable modeling was carried out by backward elimination strategy and, in addition, involved checking of confounding and relevant interaction terms. During modelling, the statistical significance of the contribution of individual predictors (in univariable analyses) or groups of predictors (in multivariable analyses) to the model was tested using the likelihood ratio test.
The models were assessed for overall fit using χ2 goodness-of-fit tests computed as the sum of the squared deviance or Pearson residuals. The values of the two test statistics were compared (as they can be quite different) to assess lack of fit. As with all overall goodness-of-fit statistics, a P>0.05 (non-significant) indicates that the model fits the data well.
This study was registered under Kenya Medical Research Institute (KEMRI), and was approved by Scientific Ethical Review Committee, KEMRI number KEMRI /RES/7/3/1. Informed consent was obtained from each study participant after reading and or providing a detailed oral explanation to all potential participants with the help of an assistant research officer. The participants were given a chance to decide whether to voluntarily participate in the study. Those willing to join the study were requested to sign on the consent form. Those with reduced capacity to sign were allowed to put a thump print on the consent form to prove consent. In individuals aged below 18 years, informed consent was obtained and interview conducted to the parents or legal guardian and the process continued as above depending on whether they were able to put a signature or thumb print on the consent document. In addition, there was a verbal assent between the researcher and children aged 13–17 years. Participants between 13–17 years of age, who accepted to participate in the study signed or put a thump in the assent form. The research assistant and the principal investigator signed the consent form as witnesses. The process of oral informed consent as well as the consent procedure was approved by the Scientific Ethical Review Committee, KEMRI.
A total of 385 participants aged 5 years and above were included in the study. The participants had a mean (standard deviation) age of 44.8 years (19.8). The sample comprised of 108 (28.1%) males and 277 (72.0%) females. Most of these participants were aged between 29 and 38 (20.5%) years old. Most of the study participants had formal education (72.0%). Of the study participants, 65.2% lived with a spouse while 34.8% lived without a spouse. Majority of the study participants were farmers; n = 291 (75.6%). More than 50% of the participants had their house floor made of earth. In addition, most of the participants washed their legs daily (99.5%) and wore shoes daily (76.9%) as shown in Table 3.
The concentration of silicon, aluminium, sodium, iron, potassium and magnesium was determined in five soil samples—each sample from each study village. The pH was also determined and the results shown in Table 4. The mean values of silicon, aluminium, sodium, iron, potassium and magnesium were 186.99mg/kg, 10303.82mg/kg, 264.45mg/kg, 15011.95mg/kg, 2121.99 mg/kg and 787.81mg/kg respectively. The mean PH was 7.04 as illustrated in Table 5.
Of the total sample size, 13 (3.4%, [95% CI, 1.8%, 5.7%]) were found to be clinically positive for Podoconiosis. Fig 1 shows the distribution of households with clinically positive and negative study participants. Table 6 shows univariable analyses of social-demographic and socio-economic factors, village, soil pH and soil element concentration associated with Podoconiosis. The prevalence ranged between 0% and 18.8% across the five villages with Githarani, Scheme, Ereri. Lower Kiambogo and Upper Kiambogo reporting 18.8%, 0.0%, 1.6%, 3.2% and 2.0% respectively (Table 6). Univariable analyses screened 16 variables but returned 12 significant variables (P≤0.1) associated with the count of Podoconiosis in the study area. These variables included age, gender, education level, and frequency of wearing shoes, village of residence and frequency of washing legs.
If a study participant were to increase the age by one year, the difference in the log of expected counts of Podoconiosis was expected to increase by 0.04 (Table 6), i.e. increased age was associated with Podoconiosis occurrence. The difference in the log of expected counts of Podoconiosis was expected to be -1.5 units lower for males compared to females (Table 6), i.e. Podoconiosis was more likely to be found in females relative to males. Additional risk factors (interpreted in the same way) included non-formal education, walking and working bare feet, failure to wash legs daily, and village (Table 6). In addition, the concentrations of silicon, aluminium, iron, magnesium and potassium were separately found to be significantly associated with the log of expected counts of Podoconiosis cases. However, this relationship was protective to the occurrence of Podoconiosis cases. Lastly, if the soil pH were to increase by one unit, the log of expected counts of Podoconiosis was expected to increase by 6 (Table 6), i.e. alkaline soil was a risk factor of Podoconiosis occurrence.
In the multivariable analyses (P≤0.05), only two variables of the 12 screened in the univariable model remained significant (Table 7). This included frequency of wearing shoes and village. The logs of expected counts of Podoconiosis was expected to be 2.7 units higher for study participants who rarely wore shoes compared to those who wore shoes daily while holding the village variable constant in the model (Table 7). In addition, the difference in the logs of expected counts of Podoconiosis was expected to be lower for study participants from all villages compared to study participants from Githarani village holding the wearing shoes variable constant in the model. The risk of Podoconiosis declined in this order in the villages: Githarani, Lower Kiambogo, Ereri, Upper Kiambogo and Scheme (Table 7).
As soil samples had been collected from the villages, they could not be modeled together with the village variable. In the second set of modeling, the variable village was dropped and the different soil mineral concentrations (silicon, aluminium, potassium, magnesium and iron) and pH modelled while adjusting for the variable frequency of wearing shoes. This stage of modeling was implemented to tease out the effect of the village. The element concentrations represented the study villages since each soil sample represented one village.
Adjusting for frequency of wearing shoes, only the iron concentration was significant (P≥0.05) (Table 8). If the concentration of iron were to increase by 1 mg, the difference in the logs of expected counts of Podoconiosis would be expected to decrease by 0.0003 units, while holding the variable frequency of wearing shoes in the model constant.
Extensive analyses were carried out first to assess confounding of all soil minerals in presence of iron. Interestingly, in the presence of iron, the effect aluminium concentration on Podoconiosis changed from negative dimension (in the univariable analyses) to positive. This suggested that iron and aluminium could be related and acting as confounders for each other. To investigate this finding further, an interaction term was generated by adding the cross-product term (iron concentration*aluminum concentration) and testing if the coefficient term was statistically significant. The interaction term was significant independently but not in presence of iron or frequency of wearing shoe in the multivariable analyses.
The parsimonious model of the data is illustrated in the model in Table 8. To assess the overall fit of the model, the chi-square goodness-of-fit tests were computed as the sum of the squared deviance and Pearson residuals. The resulting test statistic has an approximately χ2 distribution in presence of multiple observations within each covariate pattern defined by the predictors in the model (if it is significant, it indicates lack of fit). For this data, Deviance statistic had a P = 1.0000 whereas the Pearson statistic had P = 0.17 indicating that the model fit the data well.
This study provided a preliminary but detailed quantitative assessment of prevalence and factors associated with Podoconiosis occurrence in Mount Longonot region in Nakuru County in the Great Rift Valley in Kenya. Data was collected and triangulated by clinical investigation, responses from a structured questionnaire, molecular assaying to rule out infectious elephantiasis and soil mineral concentration analyses. The findings strongly pointed to possible occurrence of Podoconiosis in the region with a prevalence of 3.4%. According to our knowledge, this is the first field observational research of Podoconiosis prevalence in Kenya. This prevalence was similar to recent prevalence reports in different Podoconiosis endemic areas in Ethiopia ranging from 3.3% in Debre Eliyas and 3.4% in Dembecha woredas [17] with overall prevalence being 3.3%. An extensive evaluation documented a national prevalence of 3.4% in Ethiopia [20] similar to our study using the elimination method for diagnosis, with older studies reporting a prevalence of 2.7% [13] and 2.8% [1] in Ethiopia. Some studies in Ethiopia and elsewhere show a high prevalence between 2.8% and 7.4% [14][15][43]. A study carried out in Uganda in high-risk communities reported a prevalence of 4.5% [3], whereas another in Cameroon reported a prevalence of 8.1% [7]. Variation between our prevalence value and values from other reports are most likely due to differences in sampling methods, sample sizes, location, time and the level of risk.
Consistent with published findings, frequency of wearing shoes and soil mineral concentration (particularly iron) were linked with burden of Podoconiosis in this study. These variables show effect after a long period of exposure to reactive alkaline volcanic soils [10][11]. Minute mineral particles enter into the skin due to long-term exposure of uncovered feet to the reactive soil. This triggers a provocative reaction in the lymphatic system which causes thickening and subsequent obstruction of lymphatic system [9]. Emerging evidence also suggests that genetic susceptibility may play a role [44]. Indeed, a study has estimated that an offspring from an affected parent is five times more likely to develop Podoconiosis compared to an individuals selected randomly from the general population [10]. This estimate was not only due to shared environment alone, but surveys indicate that 63% of Podoconiosis prevalence is attributed to inheritance of susceptibility [10]. An interaction between genetic susceptibility and irritation from mineral particles has not been studied and this is a front for future research.
However, under univariable analyses, increasing age, female gender, low education level, low frequency of wearing shoes, low frequency of washing legs, high soil pH and some soil elements showed a statistically significant relationship with the prevalence of Podoconiosis.
Age, gender and education level are individual-level variables widely reported to be associated with prevalence of Podoconiosis. Increasing age is expectedly associated with occurrence of Podoconiosis most likely because age is a proxy for exposure time (cumulative exposure). Previous work reported that the disease mostly develops in the agriculturally productive ages of 16 to 54 years which could explain the cumulated exposure [20][32]. Consistent with our study findings, being a female would increase the chances of an individual having Podoconiosis [20]. Gender differences may exist in terms of preventive behaviours such as shoe ownership and wearing practices, risk behaviours such as kitchen and farm gardening as well as access to personal resources such as socks and shoes [20]. New areas for research include possible differences in genetic susceptibility [25] and biological susceptibility [20] which may be hormonal-based and, in addition, how gender roles may modulate the risk of the disease [20].
In this study, formal education was associated with decreased risk of Podoconiosis. This shows similarity with research done in Ethiopia [20] which reported that secondary and higher education was associated with decreased risk of Podoconiosis. Moreover, another study in Ethiopia [45] showed that the illiterate participants were 11 times more likely to develop Podoconiosis relative to those who had secondary education and above. This implies that Podoconiosis occurrence can predict low education level. On the other hand, low education level can predict Podoconiosis occurrence. These findings may mutually reinforce in a positive feedback loop with low education level and Podoconiosis ultimately converging leading to stigma [15]. On the other hand, formal education may prevent the occurrence of the disease by default due to better lifestyle arising from formal employment or by design where educated people are better informed of soil-borne infectious and non-infectious sources of inflammation and mostly live on non-earthen floored houses. Although age and gender are non-modifiable factors, health promotion and education are essential in endemic areas to empower residents in the control over, and to improve their feet health.
Foot hygiene practices including low frequency of wearing shoes and low frequency of washing legs were associated with increased prevalence of Podoconiosis. This is consistent with known strategies of Podoconiosis prevention. At early stages, one is capable of managing Podoconiosis and stop further disease development by regular cleaning of feet and consistent wearing of shoes [11]. However, Deribe et al [20] reported that there was no association between wearing shoes and Podoconiosis in Ethiopia. The difference between our finding and Deribe et al [20] could be partly due to individuals with disease starting to protect their legs after Podoconiosis has developed because of lack of awareness and increased stigma. In Kenya, the opposite is true as the awareness is very little or non-existent. Deliberate water, hygiene and sanitation (WASH) education and promotion should be extended to include feet in addition to hand hygiene in endemic areas in Kenya. Specific interventions are currently being tried in the field in Ethiopia including ‘foot hygiene’ [46], hence similar measures need to be introduced in Kenya.
In this study, increasing soil pH was associated with increasing prevalence of Podoconiosis and some soil elements showed a statistically significant relationship with the prevalence of Podoconiosis. Our findings are consistent with observations that the disease occurs as a result of exposure to alkaline clay soils [12][47]. In turn, the determinants of soil formation and characteristics include environmental variables such as climate and geology including weathering of rocks. Ecologically, local properties of soil are very crucial in Podoconiosis development [48]. A recent survey in Ethiopia was carried out using historical data to explore the distribution and environmental variables of Podoconiosis [48]. In the latter study, high prevalence of Podoconiosis was noted in areas of altitudes more than 1500m above seas level, with rainfall more than 1500mm per year, and temperatures ranging between 19–21°C annually. The significant village effect found in our study most likely reflected micro-differences in environmental factors that have hitherto been reported. For instance, in 1984, Price [25] noted that Podoconiosis prevalence reduced to almost zero at a distance of 25km from the point of high red clay soil. Recent incremental knowledge showed that higher levels of minerals within soil such as mica, quartz and smectite affected the occurrence of Podoconiosis [49]. Such minerals aid in water uptake, accelerate Podoconiosis occurrence by inducing pathology in individuals’ body leading to acute adenolymphangitis, a cause of severe morbidity among those affected. An important finding in this research is that the hitherto assumption which relates the availability of compounds and elements found in soil and Podoconiosis incidence [22] was not consistent. Presence of such types of data allows for environmental risk mapping in a spatial analytical framework to produce maps of reported cases and environment-based maps of the areas at risk of Podoconiosis. This approach can be augmented with Ecologic niche modeling (ENM) whose principle is to evaluate the potential spatial distribution of species. In the same way, the ENM approach would be used to define and delineate the niche of Podoconiosis as well as foresee its probable geographic plus ecologic distribution by the analysis to determine the association between environmental variables [50] for targeted intervention. A study in Ethiopia by Deribe et al [51] has initiated this promising area for research of Podoconiosis.
Diagnosis of Podoconiosis was based on differential diagnosis as done in other studies [17][20][47]. These comprise of leprosy, filarial elephantiasis and mycetoma pedis which are associated with tropical lymphodema [47]. Clinical diagnosis has been reported to be precise in endemic settings [34]. However, as opposed to filariasis where the primary swelling can occur anywhere in the inferior extremities, Podoconiosis exclusively commence in the feet. Secondly, Podoconiosis mostly occurs in both legs, but the swelling can differ in size [49] while mycetoma and filariasis is mostly unilateral. Moreover, groin involvement which is mostly an indication of filarial elephantiasis is very rare in Podoconiosis. This study uniquely triangulated Podoconiosis clinical picture, responses from structured questionnaire, molecular assay to rule out infectious elephantiasis and soil mineral concentration analyses. Local eco-epidemiology can also be a clue to diagnosis as Podoconiosis is typically found in higher altitude areas with volcanic soil simultaneously with higher rainfall [51] whereas filariasis is uncommon at higher altitudes and other environments in which the mosquito vector of filariasis is less prevalent.
This study is not without limitations. The study utilized a cross-sectional study design. These findings therefore need to be interpreted with caution and further and more intensive cross-disciplinary studies are needed to authenticate them. This is because cross-sectional studies are subject to problems of undocumented confounders operating at different scales. Furthermore, the time-relationship between the factors and the disease is not known with such a design. However, in this case, a cross-sectional study design was appropriate as the disease is mostly non-fatal and the associated variables do not have a clear time-onset. Secondly, Podoconiosis is thought to cluster within families/households, partially associated with common environmental exposure or shared genetic susceptibility. This approach could have underestimated the prevalence in the region. However, ours was a cross sectional study in the region with a descriptive purpose with respect to the outcome and a set of risk factors. Additionally, we did not count any additional Podoconiosis suspected cases in the households that were recruited for the study. Future research in the region should investigate the existence of disease clustering at the household and the spatial level to increase the understanding of the disease mechanisms. Distinct study designs particularly case-control studies are also warranted as they are appropriate for studying rare conditions or diseases such as Podoconiosis.
This study reports the occurrence of Podoconiosis in the Mt Longonot area in the Rift Valley in Kenya. Soil Iron and Aluminium concentrations and feet hygiene were identified as possible predictors of Podoconiosis occurrence in Kenya. Because of the possible spatial restriction of the exposure, external validity to other areas may not be feasible and, therefore, we restrict our conclusions to the target (source) population. The findings in this study gives the baseline knowledge regarding the occurrence of non-filarial elephantiasis in Kenya and providing a fresh beginning in cross-disciplinary research of Podoconiosis using socio-ecology, environmental health and ecological niche and geo-spatial modeling and prediction.
|
10.1371/journal.pntd.0002799 | Knowledge, Attitudes and Practices Related to Visceral Leishmaniasis in Rural Communities of Amhara State: A Longitudinal Study in Northwest Ethiopia | In the northwest of Ethiopia, at the South Gondar region, there was a visceral leishmaniasis (VL) outbreak in 2005, making the disease a public health concern for the regional health authorities ever since. The knowledge on how the population perceives the disease is essential in order to propose successful control strategies.
Two surveys on VL knowledge, attitudes and practices were conducted at the beginning (May 2009) and at the end (February 2011) of a VL longitudinal study carried out in rural communities of Libo Kemkem and Fogera, two districts of the Amhara Regional State. Results showed that VL global knowledge was very low in the area, and that it improved substantially in the period studied. Specifically, from 2009 to 2011, the frequency of proper knowledge regarding VL signs and symptoms increased from 47% to 71% (p<0.0001), knowledge of VL causes increased from 8% to 25% (p<0.0001), and knowledge on VL protection measures from 16% to 55% (p<0.0001). Moreover, the improvement observed in VL knowledge was more marked among the families with no previous history of VL case. Finally, in 2011 more than 90% of the households owned at least an impregnated bed net and had been sprayed, and attitudes towards these and other protective measures were very positive (over 94% acceptance for all of them).
In 2009 the level of knowledge regarding VL was very low among the rural population of this area, although it improved substantially in the study period, probably due to the contribution of many actors in the area. VL patients and relatives should be appropriately informed and trained as they may act as successful health community agents. VL risk behavioural patterns are subject to change as attitudes towards protective measures were very positive overall.
| Visceral leishmaniasis (VL) is a vector borne disease that can be fatal if left untreated. In northern Ethiopia there was a VL outbreak in 2005, making the disease a public health challenge ever since. In order to promote the participation of communities in the control of the disease, it is essential to know how they perceive the disease and its management. There is a paucity of studies dealing with the knowledge, attitudes and practices (KAP) towards VL in the world in general and in rural Ethiopia in particular. We conducted two KAP studies at the beginning and at the end of a VL longitudinal study carried out between 2009 and 2011. The project included VL community talks and sensitization, and there were other interventions implemented by different actors in this period. Our results showed that, among the rural communities surveyed, the knowledge regarding signs and symptoms, causes, and protective measures of the disease was very low. However, it improved substantially in the period studied, suggesting that knowledge was subject to change by community interventions. It also showed that VL patients and relatives can act as successful health agents and that the population had positive attitudes towards the implementation of preventive actions.
| Visceral leishmaniasis (VL) (also known as kala-azar) is a vector-borne neglected disease caused by the protozoan parasite Leishmania donovani in East Africa, and transmitted by the bite of female phlebotomine sand fly. Clinical signs and symptoms often include long lasting and irregular fever, weight loss and hepato-splenomegaly; and it is fatal if left untreated [1]. More than 90% of global VL cases occur in six countries: India, Bangladesh, Sudan, South Sudan, Ethiopia and Brazil. Globally, 200,000 to 400,000 new cases of VL occur every year, and only in Ethiopia it is estimated an annual incidence of 4,000 new cases [2]. The principal foci in Ethiopia are the one in the Northwest border with Sudan (Metema and Humera), and the one located in the South, in the Segen and Woito river valleys [3]–[6].
In Libo Kemkem and Fogera (highland districts in South Gondar, Amhara Regional State) VL had never been reported until May 2005 when a large VL outbreak was identified, with more than 2,500 cases treated. A high mortality rate was reported initially, probably due to the long time required for the recognition of the epidemic [7]. Migration of laborers coming from endemic neighboring areas (border of Sudan) is one of the hypotheses for the introduction of VL in the region [8], [9] that has become a public health concern for the Amhara Regional State Health Bureau ever since.
In order to elaborate successful VL control programs it is essential to know the risk factors associated with it, and to understand the disease-related knowledge, attitudes, and practices (KAP) of the population [10]. The factors associated with Leishmahia infection in this area have already been described, being related to past history of VL in the household, house conditions or behaviors like sleeping outside, among others [11]. The factors associated with the VL clinical manifestation in this area were sleeping outside or under an acacia tree were among others [8].
However, little is known about how individuals in rural communities of this region perceive the disease and its management. There is a paucity of VL KAP studies in the New World [12], [13] and in the Old World [14]–[18] in general. And in Ethiopia, to the best of our knowledge, there are no published studies that have focused on these aspects in a rural setting. Only recently it has been published a VL KAP study conducted in Addis Zemen, the urban centre of Libo Kemkem [19]. We expect our study, focused in the rural, to contribute to those urban results, in order to help the Amhara health authorities to promote the involvement of the communities in the control of the disease, a priority for the government of Ethiopia [20].
Health education campaigns should be adapted, in contents, type, and format to the target population [1], [21]. In other settings it has already been proven that educational strategies with informative materials can contribute to VL control programs [22]–[24], but written materials in rural communities of Ethiopia with high levels of illiteracy may not be appropriate. In the area of study, since the 2005 outbreak, there have been different actors implementing outreach activities with health education and case screening. And the research study that we conducted included informative and sensitization talks, which may be more appropriate for this population. By carrying out two KAP surveys at the beginning and at the end of the longitudinal study we look forward to assess baseline VL knowledge attitudes and practices, as well as the change in VL knowledge along the study period.
Furthermore, as results from other Leishmaniasis KAP studies have suggested that the knowledge of the disease is restricted to those that have suffered from it personally or in a person closely related [25], we wanted to differentiate the results regarding VL knowledge by households with and without a positive history of VL.
Therefore, the aims of this study are 1) to assess the knowledge, attitudes and practices of VL in households of a rural endemic area of Amhara Regional State, Ethiopia and 2) to evaluate the impact of community interventions in the VL knowledge at household level between 2009 and 2011, taking into account the previous VL history of the participant households.
The study was approved by the ethical advisory boards of the Armauer Hansen Research Institute and the Ethiopian National Ethical Committee in Ethiopia, and the Instituto de Salud Carlos III in Spain. Support letters were obtained from the Amhara Regional State Health Bureau and the different districts' Health Offices. All parents/guardians gave written informed consent prior to responding to the questionnaires directed to them and to the enrolment of their children in the study, and assent was also obtained from children ≥11 years of age.
The area of study was located in the Amhara Regional State, South Gondar, Northwest Ethiopia (See Custodio et al. [11] for geographical location of study site), and comprised two districts (weredas) mainly rural: Libo Kemkem (being Addis Zemen town its capital) and Fogera (being Woreta town its capital). These are adjacent districts most affected by the outbreak of VL occurred in 2004–2005 [26]. According to year 2009 census, the population was 198,374 and 226,595 for Libo Kemkem and Fogera respectively.
The KAP surveys presented in this work were carried out within the framework of a prospective longitudinal study entitled “Visceral Leishmaniasis and Malnutrition in Amhara State, Ethiopia”. The study involved a cross sectional survey conducted in May 2009 to estimate the prevalence and associated factors of VL and malnutrition in school-aged children, and a cohort study that was carried out between May 2009 and February 2011 in order to elucidate the relationships between malnutrition and Leishmania infection in this same age group. The study consisted of four surveys that were carried out in May 2009, December 2009, May 2010 and February 2011. In the first and last surveys questions related to the knowledge, attitudes and practices (KAP) towards visceral leishmaniasis and Leishmania transmission were addressed to the care providers (present at the household at the time of the survey) of the children participant in the cohorts' study.
Population sampling was carried out by multi-staged cluster survey being the primary sampling units the sub-districts (kebeles) with high incidence of VL: Bura, Yifag Akababi, and Agita from Libo Kemkem; and Sifatra and Rib Gebreal from Fogera. Secondary sampling units were randomly selected villages (gotts) in each of the selected sub-districts, and third sampling units randomly selected households in each of the villages. The sample size was calculated according to the objective of the original project, described in detail elsewhere [9], [11], [27].
In May 2009 the care providers of the children recruited for the cohort study were interviewed by trained local personnel using a standardized structured questionnaire that included questions on demographics, household characteristics, child health, VL risk factors and VL KAP. A question regarding if someone in the household had suffered VL in the past was included in order to elaborate the variable Household (HH) with positive history of VL. This variable was based on the interviewee's report, but not verified by treatment or medical record. The use and the number of bed nets owned by the household was also reported but not verified by the interviewer.
In February 2011 the same households were visited, and the interview consisted of a standardized questionnaire that included questions related to child health and a more extensive VL KAP. However, the question regarding if someone in the household had suffered VL in the past was not included in this last interview. Care providers present in the house at the time of this visit were not necessarily the same who were interviewed in the first survey (only 40% of them were the same person in the May 2009 and in the February 2011 surveys). Therefore, we assess knowledge at household level and not at the individual one.
Out of the 276 households visited in May 2009 we were able to collect data on 218 when revisited in February 2011, which, for the purpose of this study, were the ones to be kept in the analysis.
All questionnaires were translated in to Amharic, the main local language.
The outcome variables regarding “Awareness” were based on self-perceived knowledge related to VL sings and symptoms, causes, or protective measures respectively (as an example: Do you know any sign or symptom of VL? Yes/No//If yes, which ones?). The different answers were thereafter converted into dichotomous variables.
And the outcome variables regarding “Proper knowledge” were created as follows:
Before the starting of the project, a two days consultation was made with the local community leaders, sub-district (kebele), and district administrators in order to approach why the project was relevant, what was the VL situation in the area, and also to cover VL general information.
During the cohorts study four surveys were conducted. The day before visiting the community for the first data collection (May 2009 survey), the supervisor together with the kebele administrator and the community leader conducted a one day sensitization talk to every elder in the community. And before each of the following surveys the supervisor talked to the household head or adult present in the house at the time of the visit. The talks covered VL general signs and symptoms as well as Leishmania infection ways of transmission and protective measures. In addition, in January 2011 a special informative meeting was held with the community leaders of all participant gotts and with kebeles administrator in order to promote leader's encouraging to families to participate in the fourth and last survey of the project.
During the data collection process questionnaires were checked on site by the supervisor and, once they were completed, were submitted to the data processing unit of the Armauer Hansen Research Institute (AHRI) in Addis Ababa, Ethiopia, where they were double entered in ACCESS and cross checked for consistency. Joint data analysis was conducted in the Spanish National Centre of Tropical Medicine in Madrid, Spain, where data was rechecked and cleaned. Finally, data analysis was performed using STATA version 11 (Stata Corp., College Station, TX, USA).
Descriptive statistics were performed and the Chi square test was used for comparisons between HH with and without positive history of VL, except when the number in any of the categories analysed was below 5, that the Fisher's exact test was applied.
Differences in results pre (2009) and post (2011) implementation of the study were examined by the McNemar test for matched data.
All p-values were two tailed and a p-value of ≤0.05 was taken as significant.
A total of 218 households were surveyed, all of them from rural environment with uniform low socioeconomic conditions, described in Table 1. In 2009, the majority of the heads of the households were male (91.3%), illiterate (78%) and had a principal occupation related to the cultivation of land (99.8%). Among those who had their own lands (97.7%) the mean of acreage owned was 1.2 Ha (SD: 6.7) and only 11% had more than 3 Ha. The mean household size was 6.1 persons (SD: 1.7), with households size ranging from 3 to 10 persons. Radios were present only in 28% of the households.
The main source of knowledge regarding VL in 2009, before the implementation of the study, was the health centre (n = 57, 26.1%) followed by knowing someone that had suffered VL (n = 21, 9.6%).
In 2009, 47.7% of the population surveyed reported to be aware of VL signs and symptoms, versus an 84.7% in 2011 (p<0.0001). The most frequently reported signs and symptoms were abdominal swelling, fever, weight loss, and low appetite. Furthermore, in 2011 a significantly higher frequency of interviewees had “Proper knowledge of VL signs and symptoms” as compared to 2009 (71.1% to 14.6%, p = 0.0001). And, when proper knowledge on VL signs and symptoms was stratified by the variable if someone in the household had suffered VL in the past, a higher proportion of respondents living in houses with past history of VL reported correct signs and symptoms, being this difference more marked in 2009 (94% to 28%, p<0.0001) than in 2011 (89% to 64%, p = 0.001) (Table 2).
Regarding self-perceived knowledge of the possible causes of the disease, a 16.5% reported to be aware of VL causes in 2009 versus a 58.7% in 2011 (p<0.0001). The answer considered appropriate, “Insect”, was the one most frequently reported in 2009 (8.3%) and increased to 31.8% in 2011 (p<0.0001). Respondents living in houses with past history of VL reported more frequently a proper knowledge on the vector borne disease nature of VL than respondents living in houses with no history of VL, although this difference was only significant in 2009 (p<0.001)(Table 3).
In relation to VL protection measures, in 2009 only 21% of the respondents declared to be aware of how to protect from VL, in regard to 58% in 2011, p<0.0001. The most mentioned protection measures in both years were “Bed Nets” and “Environmental Sanitation”, but the probability of giving a correct answer regarding VL protective measures was almost 8 times higher in 2011 than in 2009 (p<0.0001). When stratified by houses with and without VL history, differences were found only in the responses of the 2009 survey, were proper knowledge on VL protective measures was reported more frequently among respondents of houses with positive history of VL (33.3% versus 9.0%, p<0.001) (Table 3).
In relation to attitudes and practices, in 2009 57% of the houses reported to own bed nets versus a 98% in 2011, p<0.0001. The only reason given for not owning bed nets in 2011 was “Because it is difficult to get them”. The number of households owning two or more bed nets increased from 56 (25.7%) in 2009 to 177 (81.2%) in 2011 (Table 4). In the majority of the houses, (n = 181, 85%) respondents reported that bed nets were used by all members in the family, followed by the option “Only adults” (n = 14, 6%) and “Mother and children” (n = 7, 3%). Moreover, 94% of respondents stated that they would accept using impregnated bed nets in the house.
In 2011 there were 143 (66%) houses with iron roof, an increase from the 134 (62%) in 2009 but no significant (p = 0.06) (Table 4). The main reasons reported for using iron roof were because it was more solid (n = 60, 40%) and better for weather conditions (n = 51, 36%). The reason for using straw (n = 74, 33%), the alternative roof material, was its lower price (n = 57, 77%). In relation to house conditions, more than 90% of the houses surveyed (n = 198) had cracks in the wall, and when the interviewees were asked about the optimal frequency for repairing them, the responses ranged from “Never” (n = 15, 6.9%), “Every 2 years or more” (n = 32, 15%), “Once a year” (n = 56, 26%), “More than once a year” (n = 79, 36%), to “Every month” (n = 32, 15%).
In 2011 almost every surveyed house (96%) had been sprayed compared to 64% in 2009 (p<0.0001) (Table 4). The acceptance for indoor and outdoor spraying was very high (98% to 97% respectively). And so it was the acceptance for house surroundings environmental cleaning (99%).
Of the houses surveyed, in 2011, 63% (n = 138) reported having members of the family sleeping outside, mainly due to far away herding or cattle watching in the house surroundings. However, 29 interviewees (21%) reported that family members sleeping outside made use of bed nets, and 20 (15%), declared that other protection measures like cloths, blankets or environmental sanitation were used. The reasons reported for not using bed nets while sleeping outside ranged from “Bed nets are difficult to use when sleeping outside” (40%), “Lack/shortage of bed nets” (11%) to “Bed nets are too expensive” (11%). Finally, more than 80% (n = 174) of respondents declared that at least one member of the family rested under acacia tree, a risk factor for VL in the area, and the time of the day most frequently reported for doing it was during midday (n = 144, 83%).
In 2011, the first option for VL treatment was public health facilities (n = 215, 99%), and only 3 persons (1.3%) mentioned home remedies or traditional healer as first choices, based on “Better to try home first” and “Fear of evil eye” respectively.
The present study shows that knowledge regarding visceral leishmaniasis in the rural communities of this region of Ethiopia is low, although it improved substantially among the households participating in the longitudinal research project described before, that was carried out between 2009 and 2011. The improvement in knowledge was substantially more marked among the families with no past history of VL.
This study was focused in rural population because in Africa VL is mainly transmitted in rural settings [28] and, as stated in the objectives of the Sixtieth World Health Assembly of 2007, one of the means to combat leishmaniasis is to improve knowledge about, and skills to prevent, the disease among people in rural areas [29].
In the survey conducted in May 2009, before the implementation of the study, the level of awareness related to VL signs and symptoms was around 50%, substantially lower than results from other rural communities of Nepal and India where it raises up to 85% [18], [30], [31]. This may be due to the fact that VL has been endemic for more than 20 years in the settings of the Indian subcontinent studied, and in Libo Kemkem and Fogera districts the disease has only recently been known, at least as a public health problem [26]. However, the level of awareness was also significantly lower than the one found in the urban population of this same area that reached 83%. Furthermore, in this urban population 62% of interviewees were able to identify more than one VL symptom [19], whereas only 47% were able to report at least one correct symptom in the rural population of our study. This could be explained by the location of the VL treatment center in Addis Zemen town, which allowed neighboring communities to get information and education related to the disease. And also by the higher level of illiteracy in the rural communities (72% compared to the 34% found in the Addis Zemen town study), that has been associated with poorer VL knowledge before [31]. Future education activities should be aimed to making rural population more knowledgeable of the symptoms of the disease, as proper perception prompts patients to seek early treatment, essential in VL cure and complete recovery [18].
Awareness of VL causes was as low as 16% in 2009, and correct knowledge of the vector-borne character of VL was also remarkably low (8% of respondents mentioned mosquitos and only 1 respondent specifically sand flies) compared to rural communities in other countries, where high percentages of respondents were able to identify sandflies as the transmitting agent (20% in Sudan and 21–88% in the Indian Subcontinent [15], [17]). Also in the urban area of Libo Kemkem knowledge regarding causes was higher and more specific, as 68% of interviewee reported sandflies as the VL causal agent [19].
As might be expected, knowledge regarding preventive measures was also higher in Addis Zemen town, where 20% of respondents identified bed-nets as a VL protection measure, compared to 12% of respondents in the rural communities of our study. Notwithstanding the knowledge regarding bed-net protection is poor in the area as a whole compared to the levels of knowledge found in other studies conducted in Sudan and in the Indian Subcontinent (30–50%) [15], [16], [31]. On the other hand, spraying was barely mentioned in our study which seems to be consistent with results from other VL KAP studies [15], [18]. Potential health education activities should deal with the causes of the disease and the existing protecting measures, which should lead to behavioral changes in the population. It is worth to highlight that in relation to attitudes the level of acceptance for all the protection measures suggested (impregnated bed nets, environmental sanitation and household spraying) was over 94%. However, we are aware that this figure can be overestimated by a possible information bias regarding the health professional's nature of the interviewers who may be known to the community and influence interviewees responses.
The main source of VL knowledge reported in the 2009 survey was the health center, followed by VL patients. Moreover our results show that in May 2009 HH with a positive history of VL in the family had better knowledge overall compared to HH with no history of VL, suggesting that VL patients played an important role as sources of knowledge. This was further supported by the fact that in the post-intervention survey the differences in knowledge between HH with VL positive history and HH with no history of VL practically disappeared, as there was an external source of knowledge common to all the participant families. VL patients, friends and relatives have been identified in other studies as main sources of knowledge, and it is recommended to take the opportunity to appropriately inform and educate them, so they can act as community health agents when they are back into their places of residence [13], [16], [31].
Other health education interventions like broadcasted radio programs have been identified as successful VL information sources [16] but may not be appropriate in this context, where barely a third of the HH own radio. On the contrary, it seems that informative actions, like the ones carried out by the research project presented here, may have a positive impact in the improvement of the disease global knowledge. Our results show that proper knowledge on VL signs and symptoms, on VL causes and on VL protective measures increased by 11, 8 and 8 times respectively between 2009 and 2011. The study design did not allow for comparison with control communities in order to monitor other external factors that could be influencing HH knowledge related to VL. However, we are aware that since the 2005 outbreak, there have been numerous community interventions including outreach activities with health education and cases screening activities. Therefore we believe that the important change in knowledge showed by our results may be a result of the participation of the households in the research project as well as the contribution of all the actors in the area during the study period, emphasizing the potential of indirect education activities in the improvement of VL knowledge in this particular context.
The interpretation of the change in practices observed during the study period is of a different nature, as it is directly related to the malaria prevention and control campaign carried out in the study area during 2009/2010 [32]. This campaign included massive distribution of bed nets and spraying, which accounts for the significant increase in the number of bed nets and in the number of sprayed households [33]. However, although the majority of the HH surveyed owned at least one bed net in 2011, its acceptability as a VL protective tool needs to be ascertained, as a few proportion of the respondents related it to VL protection. Furthermore, in this study we identified that more than 60% of HH had at least one family member sleeping outside, and only a third of those reported to use any protection. They constitute and important group of hosts susceptible to sand fly bites, that need to be aware that sleeping outside without taking appropriate personal vector control measures exposes them to VL infection.
Finally, in 2011, public health facilities were the first choice of treatment in the area (also in Addis Zemen town) [19]. This is probably owed to the quality and free-of-charge nature of health services provided at the VL treatment center located in Addis Zemen Health Center, managed by MSF-Greece during the outbreak, and supported by the Amhara Region Health Bureau and the WHO Leishmaniasis Control Program thereafter. Other studies have shown contrasting results in VL treatment choices, and have associated them with geographical accessibility, treatment costs, confidence in service providers and perceived staff attitude [15], [31], [34]. Our study yielded very low use (≃1%) of traditional medicine for VL treatment compared to other studies in which local healers were consulted by 20 to 50% of the population [17], [35]. These contrasting results may be due to the quantitative methodology used, that may yield different results than qualitative research approaches [14], [34], and may also be influenced by the possible information bias introduced by interviewers being health workers of the area.
Admittedly, our study had some limitations. One of them is that the random sample of HH was taken among kebeles highly affected by VL, and therefore we cannot readily extrapolate our findings to the entire population of Libo Kemkem and Fogera districts. However, we believe that the VL knowledge to be detected among the remaining communities (not so affected by VL) would have been even lower. Another limitation is that in order to compare households with and without past history of VL we stratified according to HH status as reported in May 2009 for both surveys, due to the lack of updated information in February 2011. However, the time period of our study coincided with a low plateau in the transmission of Leishmaniasis in the area [9]. Therefore, we believe that only a small number of HH would have proven misclassified, and the general conclusions of our study would have remained the same. Finally, it is possible that an information bias was introduced by having interviewers that were health professionals of the area. However, we believe that this bias would have affected only to responses related to attitudes, as under our consideration, proper knowledge and the practices observed in the study were not subject to change by the influence of an interviewer.
The VL knowledge in the rural communities of Libo Kemkem and Fogera districts is globally poor, and it should be improved through community strategies. Recommendations are: 1) to conduct sensitization talks in the affected communities, 2) to instruct VL patients and relatives while their stay in the hospital so they can act as health agents in their communities and 3) to follow up the maintenance of bed nets and the use of any other prevention measure like household spraying or environmental sanitation as the high level of acceptance perceived suggests that changes in behaviour are possible.
|
10.1371/journal.pgen.1008292 | polyamine uptake transporter 2 (put2) and decaying seeds enhance phyA-mediated germination by overcoming PIF1 repression of germination | Red light promotes germination after activating phytochrome phyB, which destabilizes the germination repressor PIF1. Early upon seed imbibition, canopy light, unfavorable for photosynthesis, represses germination by stabilizing PIF1 after inactivating phyB. Paradoxically, later upon imbibition, canopy light stimulates germination after activating phytochrome phyA. phyA-mediated germination is poorly understood and, intriguingly, is inefficient, compared to phyB-mediated germination, raising the question of its physiological significance. A genetic screen identified polyamine uptake transporter 2 (put2) mutants that overaccumulate polyamines, a class of antioxidant polycations implicated in numerous cellular functions, which we found promote phyA-mediated germination. In WT seeds, our data suggest that canopy light represses polyamines accumulation through PIF1 while red light promotes polyamines accumulation. We show that canopy light also downregulates PIF1 levels, through phyA; however, PIF1 reaccumulates rapidly, which limits phyA-mediated germination. High polyamines levels in decaying seeds bypass PIF1 repression of germination and stimulate phyA-mediated germination, suggesting an adaptive mechanism promoting survival when viability is compromised.
| Canopy light, unfavorable for photosynthesis, elicits paradoxical responses in seeds. Depending on the timing of irradiation upon seed imbibition, canopy light can either block or promote germination. The promotion effect is mediated by the light sensor phyA and, intriguingly, it is poorly efficient when compared to that of red light, which is favorable for photosynthesis. Why would germination be stimulated by canopy light, which is not favorable for young seedling photosynthesis? We show that phyA-mediated germination is greatly enhanced in decaying seeds, which overaccumulate polyamines, a class of antioxidant molecules that promote germination. We propose an adaptive rationale for phyA-mediated germination: in decaying seeds it becomes advantageous to germinate even under canopy light as a last chance for seedling survival.
| Seeds are capsules maintaining the plant embryo in a resistant state and promoting plant dispersal. To successfully produce a seedling, seeds are able to remain viable over time and block their germination under potentially fatal conditions for the seedling.
However, over time seeds irremediably accumulate oxidative events, which eventually will compromise their viability [1,2]. Seeds limit oxidative damage through antioxidants or physical barriers limiting oxygen access to the embryo [3–5]. Polyamines (PAs) are small polycations ubiquitously present in all living organisms where they regulate numerous fundamental cellular processes, including DNA replication, transcription, translation and post-translational modification, cell proliferation, cell cycle regulation and programmed cell death [6–8]. However, how PAs perform these functions is poorly understood. In response to oxidative stress PA levels rise and protect cells by scavenging reactive oxygen species (ROS) and by increasing antioxidant enzymes activity in leaves [9–13]. However, whether PAs accumulate in Arabidopsis seeds after oxidative stress is unknown.
Arabidopsis seeds control their germination after perceiving abiotic parameters in their environment. This triggers signaling responses leading to opposite level changes of abscisic acid (ABA) and gibberellic acid (GA), two hormones that repress and promote germination, respectively [2].
The quality of light perceived by the light receptor phytochromes phyB and phyA exerts a profound influence on seed germination. Phytochromes control the abundance of the transcription factor PHYTOCHROME-INTERACTING FACTOR 1 (PIF1), a key germination repressor regulating ABA and GA levels in seeds [14]. Red (R) light, favorable for photosynthesis, promotes germination while canopy light, enriched in far red (FR) light, represses germination. Early upon seed imbibition, a R pulse photoconverts phyB into its active PfrB form that interacts with PIF1 and promotes its destruction, thus promoting germination. In contrast, a pulse of FR light photoconverts phyB into its inactive PrB form, which leads to PIF1 stabilization, thus preventing germination [15].
Paradoxically, a second FR light pulse applied later on upon imbibition promotes germination by activating phyA [16–18]. Intriguingly, unlike R light, FR light promotes germination inefficiently and erratically among seed batches. This low efficiency was linked to limiting phyA levels and strong ABA-dependent responses early upon seed imbibition [16,17]. However, regulation of endogenous PIF1 levels by FR light through phyA has not been reported. phyA-mediated germination was interpreted as a last chance to form a seedling despite the presence of unfavorable canopy light [17,19]. Yet, its low efficiency suggests that its physiological significance is not fully understood.
Here we found that recessive mutations in POLYAMINE UPTAKE TRANSPORTER 2 (PUT2) enhance phyA-mediated germination. put2 seeds overaccumulate PAs and we show that PAs stimulate phyA-mediated germination. In WT seeds, our data suggest that upon phyB inactivation after an early FR pulse, PIF1 represses PAs accumulation. phyB activation by R light promotes PIF1 downregulation and PAs accumulation. Upon phyA activation by a second FR pulse, PIF1 is downregulated but, surprisingly, increase in PAs accumulation does not take place. We propose that this differential regulation of PAs levels arises from the duration of PIF1 extinction time, which is longer after R than after FR light irradiation. Accelerated aging procedures stimulated PAs accumulation and markedly enhanced phyA-mediated germination without downregulating PIF1 levels. Our results suggest that decaying seeds bypass PIF1-dependent repression of PAs accumulation to enhance phyA-mediated germination even under unfavorable canopy light conditions.
Hereafter a “FR assay” refers to the procedure where seeds are irradiated with a single far red (FR) pulse early upon imbibition (Fig 1A). In a “FR/Nh/FR assay” seeds are irradiated with two FR pulses separated by an N hour (h) time interval, typically 48 hours (N = 48) (Fig 1A). In a “FR/R assay” seeds are irradiated with a FR pulse early upon imbibition immediately followed by a red (R) light pulse (Fig 1A; see Materials and methods).
To better understand why phyA-mediated germination is limited in seeds, we sought to identify negative regulators of phyA signaling by screening for mutants displaying enhanced phyA-dependent seed germination responses. We generated a transgenic line carrying a firefly LUCIFERASE (LUC) reporter gene under the control of GA3ox1 promoter sequences (pGA3ox1::LUC) [20]. High GA3ox1 expression is characteristic of seeds undergoing germination [21]. pGA3ox1::LUC seeds were mutagenized using ethyl methanesulfonate (EMS).
FR/12h/FR assays lead to lower phyA-dependent germination relative to FR/48h/FR assays [16,17]. Thus, to identify negative regulators of phyA-dependent germination, we screened for pGA3ox1::LUC mutant seeds displaying high LUC bioluminescence and germination in a FR/12h/FR assay (S1A Fig; see Materials and methods for details). Mutants identified in this manner were propagated and the resulting seed progeny was further studied. This led to identify 5 recessive and independent mutants, named fr/fr germination 1–5 (ffg1-ffg5) having enhanced germination in a FR/12h/FR assay relative to the parental non-mutagenized pGA3ox1::LUC (Parent.) line (Fig 1B–1D and S1B–S1D Fig).
Unsurprisingly, among these mutants, we found two mutants (ffg4 and ffg5) having high germination percentage in a FR assay (S1B Fig). ffg4 had a G-to-A transition at nucleotide 370 in PIF1/PIL5 (At2g20180), which resulted in the substitution of Trp-94 with a Stop codon (S1E Fig). ffg5 had a mutation at the splicing site of the third intron of the ABA biosynthetic gene ABA1 (At5g67030) (S1F Fig). PIF1/PIL5 and ABA were previously shown to negatively regulate phyB signaling and were therefore not further studied [17,22,23].
We also identified three independent recessive mutants that did not germinate in a FR assay but showed enhanced phyA-mediated germination (Fig 1B–1D). We selected ffg1 for further study (Fig 1E–1G).
The ffg1 locus was mapped to a 200 kbp interval on chromosome 1 (11.3 ~ 11.5 Mbp.). Sequencing analysis revealed that ffg1 mutants had a G-to-A substitution at nucleotide 376 in the Arabidopsis locus At1g31830 (PUT2 / LAT4 (L-AMINO ACID TRANSPORTER 4) / PAR1 (PARAQUAT RESISTANT 1) / PQR2 (PARAQUAT-RESISTANT 2)), which converts Gly-126 to Arg. This G126R transition is hereafter referred as a put2-1 mutant allele (Fig 2A).
PUT2 encodes an amino acid permease family protein and put2 mutants were reported to display resistance to paraquat (PQ), a methyl viologen widely used as a herbicide [24,25]. We found that put2-1 mutants were resistant to PQ, indicating that PUT2 activity is defective in put2-1 mutants (S2A Fig). This suggested that PUT2 negatively regulates phyA-mediated germination. We observed enhanced phyA-mediated germination in two previously reported independent put2 mutant alleles, par1-1 [24] and par1-5 [24] (hereafter put2-3), and in a new one, named put2-2, identified in this study after screening for PQ-insensitive mutant plants (Fig 2B–2D and S2A Fig; see Materials and methods).
Expectedly, phyAput2 double mutant seeds did not germinate in a FR/48h/FR assay, showing that the high percentage germination of put2-3 mutants in response to FR light is mediated by phyA (Fig 2E–2G and S2B Fig). We also assessed germination percentages using different FR light fluences for the second FR pulse in a FR/48h/FR assay. put2-3 seed germination was enhanced relative to that of WT seeds under all FR fluences considered (Fig 2H).
Altogether, these results confirm that PUT2 encodes a negative regulator of phyA-mediated seed germination. We next sought to understand how PUT2 represses phyA-mediated responses in seeds.
PUT2 belongs to the amino acid/polyamine/organocation (APC) transporter superfamily, which has four homologs in Arabidopsis: PUT1, PUT3, PUT4 and PUT5 [26–28]. The highest homology is found with PUT1 (75% identities, 83% positives, 3% gaps), followed by PUT3 (67%, 82%, 3%), PUT5 (53%, 69%, 7%) and PUT4 (42%, 63%, 4%) (S3 Fig). Publicly available data show that PUT2, PUT3 and PUT4 are expressed in developing, mature and imbibed seeds; however, PUT2 expression in developing seeds or upon seed imbibition is markedly higher relative to that of its homologs (S4 Fig). This suggested that PUT2 is unique among its homologs in negatively regulating phyA-mediated seed germination. Indeed, the germination percentage of put1, put3, put4 and put5 mutant seeds exposed to a FR/48h/FR assay was low and similar to that of WT seeds (Fig 3A–3C and S5 Fig).
PUT2, PUT1 and PUT3 were shown to transport polyamines (PAs) in yeast and plants [27,29,30]. This suggested that enhanced phyA-mediated seed germination of put2 seeds could result from defects in PAs cellular distribution or metabolism in seeds. The most abundant PAs in plants are putrescine (Put), spermidine (Spd) and spermine (Spm). Di-amine Put is made from arginine and is essential for plant survival [31]. Two spermidine synthases (SPDS1 and 2) are responsible for tri-amine Spd synthesis from Put, and spermine synthase (SPMS) is responsible for tetra-amine Spm synthesis from Spd. Spd, but not Spm, is essential for plant survival [32,33].
We measured free Put, Spd and Spm accumulation in seeds exposed to a FR assay 24h after FR light irradiation. In WT seeds, overall PAs levels, i.e. the sum of Put, Spd and Spm levels, were variable between seed batches ranging from about 600 nmol per gram of seed dry weight (nmol/g DW) to 1,800 nmol/g DW (S6A Fig). In every WT seed batch, Spd was the most abundant PA followed by Put and Spm. Overall PAs levels and the relative amounts of Put, Spd and Spm in put1, put3, put4 and put5 mutant seeds were similar to WT seeds (Fig 3D). In contrast, overall PAs levels in put2-3 mutant seeds were systematically markedly higher relative to WT seeds among different seed batches, mostly due to higher Spd and Spm levels (Fig 3D and S6B Fig). In dry seeds, higher PAs levels were also specifically observed in put2-3 mutants (S6C Fig). Similar results were obtained with put2-1, put2-2 and par1-1 alleles (S6D Fig).
These observations therefore suggest that high PAs levels could be responsible for the enhanced phyA-mediated seed germination in put2 mutant seeds, i.e. that PAs are positive regulators of phyA signaling in seeds.
To further study the role of PAs in regulating phyA signaling in seeds, we considered assessing phyA-mediated responses in absence of endogenous PAs. However, PAs regulate fundamental cellular processes and mutations preventing PA biosynthesis in Arabidopsis are lethal, which makes this task challenging [8,31–33].
Instead, we considered identifying seeds with higher PAs levels. Five polyamine oxidases in Arabidopsis (PAO1-PAO5) catabolize PAs by converting Spm to Spd and Spd to Put [34]. Previous reports showed that pao1—pao5 single mutants accumulate higher individual or overall PAs levels [35]. However, in some cases, as with pao4 mutants, lower levels in individual PAs were reported [36]. PAs levels in pao1—pao5 single mutant seeds were not previously reported. We observed moderately higher, i.e. less than twofold, overall PAs levels in pao1—pao5 seeds relative to WT (Fig 3E).
PAO1 and PAO5 are localized in the cytosol whereas PAO2, PAO3 and PAO4 are localized in the peroxisome [34]. We measured PAs levels in pao1pao5 and pao2pao4 double mutants, defective in cytoplasmic and peroxisomal PAO activity, respectively [37]. We found more than twofold higher PAs levels in pao2pao4 double mutants whereas pao1pao5 double mutants had moderately higher, i.e. less than twofold, PAs levels (Fig 3E). Strikingly, only pao2pao4 double mutant seed germination was markedly enhanced in a FR/48h/FR assay (Fig 3F–3H). Furthermore, we observed that addition of individual PAs in the germination plates enhanced phyA-mediated germination (Fig 3I).
Altogether, these observations support the hypothesis that PAs are positive regulators of phyA-mediated seed germination.
We sought to further evaluate this notion by identifying physiological conditions that could enhance PAs levels in seeds and whether they were associated with more efficient phyA-mediated germination.
PAs act as antioxidants in plants where they accumulate in vegetative tissues [12,13] in response to oxidative stress. Seeds irremediably accumulate oxidative events as they age and oxidative stress is a major factor affecting seed viability [38–40]. Interestingly, we observed that the percentage of phyA-mediated seed germination markedly increased with old seed batches, reaching as much as 60% with five-year old seeds (Fig 4A). This experiment was performed with seeds produced at different times, which could lead to differences in germination among seed batches. Nevertheless, these observations are consistent with the notion that phyA-mediated germination increases with oxidative stress.
Whether oxidative stress promotes PAs accumulation in seeds was not previously investigated. To address this question, we subjected WT seeds to a controlled deterioration treatment (CDT), which promotes oxidative stress and artificially accelerates seed aging [4,40]. Increased oxidative stress upon seed exposure to CDT was verified by measuring superoxide O2- levels in seeds (S7A Fig). Next, seeds that had undergone CDT were exposed to a FR/48h/FR assay and PAs levels were measured 24h after the second FR light pulse. As anticipated, PAs levels increased after exposure to 6 days of CDT (Fig 4B and S7B Fig). The increase in PAs levels was not as pronounced as in put2 mutants (Fig 3D).
Consistent with previous results, seeds exposed to white light or a FR/R assay decreased their seed germination percentage from 100% to 90% after exposure to CDT, indicating that seeds that had been exposed to CDT started to lose their viability [4,41] (Fig 4C and S8A, S8B and S8C Fig). In contrast, the germination percentage in a FR/48h/FR assay markedly increased from 12% to 67% after exposure to CDT, i.e. nearer to the germination percentage observed in a FR/R assay (Fig 4C). A similar trend was obtained with independent seed batches exposed to CDT (S8B and S8C Fig).
Altogether, these data further strengthen the notion that high endogenous PAs levels in seeds enhance phyA-mediated germination.
We next sought to better understand 1) what makes phyA-mediated germination less efficient than phyB-mediated germination and 2) whether this low efficiency reflects how PA levels are regulated by light in seeds. These questions were addressed by studying the role of PIF1, a key light regulated germination repressor, in regulating endogenous PAs levels. We first monitored endogenous PIF1 accumulation in seeds exposed to light treatments conducive of phyA-mediated germination, which was not previously reported.
In a FR assay, which blocks WT seed germination, PIF1 levels rapidly increased between 1h and 6h (Fig 5A and S9 Fig), remained high between 6h and 24h and slowly decreased thereafter, consistent with previous reports [42]. In contrast, phyA levels slowly increased between 1h and 12h and remained roughly similar between 24h and 48h, consistent with previous reports [17], and further slowly increased between 48h and 96h (Fig 5A).
The phyA-mediated germination percentage of WT seeds exposed to a FR/12h/FR, and FR/48h/FR was 0% and 2%, respectively, whereas in seeds exposed to a FR/96h/FR assay it jumped to 37% despite the modest increase in phyA levels between 48h and 96h (Fig 5B). This is consistent with previous results showing that phyA levels and the percentage of phyA-mediated germination are not linearly correlated [17]. In contrast, and as expected, close to 100% phyB-mediated germination was observed in WT seeds exposed to a FR/R assay (Fig 5B).
We monitored endogenous PIF1 accumulation in seeds exposed to the different above FR/Nh/FR assays (N = 12, 48, 96), conducive of phyA-mediated germination. Irrespective of the time of its application, the second FR pulse did not affect markedly phyA protein levels (Fig 5C). Unexpectedly, however, the second FR pulse triggered, within one hour, rapid PIF1 downregulation in all assays (Fig 5C). As expected, endogenous PIF1 downregulation was not observed in phyA mutant seeds but observed in phyBCDE mutant seeds, showing that PIF1 downregulation is driven by phyA [43] (Fig 5C).
Thus, and surprisingly, FR and R light are similarly able to downregulate PIF1 levels even though they do not stimulate germination with the same efficiency. We therefore wondered whether the duration of PIF1 extinction time could be different after R and later FR light irradiation. Indeed, the duration of PIF1 extinction time was about 12h longer after a R pulse than after a second FR pulse in both a FR48hFR assay or a FR96hFR assay (Fig 5D).
Altogether these observations show that phyA levels in seeds are not limiting to promote PIF1 downregulation in response to a second FR light pulse applied at different times upon imbibition. They rather suggest that phyA-mediated germination is inefficient, at least in part, due to the short duration of PIF1 extinction following FR irradiation.
Interestingly, publicly available data indicate that expression of four polyamine biosynthesis genes is higher in pif1 mutant seeds relative to WT seeds exposed to a FR assay [44] (S10A–S10D Fig). In contrast, expression of PUT2 is low in pif1 seeds relative to WT seeds exposed to a FR assay [44] (S10E Fig). Furthermore, expression of these genes was similar in WT and pif1 seeds exposed to a FR/R assay (S10A–S10E Fig). This suggested that PIF1 represses PAs accumulation in seeds exposed to a FR assay.
To evaluate this possibility, we measured PAs levels in WT and pif1 seeds exposed to a FR assay. Consistent with this hypothesis, PAs levels were higher in pif1 mutant seeds relative to WT seeds exposed to FR assay (Fig 6A). These data are consistent with the view that PIF1 represses PAs accumulation in seeds after an early FR light pulse.
Next, we investigated whether PAs levels increase under conditions conducive of phyA-mediated germination. Unlike WT seeds exposed to a R pulse, PAs levels did not change after a second FR pulse relative to seeds exposed to a single FR pulse (Fig 6B and S11 Fig).
Altogether, these results suggest that the short duration of PIF1 extinction upon a second FR light pulse irradiation does not permit the elevation of PAs levels in seeds. As a result, low PAs levels would contribute to the low efficiency of phyA-mediated germination (see model below).
We next investigated whether increased phyA-mediated germination in put2 mutants or in WT seeds that had undergone CDT was linked to changes in phyA and PIF1 levels in seeds. In a FR assay, PIF1 levels in put2-3 mutant seeds were similar to those in WT seeds (Fig 7A). put2-3 mutant seeds accumulated normal phyA levels up to 12h after imbibition (Fig 7A). At 48h phyA levels were higher in put2-3 seeds relative to WT seeds; we did not further investigate this matter.
Strikingly, a FR/12h/FR assay markedly stimulated germination of put2-3 seeds relative to WT seeds (Fig 7B), even though they accumulated similar phyA levels (Fig 7C; 0 h and S12A and S12B Fig). Higher germination percentage of put2-3 seeds was not associated with obvious differences in PIF1 levels, as assessed by data quantification in experiments with biological replicates, including in the duration of PIF1 extinction time after the second FR pulse (Fig 7C and S12A and S12B Fig). Furthermore, in a FR/48h/FR assay the duration of PIF1 extinction time after the second FR pulse was similar between WT and put2-3 seeds even though the percentage of put2-3 seed germination was 100% whereas that of WT seeds was only 2% (Fig 7B and 7D and S13A–S13C Fig).
After an early FR pulse, WT seeds that had undergone CDT had no obvious changes in phyA or PIF1 levels relative to untreated WT seeds up to 48h after FR light irradiation, as assessed by data quantification in experiments with biological replicates (Fig 7E and S14A Fig). After a second FR pulse (FR/48h/FR assay), the percentage of germination of WT seeds exposed to CDT was enhanced compared to unexposed WT seeds without obvious changes in phyA or PIF1 levels, as assessed by data quantification in experiments with biological replicates (Fig 7E and 7F and S14B–S14D Fig).
These observations show that enhanced phyA-mediated germination in put2 seeds or WT seeds exposed to CDT can take place without marked changes in PIF1 levels. They therefore suggest that oxidative stress promotes PA accumulation, which in turn promotes phyA-mediated germination downstream of PIF1 or else independently of PIF1 (Fig 8).
Here we sought to better understand how seed germination is promoted by canopy light through phyA. We focused on understanding why this process is poorly efficient and its physiological relevance.
Downstream phyA signaling components in seeds are poorly characterized. Their identification and study is rendered difficult by the fact that observing phyA-mediated germination requires first blocking germination through phyB inactivation. In the case of pif1 mutants, which fully germinate after an early FR pulse, germination mediated by phyA cannot be observed. Nevertheless, it is generally assumed that PIF1 represses germination downstream of phyA [23,43].
Here we found that put2 mutants are specifically enhanced in phyA-mediated germination (Figs 1 and 2), which was not previously reported. PUT2 encodes a PAs transporter and we showed that among other mutants deficient in homologous transporters only put2 mutants accumulate high PAs levels in seeds (Fig 3). Similarly, mutants deficient in PAs catabolism or WT seeds exposed to CDT increased PAs levels in seeds and enhanced phyA-mediated germination (Figs 3 and 4). Thus, our study provides correlative evidence suggesting that increased PAs levels in seeds positively regulate phyA-mediated germination.
High PAs accumulation in put2 seeds could result from inappropriate distribution of PAs in cells. PUT2 is localized in the Golgi apparatus and the chloroplast [24,45]. A put2 mutant cell sensing low PAs accumulation in the Golgi or chloroplast might respond with increased PAs synthesis as a compensatory mechanism. put2 had increased Spd and Spm levels whereas WT seeds exposed to CDT and pao2pao4 mutants had mainly higher Spd levels. Furthermore, PUT2 preferentially transports Spd in yeast cells [29]. This could suggest that Spd is important to enhance phyA-mediated germination. However, exogenous application of Put, Spd and Spm similarly promoted phyA-mediated germination (Fig 3). Thus, our study does not pinpoint a particular individual PA specifically promoting phyA-mediated germination. Understanding the role of an individual PA in phyA signaling using genetic approaches is difficult since, to our best knowledge, there are no reported PA biosynthesis mutants accumulating a single PA.
It remains to be understood how PAs promote phyA signaling in seeds. L-arginine is a precursor of Put and S-adenosylmethionine is a precursor of Spd and Spm. However, they are also precursors of nitric oxide (NO) and ethylene, respectively. Indeed, enhanced PAs accumulation was linked to an increase in NO levels and associated with both an increase and decrease in ethylene signaling [46–48]. NO and ethylene repress ABA responses in seeds and therefore PAs could enhance germination by repressing ABA signaling through NO or ethylene [49]. Against this possibility, we found that put2 mutants have normal ABA responses in seeds, consistent with a previous report showing that par1 mutants have normal ABA responses [24] (S15 Fig). PAs are involved in a wide range of fundamental cellular processes including DNA replication, transcription, translation and post-translational modification [7]. PAs were also speculated to participate in abiotic stress signaling in plants [50]. In all cases, the mechanism by which PAs act is poorly understood as they are difficult to study due to their ubiquitous presence in cells and their essential function for survival [6].
Our data suggest that endogenous PAs accumulation is repressed by PIF1 in WT seeds exposed to an early FR pulse. This indicates that PAs biosynthesis is regulated downstream of PIF1. The biological significance of the regulation of PAs biosynthesis gene expression and PAs levels in seeds by light remains to be understood. On the other hand, PAs levels were higher in put2 mutants or in WT seeds exposed to CDT and the resulting increase in phyA-mediated germination took place without changes in PIF1 levels. Therefore, this suggests that PAs levels can be regulated independently of PIF1 to regulate phyA-mediated germination. PAs could promote phyA-mediated germination downstream of PIF1. However, transcriptomic studies have suggested that phyA activation by FR also triggers gene expression changes independently of PIF1 [18]. Thus, PAs could promote phyA-mediated germination in a PIF1-independent manner (Fig 8; see putative “X” pathway in the model).
PAs levels in seeds increased in response to R light but not in response to a second FR light pulse (Fig 6). We propose that this is due to the short PIF1 extinction time following FR irradiation. This shorter time was not due to limiting phyA levels since it remained unchanged in WT seeds exposed to a FR/96h/FR assay or in put2 seeds exposed to a FR/48h/FR assay, which had higher phyA levels (Figs 5 and 7). Therefore, PIF1 reaccumulation is differently regulated after R and FR light irradiation, respectively. The underlying mechanism accounting for this differential regulation remains to be identified. It is also unknown why put2 mutants accumulate higher phyA levels at later time points upon seed imbibition.
It is generally accepted that PAs act as antioxidants in plants [51]. Seed oxidation is an unavoidable process compromising seed viability in the dry seed state. It is therefore expected that seeds have evolved adaptive mechanisms to sense oxidative damage and adapt their behavior such as their control of seed germination. In newly produced seeds, the first level of germination control is that of primary seed dormancy, a trait whereby germination is blocked even under favorable conditions. Dormancy prevents germination out of season. Seeds lose dormancy during a period of dry storage called dry after-ripening [2]. Seed oxidation is known to accelerate the release of seed dormancy during after-ripening [1]. Dormancy was shown to inhibit R- and phyB-mediated germination [52] and therefore it is expected that oxidation promotes phyB-mediated germination. However, the oxidation events that release dormancy are not sufficient to promote germination mediated by phyA: 18-month-old WT seeds, which have fully lost dormancy and fully germinated after a R pulse, germinated at less than 5% in a FR/48h/FR assay (Fig 2).
Our results therefore suggest that there are two levels of seed germination regulation through oxidative stress. Younger and still viable seeds, with moderate levels of oxidative damage, have an advantage to repress their germination under canopy light, which is unfavorable for photosynthesis. As seeds age, however they decay as a result of continuous accumulation of oxidative damage, which compromises their capacity to form a viable seedling. It would then become advantageous to germinate under a broader range of light wavelengths, including canopy light [19]. In this context, increased PAs levels upon seed imbibition could serve a dual function: protect the decaying seed from the oxidative damage that accumulated during dry after-ripening while promoting in parallel phyA-mediated germination under unfavorable light cues such as canopy light. This would represent a mechanism providing a last chance for plant survival (Fig 8).
Arabidopsis T-DNA insertion lines, all in the Col-0 background, were obtained from the Nottingham Arabidopsis Stock Centre with the following accession numbers: put2-3; SALK_119707, put1; SAIL_270_G10, put3; SALK_206472, put4; SAIL_1275_C06, put5; SALK_122097, pao1-2; SAIL_882_A11, pao2-4; SALK_046281, pao3-1; GABI_209F07, pao4-1; SALK_133599, pao5-2; SALK_053110 and phyA-211; N6223. phyA-211put2-3 double mutants were generated after crossing phyA-211 and put2-3 plants. par1-1 [24], pao1pao5 [37], pao2pao4 [37] and phyBCDE [53] seeds were kindly provided by Jianru Zuo (Chinese Academy of Sciences, China), Tomonobu Kusano (Tohoku University, Japan) and Pablo D. Cerdán (Fundación Instituto Leloir, Argentina), respectively.
All genotypes tested in each experiment were grown together under the same conditions and seeds were harvested the same day and allowed to after-ripen at room temperature for at least one month. For the germination assays, seeds were surface sterilized and 50–60 seeds of each genotype were sown on MS medium (Sigma) containing 0.8% (w/v) agar without seed stratification. For the germination assays in presence of polyamines, individual polyamines (Sigma) were added to the germination medium. In a FR assay, seeds were irradiated with a FR pulse (3.69 μmol m-2 s-1) for 5 min after 2 h seed imbibition under white light. In a FR/Nh/FR assay, seeds were irradiated with a first FR pulse (3.69 μmol m-2 s-1) for 5 min and further irradiated with a second FR pulse (3.69 μmol m-2 s-1 or as indicated in each experiment) for 5 min after N (e.g. 12, 48 and 96) hours of dark incubation. In a FR/R assay, seeds were irradiated with a red (R) pulse (14.92 μmol m-2 s-1) for 5 min followed by a FR pulse. In all assays, light irradiated plates were kept in the dark for the indicated times. Thereafter a seed that had undergone endosperm rupture, i.e. radicle protrusion, was scored as a germination event. All the germination assays were performed with three technical replicates and the results were confirmed with at least two or three independent biological seed samples. Data value distribution among biological samples is shown by scatterplots as described in Weissgerber et al. (2015) [54].
Approximately 3kb of GA3ox1 promoter region [20] was amplified with primers (5’-CGCGGATCCCACCAGAGTGTGTGCTACATGC-3’ and 5’-CCGCTCGAGAACACAGCAGGCAGCTTGCTC-3’). BamHI and XhoI restriction sites were used for cloning into binary vector pGPTVII [55]. WT (Col-0) plants were transformed with a firefly luciferase (LUC) reporter gene under the control of GA3ox1 promoter sequences (WT/pGA3ox1::LUC). With the aim of identifying mutants displaying enhanced phyA-mediated seed germination responses, a population of 20,000 WT/pGA3ox1::LUC seeds (M0) was chemically mutagenized using 0.3% ethyl methanesulfonate (EMS) as previously described [56]. In M2 populations, mutants able to germinate in a FR/12h/FR assay or displaying high LUC bioluminescence were selected for further analysis (Fig 1 and S1 Fig). LUC bioluminescence was performed as previously described [55]. Briefly, plants exposed to FR/12h/FR assay were sprayed with a luciferin (Biosynth) solution (315 μg/ml), under green safety light, 24h after the second FR pulse and examined using an Aequoria dark box with a mounted ORCAII CCD camera (Hamamatsu). This led to identify three recessive and independent mutants (ffg1—ffg3) having enhanced bioluminescence and germination in a FR/12h/FR assay relative to the parental non-mutagenized WT/pGA3ox1::LUC line. The same mutagenized seed population was used to identify put2-2 as previously described [24,25]. Briefly, the same mutagenized seed population was sown on a germination medium with 0 and 10 μM of PQ and cultured under WLc for 6 days to reveal PQ-insensitive mutants, which led to the identification of the put2-2 allele.
For map-based cloning and whole genome sequencing, the ffg1 mutant was outcrossed to Ler. The ffg1 locus was mapped as previously described [57]. Briefly, a combination of cleaved amplified polymorphic sequences (CAPS) markers and simple sequence length polymorphisms (SSLPs) markers was used for fine mapping. The ffg1 locus was mapped to a 200 kbp interval on chromosome 1 (11.3 ~ 11.5 Mbp). Identification of mutations in this interval was done after sequencing the ffg1 mutant genome. Genomic library preparation was performed using TrueSeq® DNA Library Prep Kit (Illumina) according to manufacturer’s instructions. Sequencing was performed using HiSeq2000 (Illumina). This led to the identification the put2-1 allele in the 200 kbp interval.
To avoid detecting differences in PAs levels arising from plants at different development stages, we measured endogenous PAs levels in dry or non-germinated seeds that are harvested under green safety light. Standards of polyamines as well as the other chemicals and reagents were purchased from Sigma Aldrich chemical company (St. Louis, MO, USA). Free PAs were isolated and derivatized by slightly modified method of previous report [58]. A 250 mL of 5% trichloroacetic acid (TCA) was added to a 5 mg of lyophilized seeds and homogenized using ZrO2 beads (3 mm) in mixer-mill for 5 min at 27 Hz. Sample was then sonicated for 10 min at 25 °C, and after centrifugation at 12,400 × g for 5 min, supernatant was quantitatively transferred into another vial.
A 500 μL of 2M NaOH was added following with 2.5 μL of benzoyl chloride (in methanol 50:50, v:v), and after vortexing for 5 sec reaction mixtures are left for 40 min at 25 °C. A 500 μL of saturated NaCl was added and benzoylated polyamines were extracted with 2 × 500 μL of diethyl ether. Ether was evaporated and dry samples were stored at -80 °C until analysis.
All samples were dissolved in 50 μL of mobile phase (45% methanol in 15 mM formic acid, pH 3.0), sonicated for 15 min, and centrifuged for 5 min at 12,400 × g prior to the analysis. Diaminohexane (DAH) was used as internal standard.
Ultra-high performance liquid chromatography-tandem mass spectrometry (UHPLC-MS/MS) was performed on UltiMate™ 3000 liquid chromatographic system consisting of binary pumps, an autosampler and a column thermostat coupled to a TSQ Quantum Access Max triple quadrupole mass spectrometer (Thermo Fisher Scientific, Waltham, MA, USA).
Chromatographic separation was performed on an Acquity UPLC BEH C18 (50 × 2.1 mm; 1.7 μm particle size) column (Waters, Milford, MA, USA) with appropriate pre-column kept at 40 °C. The mobile phase consisted of a mixture of aqueous solutions of 15 mM formic acid adjusted pH 3.0 with ammonium hydroxide (Solvent A) and methanol (Solvent B). The analytes were separated using a binary gradient starting at 45% of B for 2.7 min, then increase to 57% for 0.3 min, isocratic at 57% for next 2.5 min, increase to 100% B for next 0.1 min, isocratic at 100% B for next 1 min, and decrease to 45% B for next 0.1 min. Finally, the equilibration to the initial conditions took 2.3 min. The flow rate was 0.4 mL/min and the injection volume 5 μL. Benzoylated PAs were detected in positive ionization mode electrospray ionization (ESI+). The selected reaction monitoring (SRM) transitions for benzoylated putrescine (Put) were 297 > 105, and 297 > 176 at 20 eV collision energy (CE), for spermidine (Spd) 458 > 175, and 458 > 233 at 25 eV CE, and for spermine (Spm) 619 > 162, 619 > 337, and 619 > 497 at 30 eV CE. The spray voltage was set to 3 kV, the vaporizer temperature to 350 °C, and the ion transfer tube temperature to 320 °C, respectively.
CDT was performed as described in a previous report [4]. Briefly, Col-0 dry seeds were stored in a closed container during the time period indicated in each experiment. The container was maintained at 37 °C and its interior had around 82% of relative humidity imposed by the presence of a saturated salt (KCl). Then, seeds were dried back at 30% relative humidity (room temperature) and stored at -80 °C until they were used for superoxide O2- levels measurements, germination assays or protein gel blots analysis.
The levels of superoxide O2- were determined based on its ability to reduce nitro blue tetrazolium (NBT) to blue formazan as previously described [59,60]. Dry seeds (0.035g) were grinded in 1.3 mL of incubation solution (10 mM K-phosphate buffer pH 7.8, 10 mM NaN3, 0.05% NBT). After 30 min of incubation at room temperature, grinded tissue was collected at the bottom of the tube by centrifugation (8000 x g, 5 min, room temperature). Supernatant was diluted 10 x in incubation buffer, heated 85 °C for 15 min and then cooled on ice. Absorbance at 580 nm was measured by spectrophotometer to quantify NBT levels.
PIF1 recombinant proteins were prepared using PIF1-his DNA (pET21a) provided by Enamul Huq (University of Texas at Austin, USA), and induced and purified using a commercial kit according to manufacturer’s instructions (Amersham). Polyclonal anti-PIF1 was obtained from rabbits immunized with PIF1 recombinant protein. PIF1 antibodies were further affinity-purified using PIF1 recombinant protein immobilized on nitrocellulose filters as described [61]. FR or R light irradiated WT and mutant seeds were harvested under green safety light at the time indicated in each experiment. 20 seeds were homogenized with homogenization buffer (0.0625 M Tris-HCl at pH 6.8, 1% [w/v] SDS, 10% [v/v] glycerol, 0.01% [v/v] 2-mercaptoethanol), and total proteins were separated by SDS-PAGE gel and transferred to a PVDF membrane (Amersham). PIF1 and UGPase proteins were detected using 1:500 dilution of anti-PIF1 or 1:10,000 dilution of anti-UGPase (Agrisera), and anti-rabbit IgG HRP-linked whole antibody (GE healthcare) in a 1:10,000 dilution was used as a secondary antibody. PHYA proteins were detected using 1:5,000 dilution of anti-phyA (kindly provided by Akira Nagatani; Kyoto University, Japan) and the anti-mouse IgG HRP-linked whole antibody (GE healthcare) in a 1:10,000 dilution was used as a secondary antibody [17]. Quantification of band intensity was performed using imageJ.
|
10.1371/journal.pcbi.1005364 | Two dynamic regimes in the human gut microbiome | The gut microbiome is a dynamic system that changes with host development, health, behavior, diet, and microbe-microbe interactions. Prior work on gut microbial time series has largely focused on autoregressive models (e.g. Lotka-Volterra). However, we show that most of the variance in microbial time series is non-autoregressive. In addition, we show how community state-clustering is flawed when it comes to characterizing within-host dynamics and that more continuous methods are required. Most organisms exhibited stable, mean-reverting behavior suggestive of fixed carrying capacities and abundant taxa were largely shared across individuals. This mean-reverting behavior allowed us to apply sparse vector autoregression (sVAR)—a multivariate method developed for econometrics—to model the autoregressive component of gut community dynamics. We find a strong phylogenetic signal in the non-autoregressive co-variance from our sVAR model residuals, which suggests niche filtering. We show how changes in diet are also non-autoregressive and that Operational Taxonomic Units strongly correlated with dietary variables have much less of an autoregressive component to their variance, which suggests that diet is a major driver of microbial dynamics. Autoregressive variance appears to be driven by multi-day recovery from frequent facultative anaerobe blooms, which may be driven by fluctuations in luminal redox. Overall, we identify two dynamic regimes within the human gut microbiota: one likely driven by external environmental fluctuations, and the other by internal processes.
| Dynamics reveal crucial information about how a system functions. In this study, we develop an approach for disentangling two types of dynamics within the human gut microbiome. We find that autoregressive dynamics involve recovery from large deviations in community structure. These recovery processes appear to involve the blooming of facultative anaerobes and aerotolerant taxa, likely due to transient shifts in redox potential, followed by re-establishment of obligate anaerobes. Non-autoregressive dynamics carry a strong phylogenetic signal, wherein highly related taxa fluctuate coherently. These non-autoregressive dynamics appear to be driven by external, non-autoregressive variables like diet. We find that most of the community variance is driven by day-to-day fluctuations in the environment, with occasional autoregressive dynamics as the system recovers from larger shocks. Despite frequently observed disruptions to the gut ecosystem, there exists a returning force that continually pushes the gut microbiome back towards its steady-state configuration.
| Microbial ecology has become an important branch of medical science [1]. Recent work has shown how each person maintains a fairly unique microbial fingerprint [2–4], and that microbial dysbioses are often associated with shifts in health-status [5–8]. We now recognize that our microbiota are highly dynamic, and that these dynamics are linked to ecological resilience and host health [9–11]. The field has not yet settled upon whether gut microbial community structure varies continuously or if it jumps between discrete community states, and whether these states are shared across individuals [12–14]. In particular, some researchers suggest that gut communities can be binned into discrete ‘enterotypes’ [12], while others argue that gut communities vary along multidimensional continua [13]. If the ultimate goal of microbiome research is to improve human health by engineering the ecology of the gut, we must first understand how and why our microbiota vary in time, whether these dynamics are consistent across humans, and whether we can define ‘stable’ or ‘healthy’ dynamics.
Gut microbiota are continually buffeted by external factors like diet and host behavior [10, 15]. Internal species-species (e.g. cross-feeding or successional turnover) and host-species (e.g. immune system regulation or host physiology) interactions can also influence the structure and stability of the gut ecosystem [16–18]. Lotka-Volterra (LV) models, which predict changes in community composition through defined species-species or species-resource interaction terms, are popular for describing these internal ecological dynamics [19–22]. LV models are deterministic and fairly straight-forward to interpret, but little is known about the relative importance of these purely autoregressive factors in driving gut microbial dynamics (see Theoretical Considerations section below for a more detailed comparison of LV and VAR models). More recently, a model-free approach to forecasting non-linear dynamics—called convergent cross-mapping (CCM)—has been applied to ecological time series data [23]. While extremely useful, CCM can be difficult to interpret [24, 25], and may not be appropriate for high-dimensional systems with weak coupling between components (e.g. the gut).
Long, dense time series data are becoming increasingly available to microbial ecologists [10, 11, 26]. These temporal data are invaluable for understanding the behavior of microbial communities but require special care during analysis due to the non-independence of temporally adjacent samples [27]. In this paper we conduct a meta-analysis of the four longest, densest human gut time series currently available [10, 11].
We separated microbial dynamics into autoregressive and non-autoregressive components by applying vector autoregressive (VAR) models, which were originally developed for econometrics [28–30]. We took this approach because we found that substantial autocorrelation persisted in most microbial time series for at least 3 days (i.e. past values of an OTU were predictive of its current value), which meant that temporally adjacent samples were not independent. VAR models are standard for analyzing stationary multivariate time series with autocorrelation, cross-correlations, and noise. Time series are considered to be stationary if they appear to be sampled from the same probability distribution through time (i.e. the mean and variance, along with the other moments of the distribution, do not change through time; Fig 1). VARs model each element as a linear function of lagged values of other elements in the time series [29]. In order to reduce the number of coefficients generated by classical VAR models (i.e. with n species in a VAR model with p lags, n2 x p coefficients are generated) and avoid over-fitting we apply regularized estimation, resulting in a sparse VAR (sVAR) [30, 31]. The residual variation of an sVAR model is stripped of much of its autoregressive structure, which allows for the application of standard statistical techniques that assume sample-to-sample independence. sVARs have the benefit of explicitly modeling error, unlike LV-type models [32], and are more straightforward to interpret than CCM forecasting [25].
We suggest that there are two differentiable sets of drivers generating autoregressive and non-autoregressive microbial community dynamics in the gut. The first set of drivers induce multi-day recovery processes, where the past state of the system is predictive of the future state [33]. As mentioned above, LV models are differential equations that can incorporate species-species and species-resource interactions. These models are inherently autoregressive and are the dominant workhorses of ecological time series modeling [19, 26, 34]. The second type of driver is non-autoregressive, and likely includes dietary factors and other external perturbations [10]. In this paper, we show that there are indeed two dynamic regimes: auto-regressive and non-autoregressive. These dynamic regimes appear to reflect internal and external drivers, respectively. The emerging picture of the gut microbiome shows a dynamically stable system, continually buffeted by internal and external forces and recovering back toward a conserved steady-state.
In this section, we directly compare our VAR modeling approach to the more common generalized Lotka Volterra family of models (gLV). gLVs are first-order differential equations that model growth rates as a non-linear function of community composition, and thus assume the existence of mechanistic coupling between variables in the system. VAR models, by contrast, assume linear dynamics, but can only be applied when the observed data are, or have been transformed such that, the time series are stationary. In both gLV and VAR models, dynamics are defined by species-species (or species-resource) interaction terms. In this study, we find that a linear VAR-based approach was sufficient to extract essential dynamics of the system without the need for a nonlinear mechanistic framework (i.e. as implemented in gLV).
gLVs have the following structure:
dXidt=aiXi(t)(1−Xi(t)K)+Xi(t)∑j=1(j!=i)nBijXj(t)
where t is time, ai is the self-interaction term for organism i, Xi is the abundance of organism i, Bij is the interaction term between organisms i and j in a community composed of n organisms.
Dividing by abundance and converting to difference equation form allow for gLV parameters to be solved with a system of linear equations as for a VAR(1) process [22]:
log(Xi(t))−log(Xi(t−1))=ai−aiK*Xi(t−1)+∑j=1(j!=i)nBijXj(t−1)
We can now compare this linearized gLV to a corresponding VAR(1) process:
Xi(t)=qi+si*Xi(t−1)+∑j=1(j!=i)nRijXj(t−1)+ei(t)
where ei(t) is the error term. Thus, both gLV and VAR(1) can be solved using systems of linear equations, but the interpretation of the coefficients will be different. For instance, we can compare ai (1/time) and for qi (abundance), to see that VAR models directly model data on observed data, whereas gLV models assume that a more appropriate model maps observed abundance data to differenced log-transformed data.
Further, VAR(p) models can include an arbitrary number of time lags (i.e. autoregressive processes can extend further back in time). gLVs do not allow for explicit inclusion of historical time series data beyond one time lag. In considering the appropriate approach for analysis of microbial community time series, we created a decision tree to guide appropriate modeling strategies (Fig 2).
Database (Greengenes) OTUs accounted for 95–99%, 93–99%, 83–97%, and 83–96% of all sequences per time point in the F4, M3, DA, and DB time series, respectively. The proportion of non-database OTUs was quite stable over time. Furthermore, the 50 most abundant OTUs in each time series were all Greengenes OTUs. Non-database OTUs tended to be low-abundance taxa.
Microbial community alpha diversity showed stable, mean-reverting behavior across all four time series (Fig 4). The average effective number of species (Neff = e[Shannon diversity]) was between 28–50 for each time series, which indicated that compositional effects (i.e. spurious correlations caused by non-independence between relative abundances) were not a major concern in this analysis (Fig 4) [51]. Friedman and Alm (2012) found that, in simulated data, for a Neff of ~30 or more, conventional statistics gave the same result as their compositionally aware method. We found that large deviations in alpha diversity were strongly correlated with a set of conditionally rare taxa (CRTs) that occasionally bloomed to as much as ~30% of the community (dominant taxa at steady-state were usually between 10–20% of the community), but were usually found at very low abundances (Fig 4). In order to test whether CRT blooms drove significant compositional effects in the correlation structure of our time series, we calculated Spearman’s correlations between the 100 most abundant taxa in the M3, DA, and DB time series with and without the CRT time points (i.e. time points where CRTs rose above 10% relative abundance were removed; S1 Fig). Overall, we saw no compositional effects due to CRT blooms (S1 Fig). A recent meta-analysis by Shade and Gilbert (2015) found that CRTs are responsible for a significant fraction of overall community dynamics in many different ecosystems [43]. Abundant gut CRTs—Prevotella, Bacteroides fragilis, Akkermansia muciniphila, Lachnospiraceae, Enterobacteriaceae and Haemophilus parainfluenzae—were present in M3, DA, and DB time series (S1 Table; Fig 4). Abundant CRTs (i.e. peak abundance ≥ 10% of sequence reads) were not identified in the F4 time series, consistent with its relatively stable alpha-diversity trace (Fig 4). However, there were several Enterobacteriaceae blooms in the F4 time series that fell beneath our abundance threshold. The M3 time series was an outlier, with much more frequent CRT blooms than the other time series. These bloom events likely represent opportunistic or pathogenic organisms that are either the cause or symptom of a disruption in the normal steady-state gut environment. Many CRT OTUs are facultative anaerobes or aerotolerant taxa (e.g. Enterobacteriaceae OTUs and other Proteobacteria or OTUs in the B. fragilis group, like B. uniformis, B. ovatus), which are probably responding to changing redox potential in the gut following some disturbance (e.g. inflammation) [52].
After first-differencing (i.e. differencing OTU abundances at adjacent time points), abundant OTUs showed completely stationary dynamics, implying that the mean, variance, and autocorrelation structure do not change over time (ADF and KPSS tests; Table 1). Both ADF and KPSS Level tests (Table 1) assess whether or not a process has a unit root, which implies that the mean changes through time and that the system does not recover to the mean trend in the presence of a shock (i.e. the process is integrated). The KPSS Trend test (Table 1) assesses whether or not a process is trend-stationary, which implies a mean trend with stationary error that is capable of recovering to the trend line following a shock. Non-stationary OTUs are likely not stable members of the steady-state community. Prior to first-differencing, most OTU trajectories showed ADF-stationary dynamics (Table 1). There was no apparent enrichment for particular taxonomic groups among non-stationary OTUs. In addition to OTU abundance trajectories being largely stationary, we also saw a range of autocorrelation decay curves for the top 50 most abundant OTUs (Fig 5). Some OTUs showed strong, persistent autocorrelation, while others did not (Fig 5). The amount of autocorrelation decay also varied across time series. In particular, the DA time series showed much less autocorrelation than the other three time series (Fig 5).
Stationarity implies that there is a restoring force on an OTU's abundance over time, so that it returns to a mean value after a perturbation (i.e. a steady-state population size, or ‘carrying capacity’). Indeed, we found a significant negative correlation between the change in OTU abundances between time t and time t+1 and the abundances of OTUs at time t (Fig 6). Thus, by converting abundance dynamics to rate dynamics, we can achieve stationarity across all abundant time series in the system and preserve correlation structure between OTUs, as long as we exclude non-stationary windows that contain large disturbances (e.g. food poisoning in the DB time series on day 150). Furthermore, the strong correlation between rates and OTU abundances at the prior time step indicates that much of the variance in these data is linear (Fig 6). The fact that these rate dynamics can be modeled as linear, stationary processes opens up a wide array of time series models to gut data sets.
Prior work has argued for the existence of discrete gut community configurations across humans, termed ‘enterotypes’, that may be associated with health and disease [12, 14]. However, longitudinal data has provided evidence that these enterotypes may arise from undersampling individuals through time. For example, the M3 time series is known to moved fluidly through all three putative enterotypes from Arumugam et al. (2010) during the course of a year [13], which suggests that gut communities vary along a continuum. We expanded upon this analysis using an improved Dirichlet multinomial mixture model (DMM) clustering method [14, 42], which assesses whether samples appear to be pulled from a common Dirichlet multinomial distribution. In general, clustering methodologies are plagued by over- and under-sampling issues. In a time series, samples that are taken close together in time are likely to be similar to one another (i.e. they are autocorrelated). If the community is sampled densely enough as it moves through state space, then packets of samples that happen to be temporally adjacent may be grouped into pseudo-clusters (i.e. over-sampling). On the other hand, if only a handful of samples are taken from an individual, then outlier points may give rise to pseudo-clusters because the state space was not sufficiently sampled (i.e. under-sampling). If there are no discrete states to be found, then the number of states should decay smoothly to 1 as the sampling sparcity is increased. However, if discrete states do exist, then the number of states should reach a plateau that is stable across a range of sampling densities.
We aimed to address whether gut communities can be grouped into discrete states and whether these states are shared across individuals given a range of intra-individual sampling effort. DMM states were fit to the full-length time series, which included the food poisoning event in the DB time series. Independent of our OTU filtering method (50 most abundant OTUs in each time series, or 76 abundant OTUs shared across all time series), we found that the number of DMM states decayed rapidly from 15 to a plateau of 6 (Fig 7A). These states were almost entirely unique to individuals (Fig 7B). There was only one case where a sample from DA was assigned to a state largely associated with DB. In addition, only DB and M3 time series harbored multiple states, which is consistent with major perturbations in the gut community within these time series (i.e. food poisoning and frequent CRT blooms, respectively). Overall, barring major perturbations, we conclude that individual humans—given sufficient sampling density—can be distinguished by unique Dirichelet multinomial distributions (Fig 7). Because almost every time point from a subject’s time series is assigned to the same one or two DMM states, continuous methods are necessary for exploring the dynamics of microbial communities within an individual. We believe this result will hold as larger numbers of long-term human gut time series become available.
Despite the clustering of individuals into unique states, we identified 956 OTUs that were present across all four individuals. These organisms tended to be found at similar median abundances across individuals (Fig 8). In addition, these shared OTUs made up 70–80% of the sequences in each data set (Table 2). Furthermore, a single OTU (from the genus Bacteroides; Greengenes ID 850870) was the most abundant taxon in M3, F4 and DB, and was the second most abundant organism in DA (Fig 6). Thus, despite the fact that most OTUs appear to be unique to an individual’s gut, there is a set of abundant core taxa that are present at similar abundances across people. It is unlikely that compositional effects are responsible for the existence of these carrying capacities, due to the fact that there is a wide-range in of abundances at which OTUs appear to persist (Fig 8). With similar carrying capacities not only within but also across individuals, these core OTUs likely occupy similar metabolic niches across humans.
We employ an analytical approach based on continuous, multivariate, linear, autoregressive modeling tools developed for econometrics to pull apart two independent dynamic regimes in the human gut microbiome (Fig 3). Each regime, autoregressive and non-autoregressive, tells a unique story about the gut ecosystem.
OTU abundance trajectories showed fairly strong autocorrelation structure (Figs 5 and S2), although there was no evidence for auto-covariance in a limited set of dietary metadata from the DB time series (S3 Fig). The autocorrelation decay curves showed that most of the autocorrelation was gone after a lag of 3 or 4 days in the abundance data, and most of the autocorrelation was gone from the differenced time series after 1 or 2 days (S2 Fig). Thus, we chose to fit a lag-3 sparse vector autoregressive model to all the data to account for this autoregressive signal. sVAR(3) models produced residuals with reduced autocorrelation structure (S2 Fig). The sVAR models (i.e. the linear autoregressive components of the variance) accounted for a minority of the total community variance (0–50% for any given OTU; Fig 9). The set of OTUs with strong autoregressive signals were phylogenetically heterogeneous (S4 Fig). OTUs from the Enterobacteriaceae family tended to have larger amounts of their variance explained by the sVAR than other taxa (Fig 9). These Enterobacteriaceae OTUs were also often identified as CRTs, blooming periodically from low to high abundance. The most abundant taxa also tended to show more autoregressive structure than lower-abundance taxa (Fig 9).
Many sVAR coefficients showed significant Granger causal associations (i.e. the past abundances of one OTU predict the future abundances of another OTU; Fig 10; Chi-Squared test, p < 0.05). The M3 time series had the sparsest Granger network, with no significant relationships for 3-day lags. This lack of significant Granger relationships at 3-day lags may be due to the higher frequency of CRT blooms in that time series, which may have continually disrupted community recovery and obscured successional trends. A B. fragilis OTU had the largest number of connections in the M3 Granger network. B. fragilis, B. ovatus, and Enterobacteraceae OTUs had the largest numbers of connections in the F4 Granger network. In the DA and DB Granger networks, F. prausnitzii OTUs had the largest number of significant Granger interactions (Fig 10). Overall, Bacteroides, Faecalibacterium, and Enterobacteriaceae OTUs were prevalent in all the Granger networks (Fig 10). In each Granger network, there were several Granger-causal OTUs that influence multiple downstream responder OTUs, but there were only a few responder OTUs (e.g. B. fragilis, in the F4 time series) that integrate multiple upstream Granger signals. This pattern is consistent with cascading dynamics that result from perturbing a highly connected/central node in a network with an external shock [53]. In general, facultative aerobes (e.g. Enterobacteraceae OTUs) and aerotolerant taxa (e.g. Bacteroides) were more likely to Granger-cause obligate anaerobes (e.g. most Firmicutes OTUs) than the reverse (Fig 10). These intrinsic dynamics suggest a successional process that might follow a spike in luminal oxygen levels [52].
sVAR(3) residuals showed reduced autocovariance and could thus be more appropriately analyzed using standard statistical methods that assume independence. Closely related gut bacteria were generally positively correlated with one another, but this coherence decayed rapidly with phylogenetic distance (Fig 11). There was a significant anti-correlation between OTU-OTU phylogenetic distance and OTU-OTU correlations in abundances and abundance rates through time (i.e. highly related taxa tended to be positively correlated; Spearman’s p < 0.001). This phylogenetic coherence is conserved in the sVAR(3) residuals, but is completely absent in the sVAR(3) coefficients (Fig 11). Moreover, although we did not identify strong correlations between the set of dietary variables measured for the DB time series and the DB gut community, these dietary variables showed little-to-no autocorrelation structure (S3 Fig; and see Additional File A8 in David et al., 2014) [10]. Furthermore, we fit an sVAR(3) model to the original 97% OTUs from the DA time series paper and found that none of the OTUs that had previously been shown to correlate with dietary variables had any autoregressive signal (i.e. sVAR coefficients = 0; S1 File). We suggest that unmeasured dietary variables and host behavior/physiology are the non-autoregressive drivers responsible for the pronounced phylogenetic signal in OTU-OTU residual correlation structure (i.e. related taxa are positively correlated because they share a similar host/dietary niche, which fluctuates stochastically in time). In addition to phylogenetic coherence in the correlation structure within a gut time series, we find the same phylogenetic coherence in the correlations between OTU abundances across people from the HMP gut data set (S5 Fig). Thus, microbial phylogeny is strongly coupled to host niche in the human gut. In order to assess the generality of this relationship between phylogeny and correlation, we analyzed two time series from the English Channel and from Lake Mendota (marine and freshwater environments, respectively). We found the same pattern in these non-host associated environments, suggesting that niche filtering is a strong driver of dynamics across different ecosystems (S6 and S7 Figs).
Despite distinct differences in community composition across humans, the dynamics of the gut microbiota are stable and highly conserved. Dominant microbes in the gut appear to have fixed carrying capacities (i.e. their dynamics are stationary), which opens the door to many classical time series modeling approaches. Significant time-lagged interactions between OTUs often include opportunistic, facultative anaerobic organisms like Enterobacteraceae, and obligate anaerobes like F. prausnitzii [54, 55]. These autoregressive interactions appear to be due to succession and recovery of the gut community from CRT blooms, which may result from a disruption in the luminal redox balance [52]. The largest component of community variance is non-autoregressive and appears to be driven by non-autoregressive environmental forces, like pH [56] or fiber intake [10]. Unlike the autoregressive dynamics, these non-autoregressive dynamics carry a strong phylogenetic signal, indicative of niche filtering. Our results, based on a limited number of individuals, paint a coherent picture of the gut ecosystem and the major forces underlying its structure and stability, with two distinct dynamic regimes: one driven by external factors (e.g. diet) and the other by internal autoregressive processes (e.g. recovery following a disturbance). Moving forward, it will be important to collect more time series data, from both healthy and diseased individuals, to determine how general these dynamics are and whether or not they are observed in dysbiotic gut communities.
|
10.1371/journal.ppat.0030187 | Edema Toxin Impairs Anthracidal Phospholipase A2 Expression by Alveolar Macrophages | Bacillus anthracis, the etiological agent of anthrax, is a spore-forming Gram-positive bacterium. Infection with this pathogen results in multisystem dysfunction and death. The pathogenicity of B. anthracis is due to the production of virulence factors, including edema toxin (ET). Recently, we established the protective role of type-IIA secreted phospholipase A2 (sPLA2-IIA) against B. anthracis. A component of innate immunity produced by alveolar macrophages (AMs), sPLA2-IIA is found in human and animal bronchoalveolar lavages at sufficient levels to kill B. anthracis. However, pulmonary anthrax is almost always fatal, suggesting the potential impairment of sPLA2-IIA synthesis and/or action by B. anthracis factors. We investigated the effect of purified ET and ET-deficient B. anthracis strains on sPLA2-IIA expression in primary guinea pig AMs. We report that ET inhibits sPLA2-IIA expression in AMs at the transcriptional level via a cAMP/protein kinase A–dependent process. Moreover, we show that live B. anthracis strains expressing functional ET inhibit sPLA2-IIA expression, whereas ET-deficient strains induced this expression. This stimulatory effect, mediated partly by the cell wall peptidoglycan, can be counterbalanced by ET. We conclude that B. anthracis down-regulates sPLA2-IIA expression in AMs through a process involving ET. Our study, therefore, describes a new molecular mechanism implemented by B. anthracis to escape innate host defense. These pioneering data will provide new molecular targets for future intervention against this deathly pathogen.
| All mammals are susceptible to infection by Bacillus anthracis, the etiological agent of anthrax. Infection can occur either accidentally or as a potential consequence of a terrorism threat. Pulmonary infection is the most life-threatening form of the disease, causing a near 100% mortality. Despite appropriate therapy, all forms of infection may progress to fatal systemic anthrax, characterized by sepsis and respiratory failure. Thus, it is important to understand the mechanisms of host defense against B. anthracis. We have previously shown that alveolar macrophages produce an enzyme involved in innate defense that can kill B. anthracis: the enzyme is known as secreted phospholipase A2-IIA (sPLA2-IIA). The alveolar macrophage is one of the first cell types to come in contact with B. anthracis. In this study, we show that live B. anthracis spores stimulate the synthesis of sPLA2-IIA, this stimulation being counterbalanced by the inhibitory effect of the edema toxin produced by germinated spores and bacilli. Our study suggests that inhibition of sPLA2-IIA synthesis by edema toxin is a mechanism by which B. anthracis can escape innate host defense. These pioneering data provide new molecular targets for future intervention against this deadly pathogen.
| Bacillus anthracis, the etiological agent of anthrax, is a spore-forming Gram-positive bacterium [1]. Even though anthrax is primarily a disease of herbivores, all mammals are susceptible to B. anthracis infection. Human infection can occur via cutaneous, gastrointestinal, or respiratory routes, either accidentally or intentionally as a potential consequence of a bioweapon or a terrorism threat. Whatever the infection route used by this bacterium, spores are taken up by macrophages and/or dendritic cells, and subsequently migrate and germinate in the draining lymph nodes [2,3]. The infection then spreads through the whole organism, leading to respiratory failure and multiple hemorrhagic lesions. Despite appropriate therapy, all these forms of infection may progress to fatal systemic anthrax, which is characterized by shock-like symptoms, sepsis, and respiratory failure [4]. Pulmonary infection by B. anthracis has been shown to be the most life-threatening form of the disease, causing a near 100% mortality.
Innate immune response is the first line of host defense against invading pathogens. Type-IIA secreted phospholipase A2 (sPLA2-IIA) [5,6] is one of the major components involved in innate host defense against bacteria [7,8]. This enzyme belongs to a family of enzymes catalyzing the hydrolysis of phospholipids at the sn-2 position, leading to the generation of lysophospholipids and free fatty acids [5,6]. sPLA2-IIA is produced by several cell types, including guinea pig alveolar macrophages (AMs) [9], which play a central role in innate immunity and are the first line of defense against inhaled pathogens. These cells are the major pulmonary source of sPLA2-IIA in experimental models of acute lung injury [9,10]. Besides its ability to hydrolyze pulmonary surfactant phospholipids [11] and release arachidonic acid [12], sPLA2-IIA exhibits potent bactericidal activity, especially against Gram-positive bacteria [13–15]. The bactericidal activity is exhibited through a process involving rapid hydrolysis of bacterial membrane phospholipids [16,17]. This activity is the most significant biological property of sPLA2-IIA, being observed at much lower concentrations of this enzyme than for other properties. sPLA2-IIA is constitutively present in guinea pig airways at concentrations [11] above those required for killing B. anthracis [17]. We have also shown that isolated guinea pig AMs constitutively release enough sPLA2-IIA to kill B. anthracis [17]. sPLA2-IIA is highly bactericidal for B. anthracis, either in vitro or in vivo [17,18]. This anthracidal effect occurs both against germinated spores and bacilli. A sPLA2-IIA–dependent anthracidal activity was found in human bronchoalveolar lavage fluids (BALF) of patients with acute respiratory distress syndrome [17]. In a more recent study, we showed that transgenic mice expressing sPLA2-IIA are resistant to experimental infection with B. anthracis, in contrast to sPLA2-IIA−/− mice [18]. Interestingly, treating sPLA2-IIA−/− mice with recombinant sPLA2-IIA protected them against mortality caused by B. anthracis infection.
The exact mechanisms by which B. anthracis induces anthrax are not fully understood; however, it is clearly established that this bacterium spreads rapidly in the host at the early stages of infection without a detectable immune response [2,19,20]. This allows bacteria to replicate to very high numbers in the blood, ultimately leading to death of the host [4]. This is due to the ability of B. anthracis to subvert the host immune response [19,20] by the action of B. anthracis toxins. Indeed, B. anthracis secretes a binary A-B toxin composed of a single B transporter called protective antigen (PA) and two alternative A components, lethal factor (LF) or edema factor (EF) [1,19]. LF and EF act in pairs, with PA leading to lethal toxin (LT = PA + LF) and edema toxin (ET = PA + EF), respectively. PA serves as a transporter delivering LF and EF inside host cell cytosol where they act on specific molecular targets. EF is a calmodulin-activated adenyl-cyclase (AC) leading to a sustained increase in cAMP levels [21]. However, little is known about the genes targeted by ET and their implication in the pathogenesis of anthrax.
Here, we report that ET inhibits sPLA2-IIA expression in AMs and demonstrate that this inhibition occurs through the sequential accumulation of cAMP and stimulation of protein kinase A (PKA) activity. In the absence of ET, B. anthracis was able to induce sPLA2-IIA expression via a process at least partly involving the cell wall component, peptidoglycan (PG). ET also stopped this induction. This study is important because AMs are a major source of sPLA2-IIA, a critical component in host defense against B. anthracis. Inhibition of sPLA2-IIA expression in AMs by ET may represent an effective strategy for subverting pulmonary host immune response by B. anthracis.
AMs were preincubated with ET (PA + EF) 1 h before adding lipopolysaccharide (LPS) to analyze the effect of ET on sPLA2-IIA expression. ET stopped both basal and LPS-induced sPLA2-IIA secretion in a concentration-dependent manner (Figure 1A). No effect was observed when EF or PA was added separately to AMs (unpublished data). Inhibition of sPLA2-IIA secretion by ET was also observed when AMs were stimulated by tumor necrosis factor-α (TNFα) instead of LPS (Figure 1B). We showed that LPS induced a marked increase in sPLA2-IIA mRNA levels, and that the increase was subsequently abolished by the addition of ET (Figure 1C).
We next investigated the effect of ET on two other inflammatory mediators produced by AMs, interleukin 8 (IL-8) and prostaglandin E2 (PGE2). ET failed to interfere with LPS-induced IL-8 (Figure 1D) and PGE2 (Figure 1E) secretion. We also examined the effect of ET on nuclear factor κ B (NF-κB) translocation. ET had no effect on LPS-induced NF-κB translocation (Figure 1F), as assessed by electrophoretic mobility shift assay (EMSA). These results together indicated that ET inhibits sPLA2-IIA expression in AMs through a different signaling pathway from those inducing IL-8 and PGE2 secretion or NF-κB translocation.
Because ET exhibits a calmodulin-dependant AC, we examined its effect on intracellular cAMP levels in our cell system. A 30-min incubation of AMs with ET led to an increase in cAMP levels, whereas LPS had no effect (Figure 2A). The induction of cAMP accumulation by ET was transient; cAMP levels returned to near basal levels after AMs were incubated with ET for 24 h (Figure 2A, insert). In agreement, a cAMP-elevating agent, forskolin, significantly inhibited LPS-induced sPLA2-IIA secretion (Figure 2B). AC inhibitors (adefovir and ddA) reversed ET inhibition of LPS-induced sPLA2-IIA secretion (Figure 2C).
cAMP is known to activate protein kinases, such as PKA; thus, this kinase may be involved in the inhibition of sPLA2-IIA expression by ET. ET induced a marked and transient activation of PKA in AMs (Figure 2D). Indeed, this activation was observed 2 h after adding ET, and was undetectable 20 h later (unpublished data). To mimic the ET-induced PKA activation, we examined the effect of 6-Bnz-AMP, a specific agonist for PKA. 6-Bnz-AMP inhibited both basal and LPS-induced sPLA2-IIA expression (Figure 2E). By contrast, O-Me-cAMP, a specific agonist for the exchange protein directly activated by cAMP (Epac) [22], had no effect on sPLA2-IIA expression. Taken together, our results suggested that ET inhibits LPS-induced sPLA2-IIA expression in AMs via a cAMP/PKA-dependent process.
Because PKA is known to phosphorylate cAMP-responsive element binding protein (CREB), we examined whether this transcription factor mediates the inhibition of sPLA2-IIA expression by ET. ET induced a time-dependent CREB phosphorylation (Figure 3A), but had no effect on the total level of CREB (Figure 3B) in AMs, as assessed by western blot analysis. We also investigated the effects of ET on CREB activation using Chinese hamster ovary (CHO) cells transfected with a CREB ([CRE]4-Luc) reporter plasmid construct. ET significantly increased the CREB luciferase activity (Figure 3C). However, LPS had no effect on this activity and failed to interfere with ET-induced CREB activation. This activation was prevented by cotransfecting cells with a dominant-negative CREB construct, pGR-CREBM1, as opposed to pGR (Figure 3C). Transfection of CHO cells with a sPLA2-IIA promoter luciferase construct demonstrated that ET inhibits LPS-induced sPLA2-IIA gene transcription activity (Figure 3D). Similar results were observed when LPS was replaced by IL-1β as the inducer of sPLA2-IIA expression (unpublished data). However, cotransfection of a dominant-negative CREB construct failed to reverse the inhibition of sPLA2-IIA gene transcription activity (Figure 3D), indicating that CREB does not mediate ET inhibition of sPLA2-IIA expression.
Using a more pathophysiological approach, we examined whether infecting AMs with B. anthracis bacilli modulates sPLA2-IIA expression and whether ET participates in this modulation. AMs were incubated in an antibiotic-free culture medium for 3 h with either RP10 or RPLC2 bacilli; RP10 produces functional and RPLC2 produces inactive ET. After removing bacilli not having undergone phagocytosis, AMs were stimulated overnight with LPS in culture medium supplemented with antibiotics. The RP10 strain inhibited LPS-stimulated sPLA2-IIA secretion, whereas the RPLC2 strain had no effect (Figure 4A). Inhibition by the RP10 strain occurred in a multiplicity of infection (MOI)-dependent manner and was selective for sPLA2-IIA. Indeed, this strain failed to inhibit LPS-induced PGE2 (Figure 4B) and IL-8 (Figure 4C) production. These findings demonstrated that in LPS-stimulated AMs, B. anthracis strains producing functionally active ET down-regulated sPLA2-IIA expression.
We next examined the effect of B. anthracis on sPLA2-IIA expression in unstimulated AMs. RPLC2 bacilli induced sPLA2-IIA expression (Figure 4D), and PGE2 (Figure 4E) and IL-8 (Figure 4F) secretion. The RP10 bacilli strain induced PGE2 and IL-8 secretion, but had no effect on sPLA2-IIA expression (Figure 4D–4F). Interestingly, RP10 and RPLC2 spores induced sPLA2-IIA expression, even after 3 h of infection (Figure 4G). These findings indicate that in the sporular state, RP10 and RPCL2 strains induce sPLA2-IIA expression. However, in the bacilli state, the RPCL2 strain (devoid of ET) induced sPLA2-IIA expression, whereas the RP10 strain (producing ET) exerted an inhibitory effect. AMs were incubated with cytochalasin D (Cyto D) before adding the RP10 strain, to examine the impact of B. anthracis phagocytosis on sPLA2-IIA expression. Cyto D reduced the inhibitory effect of bacilli on sPLA2-IIA expression, but failed to interfere with the stimulatory effect of spores (Figure 4H and 4I). This suggests that both extracellular and intracellular bacilli are involved in inhibiting sPLA2-IIA expression, whereas extracellular spores seem to play a more important role in inducing this enzyme.
As RPLC2 strain induces sPLA2-IIA expression in AMs, we searched for which B. anthracis component was involved in this induction. PG purified from B. anthracis stimulated sPLA2-IIA expression (Figure 5A). PG, as well as LPS, induced NF-κB translocation, as assessed by EMSA (Figure 5B). PG-induced sPLA2-IIA expression was abolished if AMs were pretreated with the NF-κB inhibitor CAPE (Figure 5C). Interestingly, pretreating AMs with ET stopped sPLA2-IIA expression induced by B. anthracis PG (Figure 5D). PG-induced sPLA2-IIA expression was also inhibited by the cAMP-elevating agent, forskolin, and the PKA agonist, 6-Bnz-cAMP (Figure 5E).
In this study, we investigated the effect of B. anthracis, the causative agent of anthrax [1–4], on the expression of sPLA2-IIA, an important component of host defense against invading bacteria. This enzyme is bactericidal in vitro or in vivo, and is especially active against Gram-positive bacteria, including B. anthracis [7,8,17,18]. sPLA2-IIA is produced by AMs and found in human and animal BALF at sufficient levels to kill B. anthracis [17]; these findings are consistent with the enzyme having a role in host defense against pulmonary anthrax. However, despite the ability of lungs to produce sPLA2-IIA, the pulmonary form of anthrax has been shown to be fatal, causing almost 100% mortality [1–4]. This led us to postulate that B. anthracis may inhibit sPLA2-IIA synthesis by AMs, subvert host pulmonary defense, and allow this pathogen to spread extensively in the host.
We show here that ET inhibits sPLA2-IIA secretion by AMs, interfering with its expression at the transcriptional level. Inhibiting sPLA2-IIA secretion may decrease the capacity of AMs to kill B. anthracis bacilli and germinated spores. Indeed, AM activity against B. anthracis has been shown to be at least partly associated with sPLA2-IIA, as it was reduced by an sPLA2-IIA inhibitor [17]. This inhibition was observed whatever the stimuli used (LPS, TNFα , IL1β, or PG). We analyzed the signaling pathways by which ET down-regulates sPLA2-IIA expression; our analysis suggested that this inhibition occurs via a process involving cAMP accumulation. Our studies showed that this accumulation was transient, reaching near basal values within 24 h. This contrasts with previous studies reporting that cAMP accumulation was elevated for 48 h or more after ET incubation with NIH/3T3 fibroblasts and RAW 267 macrophages [23]. Thus, it is likely that the duration and amplitude of cAMP accumulation induced by ET may vary with the cell type considered. Because cAMP activates several kinases, we examined whether PKA and Epac, two cAMP-dependent kinases, were involved in this process. PKA but not Epac, appeared to mediate ET-induced inhibition of sPLA2-IIA expression. Our results also suggested that elevating intracellular cAMP concentrations (either by ET or 6-Bnz-cAMP) interfered with basal and LPS-induced sPLA2-IIA expression by different mechanisms. The inhibition of induced expression appeared to occur through a process that interferes, at least partly, with the sPLA2-IIA promoter, whereas the inhibition of basal expression appeared to be independent of the sPLA2-IIA transcription.
PKA phosphorylates proteins, such as CREB, that are involved in regulating gene expression in mammalian cells [24]. This factor can modulate, either positively or negatively, gene expression in several cell-activation processes [24,25]. Although ET induces CREB activation, this transcription factor does not mediate the inhibition of sPLA2-IIA expression by ET. However, it is likely that CREB activation by ET could modulate the expression of other genes involved in host defense, which remain to be identified. We next investigated whether ET inhibits sPLA2-IIA expression by interfering with the activation of NF-κB, known to be critical in inducing sPLA2-IIA expression [26]. ET had no effect on stimulated NF-κB translocation in AMs. Also, ET had no effect on the secretion of IL-8, whose expression is controlled by NF-κB. However, we cannot exclude that ET may interfere with stimulating cofactors involved in NF-κB coactivation at the sPLA2-IIA promoter level. Studies in progress in our laboratory showed that trichostatin A, an inhibitor of histone deacetylase (HDAC) activity [27], significantly decreased sPLA2-IIA expression in LPS-stimulated AMs. Because HDAC activity is altered by a PKA-dependent phosphorylation [28], it is likely that HDAC may play a role in the inhibition of sPLA2-IIA expression by ET. Further studies are required to verify this hypothesis.
In a more physiological approach, we investigated whether ET modulates sPLA2-IIA expression during infection of AMs with live B. anthracis. This bacterium inhibits LPS-induced sPLA2-IIA expression via ET. Indeed, RPLC2, the bacterial mutant with inactive ET, had no effect on this induction, whereas the RP10 strain expressing functional ET abolished LPS-induced sPLA2-IIA expression. Incubating RPLC2 bacilli, which produce inactive ET, with unstimulated AMs induced sPLA2-IIA expression. This suggested the existence of bacterial component(s) that are able to induce sPLA2-IIA synthesis, and that their actions are masked by the ET inhibitory effect produced by RP10 bacilli. Our findings showed that the cell wall PG purified from B. anthracis induces sPLA2-IIA expression via a process involving NF-κB activation. It is still not clear whether PG-induced sPLA2-IIA expression occurs via an activation of TLR2 or Nod, two PG recognition proteins [29]. A recent study has reported that Nod may be involved in cell activation by B. anthracis spores [30]. We cannot exclude, however, that other bacterial components present in the cell wall or released by B. anthracis may also be involved in inducing sPLA2-IIA expression. Interestingly, ET suppressed PG-induced sPLA2-IIA expression, confirming the relevance of our studies, and showing that ET also suppresses the sPLA2-IIA expression induced by B. anthracis itself. Therefore, during host infection, B. anthracis may modulate sPLA2-IIA expression, either positively or negatively, depending on the status of ET synthesis in the bacterium (Figure 6).
Mammalian pulmonary infection with B. anthracis is initiated by the inhalation of spores, the cell walls of which contain PG. Infecting spores therefore induce sPLA2-IIA expression in the earlier stages of infection. This is consistent with previous studies, which have reported that B. anthracis spores stimulate cytokine production in various cells [31–33]. The susceptibility of inhaled spores to the bactericidal activity of sPLA2-IIA present in airways is dependent on their germination velocity; this is because sPLA2-IIA only kills germinated spores and bacilli [17]. Previous in vivo studies [3] have shown that germination occurred rapidly upon entry into the lung (35–60 min), and that the spores were mostly found inside the AM. This was followed by a rapid onset (<3 h) of expression of genes encoding virulence factors, such as LF, PA, and EF [34]. Elimination of inhaled B. anthracis by the host would thus depend on the balance between sPLA2-IIA levels in the airways and bacterial load. If the balance favors sPLA2-IIA, germinated spores and bacilli would be killed quickly. Our previous studies have shown that the constitutive (basal) levels of sPLA2-IIA present in guinea pig airways [11] are greater than those required for killing B. anthracis [17], and that these levels were greater in inflamed lungs [11]. sPLA2-IIA was also found in BALF of patients with lung inflammatory diseases (ARDS) at sufficient levels to exert this anthracidal effect. However, it is still unknown whether BALF of healthy subjects contains functionally significant amounts of sPLA2-IIA, and whether this enzyme would be available soon after the invasion of inhaled spores. It is also possible that some germinated bacteria may escape killing by sPLA2-IIA. Therefore, bacilli derived from geminated spores rapidly produce ET, which may in turn inhibit sPLA2-IIA expression and shift the balance in favor of bacteria. In vivo studies, in which guinea pigs are infected with inhaled B. anthracis spores, are required to investigate this possibility.
Our previous studies clearly established a role for sPLA2-IIA in host defense against B. anthracis; however, it is still unknown whether this enzyme participates in killing germinated spores previously ingested by AMs. sPLA2-IIA is involved in the killing of Staphylococcus aureus ingested by neutrophils [35]. Indeed, added sPLA2-IIA binds to S. aureus before its internalization by neutrophils and participates in the cytotoxicity of these cells towards ingested bacteria. Therefore, we suggest that sPLA2-IIA is involved in killing B. anthracis ingested by AMs.
In conclusion, we report here that B. anthracis represses the expression of sPLA2-IIA, a major component in innate host response with anthracidal properties, in AMs. This inhibition occurs through a process involving ET-mediated cAMP accumulation and PKA activation, and represents a novel mechanism for evading the innate immune response of the host. Other bacteria (for example Bordetella pertussis or Yersinia pestis) [19] are known to produce toxins with AC activity; thus, we can speculate that the inhibition of sPLA2-IIA expression in AMs may be a more general process occurring during bacterial infection. Therefore, using pharmacological approaches to inhibit ACs of invading bacteria may represent a therapeutic strategy for treating not only pulmonary anthrax, but also other bacterial pulmonary infections.
Male Hartley guinea pigs were purchased from Charles River Laboratories. RPMI 1640 cell culture medium was purchased from Invitrogen, and fetal calf serum (FCS) from Hyclone. Caffeic acid phenethyl ester (CAPE), and cytochalasin D (Cyto D) were purchased from Biomol. LPS from Pseudomonas aeruginosa and 2′, 5′-dideoxyadenosine 3′-triphosphate (ddA) were purchased from Sigma Aldrich. N6-Benzoyladenosine-3′, 5′-cyclic monophosphate (6-Bnz-cAMP) and 8-(p-Chlorophenylthio)-2′-O-methyl-adenosine-3′, 5′-cyclic monophosphate (O-Me-cAMP) were purchased from Biolog. CREB and phospho-CREB antibodies were obtained from Cell Signaling Technology. EF, PA, and PG from B. anthracis were produced and purified as described previously [36]. BIS-POM-PMEA (adefovir) was provided by Dr. W. J. Tang (University of Chicago, Chicago, Illinois).
The following isogenic B. anthracis strains were studied: (1) the single mutant RP10 Δlef producing only PA-EF and (2) the double-mutant RPLC2 on lef and cya genes producing PA-LF and PA-EF, respectively, without enzymatic functions [37].
Guinea pig bronchoalveolar lavages (BAL) were performed with PBS, and AMs were isolated, as previously described [9]. AMs were then adjusted at 2.106 cells/ml in RPMI 1640 with 3% FCS and 1% of antibiotic, and were pretreated with ET (PA + EF), 6-Bnz-cAMP, O-Me-cAMP, or TSA 1 h before incubation with LPS, PG, or TNFα. In certain experiments, AMs were pretreated 5 h with adefovir before incubation with ET. In other experiments, ET was preincubated with ddA for 1 h before being added to AMs. These reagents were used at the concentrations indicated in the figures. Subsequent analyzes were performed as detailed below.
Cells were infected with B. anthracis bacilli or spores for 3 h at various MOI values. Cells were then washed twice and incubated overnight in RPMI 1640 supplemented with 3% FCS and 2.5 μg/ml gentamicin in the presence or absence of LPS. In certain experiments, AMs were pretreated with Cyto D for 30 min before adding bacteria. At the end of the incubation, media were harvested and centrifuged. The resulting supernatants were collected and stored at −20 °C for subsequent analyzes.
Cells were grown on a cell culture plate and total RNA was extracted using an RNeasy kit (Qiagen). DNase treatment was performed using 2 μg of extracted RNA, 1 μl of DNase I (Amersham Biosciences), and 0.5 μl of RNasin (Promega) in a total volume of 20 μl in the manufacturer's buffer. cDNA were obtained by incubating RNA with 1 mM dNTP (Eurobio), 1.5 μl of hexamers as primers, 20 units of RNasin (Promega), and 300 units of Moloney murine leukemia virus reverse transcriptase RNase H minus (Promega) in a total volume of 50 μl of the manufacturer's buffer; the incubation was for 1 h at 42 °C and was followed by a 10-min incubation at 70 °C. PCR was performed using specific primers (Proligo) for guinea pig sPLA2-IIA (sense, 5′-ACA AGT TAT GGC GCC TAT GG-3′; antisense, 5′-GCC CAG TGT AGC TGT GAA GC-3′). As an internal control, we used primers for the detection of guinea pig β-actin (sense, 5′-AAA CTG GAA CGG TGA AGG TG-3′; antisense, 5′-TCA AGT TGG GGG ACA AAA AG-3′). Amplifications were performed in a Peltier thermal cycler (MJ Research) using Q-BioTaq polymerase (Qbiogene). For the detection of sPLA2-IIA, PCR thermo-cycling included 30 cycles of denaturation at 95 °C for 45 s and annealing at 60 °C for 45 s.
Nuclear proteins were extracted from 2.106 AMs, as previously described [38]. The NF-κB double-stranded oligonucleotides (Santa Cruz Biotechnology) corresponded to an NF-κB binding site consensus sequence of 5′-AGT TGA GGG GAC TTTT CCC AGG C-3′. The overhanging ends were γ-32P–labeled with T4 polynucleotide kinase (Biolabs). Protein concentrations were determined using a Nanodrop spectrophotometer (Nyxor Biotech). Binding reactions were performed in a total volume of 20 μl for 20 min at room temperature, by adding 5 μg of nuclear extract, 10 μl of 2× binding buffer (40 mM HEPES [pH 7], 140 mM KCl, 4 mM DTT, 0.02% Nonidet P-40, 8% Ficoll, 200 μg/ml BSA, 1 μg of poly(dI:dC)), and 1 μl of labeled probe. The reaction mixtures were separated on a 5% polyacrylamide gel in 0.5% Tris/borate/EDTA buffer at 150 V for 2 h. Gels were dried and exposed for 2 to 12 h. We have previously shown, using supershift analysis, that antibodies directed against NF-κB's p50 and p65 subunits displaced the NF-κB band in LPS-stimulated AMs; this confirmed that the observed complexes belong to the NF-κB family [39].
Proteins from AMs were extracted in lysis buffer (10 mM Tris-HCl, 10 mM NaCl, 3 mM EDTA, 100 μM leupeptin, 100 mM aprotinin, 1 mM soybean trypsin inhibitor, 5 mM NEM, 1 mM PMSF, 5 mM benzamidine, and 1% Triton W-100 [pH 7.4]) and were run on a gel under reducing conditions. Semidry transferred proteins were applied to polyvinylidene difluoride membranes. Nonspecific binding sites were blocked overnight with 5% BSA in 20 mM Tris-HCl (pH 7.6), 140 mM NaCl, and 0.1% Tween 20. Blots were probed for 1 h with rabbit polyclonal anti-human phospho-CREB (ser 133) or CREB antibodies (1/2,000 dilution). These antibodies also recognize activating transcription factor-1 (ATF-1), which belongs to the CREB family. After washing, the immunoreactive bands were visualized using a peroxidase-conjugated goat anti-rabbit immunoglobulin G (IgG) (1/10,000 dilution) antibody and an ECL Plus Western Blotting Detecting System (Amersham Biosciences). Quantifications were carried out using the Image J software and were expressed as arbitrary units.
cAMP concentrations were measured in disrupted cells using a specific enzyme immunoassay kit purchased from Cayman Chemical Co; the concentrations were measured after incubating AMs with LPS and/or ET for 30 min or 24 h. Protein concentrations were measured in cell lysates using a kit from Pearce, and then the concentrations of cAMP were expressed in picomoles per milligram of protein. IL-8 and PGE2 concentrations were measured in culture medium after 24 h incubation of AMs with LPS and/or ET; these concentrations were measured using a specific PGE2 enzyme immunoassay (Cayman Chemical Co) and human IL-8 Kit DuoSet ELISA (R&D Systems), which cross-reacts with guinea pig IL-8 [40].
AMs were incubated with LPS and/or ET for 2 h. The PKA activity was then measured in disrupted cells using a specific enzyme immunoassay kit purchased from Promega. sPLA2-IIA activity was measured in culture medium using [3H]-oleic acid-labeled membranes of Escherichia coli, following a modification [41] of the method by Franson et al. [42].
Mutated constructs [−488; +46]-sPLA2-Luc were prepared, as described previously [43]. CHO cells were seeded on dishes cultured in HAM F12, supplemented with 10% (v/v) FCS (Gibco BRL), 4 mM glutamine, 100 U/ml penicillin, and 100 mg/ml streptomycin. CHO cells were seeded in 24-well plates at a concentration of 2.104 cells per plate at 70% confluence, 24 h before transfection. Transfections with mutated constructs [−488; +46]-sPLA2-Luc, CREB, and DN CREB were performed using 0.75 ml of LIPOFECTAMINE Plus (Invitrogen), 0.4 mg of reporter DNA, as indicated in Figure 4, and 0.1 mg of pCMV-β-galactosidase per well. The cells were incubated with HAM-F12 medium 3 h after adding the DNA, and incubation was continued for 24 h. CHO cells were incubated with PA (1 μg/ml) and EF (500 ng/ml) for 1 h. LPS (1 μg/ml) was then added, and incubation was continued for an additional 24 h. Luciferase activity was measured using a luciferase reporter assay kit, with signal detection for 12 s by a luminometer (Berthold), and was normalized by dividing the relative light units by β-galactosidase activity [43]. The degree of induction was calculated relative to the control.
Cell viability was checked by the trypan blue dye exclusion test. Cell lysis was controlled by measuring the release of lactate dehydrogenase (LDH) activity using a commercial kit from Boehringer. No cell mortality was observed in all the experiments presented in this study.
Data are expressed as the mean ± standard error of the mean (S.E.M.) of at least three separate experiments, and statistical analyzes were performed using the unpaired Student t-test. |
10.1371/journal.pcbi.1003226 | Target Inhibition Networks: Predicting Selective Combinations of Druggable Targets to Block Cancer Survival Pathways | A recent trend in drug development is to identify drug combinations or multi-target agents that effectively modify multiple nodes of disease-associated networks. Such polypharmacological effects may reduce the risk of emerging drug resistance by means of attacking the disease networks through synergistic and synthetic lethal interactions. However, due to the exponentially increasing number of potential drug and target combinations, systematic approaches are needed for prioritizing the most potent multi-target alternatives on a global network level. We took a functional systems pharmacology approach toward the identification of selective target combinations for specific cancer cells by combining large-scale screening data on drug treatment efficacies and drug-target binding affinities. Our model-based prediction approach, named TIMMA, takes advantage of the polypharmacological effects of drugs and infers combinatorial drug efficacies through system-level target inhibition networks. Case studies in MCF-7 and MDA-MB-231 breast cancer and BxPC-3 pancreatic cancer cells demonstrated how the target inhibition modeling allows systematic exploration of functional interactions between drugs and their targets to maximally inhibit multiple survival pathways in a given cancer type. The TIMMA prediction results were experimentally validated by means of systematic siRNA-mediated silencing of the selected targets and their pairwise combinations, showing increased ability to identify not only such druggable kinase targets that are essential for cancer survival either individually or in combination, but also synergistic interactions indicative of non-additive drug efficacies. These system-level analyses were enabled by a novel model construction method utilizing maximization and minimization rules, as well as a model selection algorithm based on sequential forward floating search. Compared with an existing computational solution, TIMMA showed both enhanced prediction accuracies in cross validation as well as significant reduction in computation times. Such cost-effective computational-experimental design strategies have the potential to greatly speed-up the drug testing efforts by prioritizing those interventions and interactions warranting further study in individual cancer cases.
| Selective inhibition of specific panels of multiple protein targets provides an unprecedented potential for improving therapeutic efficacy of anticancer agents. We introduce a computational systems pharmacology strategy, which uses the concept of target inhibition networks to predict effective multi-target combinations for treating specific cancer types. The strategy is based on integration of two complementary information sources, drug treatment efficacies and drug-target binding affinities, which are readily available in drug screening labs. Compared to the cancer sequencing efforts, which often result in a huge number of non-targetable genetic alterations, the target combinations from our strategy are druggable, by definition, hence enabling more straightforward translation toward clinically actionable treatment strategies. The model predictions were experimentally validated using siRNA-mediated target silencing screens in three case studies involving MDA-MB-231 and MCF-7 breast cancer and BxPC-3 pancreatic cancer cells. In more general terms, the cancer cell-specific target inhibition networks provided additional insights into the drugs' mechanisms of action, for instance, how the cancer cell survival pathways can be targeted by synergistic and synthetic lethal interactions through multi–target perturbations. These results demonstrate that the principles introduced here offer the possibilities to move toward more systematic prediction and evaluation of the most effective drug target combinations.
| Over the past decade, there has been a significant increase in the R&D cost when translating new cancer drug candidates into effective therapies in the clinic. The two single most important reasons are (i) lack of efficacy and (ii) clinical safety of the candidate drug compounds [1]. This decline in productivity of the pharmaceutical industry has greatly challenged the dominant paradigm in drug discovery, where such ‘magic bullet’ compounds are being searched that could produce dramatic treatment outcomes at a population-level by modulating one specific target only. The shortcomings of such ‘one-size-fits-all’ treatment strategies are well reflected in the disappointing outcome of the current anticancer drug development, where agents directed at an individual target often show limited efficacy and safety due to factors such as off-target activities, network robustness, bypass mechanisms and cross-talk across compensatory escape pathways [2]–[4]. Most cancers develop from specific combinations of genetic alterations accumulated in tumor cells, which are often distinct between different cancer types and result in different treatment responses to the same therapy. Moreover, the extensive mutational heterogeneity results in alterations within multiple molecular pathways, making most advanced tumors readily resistant to single-targeted agents. Therefore, rational strategies to develop multi-targeted therapies for specific cancer types are needed to attack the resistance problem and to provide more effective and personalized treatment strategies [5]. Targeted drug combinations may also overcome the side effects associated with high doses of single drugs by countering pathway compensation and thereby increasing cancer cell killing while minimizing overlapping toxicity and allowing reduced dosage of each compound [6].
Even though it is widely acknowledged that effective cancer treatments need to go beyond the traditional ‘one disease, one drug, one target’ paradigm, the major bottleneck hindering the development of combinatorial therapies is the lack of such systematic experimental-computational approaches that could pinpoint the most effective combinations [7]–[9]. While efforts based on next-generation sequencing are very successful at systematically characterizing the structural basis of cancers, by identifying the genomic mutations associated with each cancer type [10], these findings often do not lead to clinically actionable therapeutic strategies and rarely to rational targeted combinations. The large number of genetic alterations present in tumor cells makes the discrimination of the cancer-specific driver mutations and pathways highly challenging, and even when genetic aberrations with pathogenetic importance can be identified, these may not be pharmaceutically actionable. Moreover, genes not altered at the genomic level may also play essential roles in the cancer progression, hence providing additional therapeutic opportunities [11]. In contrast, systematic assessment of genes for their contribution to tumor addictions can provide functional insight into the molecular mechanisms and pathways behind specific cancer types, hence highlighting their vulnerabilities associated with driver genes, synthetic lethal interactions and other tumor dependencies [12]–[14], which are complementary to the structural information obtained from the cancer mutational landscape. Advances in high-throughput chemical and RNAi screening have now made it possible to carry out comprehensive functional screening in cancer cells, providing novel targets for the next generation of anticancer therapies for patients sharing a common genetic background [15]–[18].
However, despite the emerging possibilities for perturbing gene functions with a wide spectrum of shRNA/siRNA libraries or using diverse drug and compound collections, functional interactions between genes and/or drugs have remained extremely difficult to predict on a global scale [18]. The complex genotype-phenotype relationships behind such interactions pose modeling challenges beyond the reach of the classical linear approaches. Moreover, polypharmacologic compounds elicit their bioactivities by modulating multiple targets, which leads to a combinatorial explosion both in the pharmacological and molecular spaces. Taken together, the exponentially increasing number of possible RNAi, chemical, target and dose combinations poses great experimental challenges, and exhaustive experimentation with all the possible combinations is impossible in practice, making the pure experimental approach quickly unfeasible [19]. To meet these computational and experimental challenges, novel modeling frameworks and efficient computational algorithms are needed to effectively reduce the search space for determining the most promising combinations and prioritizing their experimental evaluation. Ideally, the experimental setup should be both economical and practical, utilizing such functional measurements and phenotypic readouts that are readily available in typical drug screening experiments. Moreover, the experimental and computational platforms should also be compatible with the eventual clinical translation in the sense that the measurements and their analysis can be made in each patient individually, and that the modeling and algorithmic predictions can be calculated in a reasonable time.
A number of computational algorithms have been developed for predicting drug combinations in silico [5], [9], [20]. Most of the approaches are based on detailed mathematical modeling, utilizing a priori knowledge extracted from databases, such as those focusing on established cancer pathways, metabolic network constructions or literature-curated models [21]–[23]. A limitation of such detailed models is that global kinetic information for many cancer-related systems are still rarely available, and reduced subsystem models are often biased toward what is already known about the cancer processes. For instance, pathway-specific models may miss important novel features, such as pathway cross-talks or novel cancer dependencies. Accordingly, although major canonical pathways involved in different cancer types are increasingly well established, individual pathway models cannot capture the complex and context-dependent cellular wiring patterns behind distinct cancer phenotypes [5]. There are also approaches that take the cell context into account by means of global gene expression or targeted phosphoproteomics profiling [24]–[27]. However, such molecular phenotypes are not routinely profiled in a typical high-throughput drug testing approaches, especially in clinical settings. Moreover, downstream changes in the expression patterns are shown to be suboptimal in distinguishing mechanism of action between different compounds [28], [29]. Perhaps more importantly, targets identified by means of genomic profiling may not be pharmaceutically actionable in clinical practice. For instance, many genes identified through expression profiling or genomic sequencing are either not druggable at all, or druggable, but not actionable, as there are no approved drugs available in the clinic.
In this article, we present an efficient model construction algorithm, named TIMMA (Target Inhibition inference using Maximization and Minimization Averaging), which makes the use of partly overlapping target subsets and supersets of promiscuous drug-target binding profiles in the estimation of anticancer efficacies for novel drug target combinations. The model construction and target combination predictions are based on functional data on drugs and their targets that are available from comprehensive target binding assays and from high-throughput drug sensitivity screens. We implemented a modified sequential forward floating search algorithm for model selection, which enables scaling-up to proteome-wide evaluation of the targets in terms of their relevance to cancer survival. Both simulation studies and an application to a canine osteosarcoma cell line data showed that TIMMA achieved improved prediction accuracy, when compared to a published algorithm [30], at significantly lower computational cost. Importantly, application case studies in MCF-7 and MDA-MB-231 breast cancer and BxPC-3 pancreatic cancer cells confirmed that TIMMA-predicted kinase targets are essential for tumor survival, either individually or in combination, as validated by independent single and pairwise target knockdowns with siRNA screening. Our model predictions, visualized as a target inhibition network, provide insights into such druggable cancer cell addictions, the inhibition of which can jointly block the survival pathways. With the increasing interest in drug combination screens, our modeling strategy can be readily used as an efficient prioritization procedure to pinpoint the most potential drug combinations based merely on their selectivity profiles and individual responses in given cancer samples.
Consider a set of drugs where the single-drug treatment efficacy on a given cancer sample is measured as a phenotypic response in a high-throughput drug screen. The drug's treatment efficacy to kill cancer cells is conventionally scored using response parameters, such as the drug concentration at which the cancer cell growth is inhibited by a certain percentage (e.g. half-maximal inhibitory concentration IC50). A drug with a smaller inhibitory concentration is usually considered as more potent. Drug treatment efficacy and potency can be also quantified based on the area under the dose-response curve, such as the activity area (AA) [31] or the drug sensitivity score (DSS), which provide summary information about the complex dose-response relationships. We denote the drug treatment efficacy data by a vector with length and scale it into the interval of [0, 1], with the minimum and maximum efficacies being 0 and 1, respectively.
To relate a drug's treatment efficacy with its underlying mechanism of action, the cellular targets of the drug need to be mapped into a drug-target inhibition profile. Let the potential target set be , where refers to the total number of targets that bind to at least one of the drugs. A target inhibition profile of a drug i can be binarized from drug-target binding affinities as a binary vector , where 0 and 1 is a result of classification of low and high binding affinities, respectively. The target inhibition profile for all the drugs is abbreviated as . An example of such binarized target annotations can be derived from quantitative binding assay measurements collected from the ChEMBL database [32], provided that knowledge of relevant binding affinity cutoffs is applicable.
Given the single drug efficacy and target inhibition profiles, our aim is to predict the treatment efficacy for novel drug combinations. We consider the target inhibition profile of a drug combination as a union of the target inhibition profiles of each component drug in the cocktail (Figure 1). However, not all the targets in the profile are essential in explaining the treatment efficacy. Ideally, an effective drug combination should affect signaling pathways involved in cell proliferation and growth of the particular cancer type. In searching for a rational design in polypharmacology, one needs to first identify a set of targets whose interactions play critical roles in delivering the anticancer efficacy [9], [33]. Therefore, a fundamental computational problem is to identify a subset of therapeutic targets whose combinatorial interaction effects can be predicted in relation to cancer survival phenotypes. Note that in an individualized experimental setting, where different cancer types are tested for drug efficacy, the therapeutic targets should be also cancer-specific.
Let denote such a cancer-specific therapeutic target set. Identification of corresponds to a partition of the potential target set into two non-empty and non-overlapping groups. Let the space of distinct partitions for be denoted by . We will learn an optimal partition from such that the cancer-specific targets can be separated from the remaining ones in . We assume that the drug target inhibition profiles and the drug treatment efficacy data can be used for evaluation of target set relevance provided that is a treatment outcome of drug perturbations on cancer survival pathways by multi-target inhibition in . A plausible assumption is that the targets of more effective drugs are more likely to be involved in cancer survival pathways than those of less effective drugs. Therefore, targets that are predictive of drug efficacy are, in general, functionally important for cancer survival and should be selected with a higher probability for drug target combinations as well. More formally, the learning procedure for identifying such a cancer-specific target set is to find a model that gives the best prediction performance. We are especially modeling multiple interactions among the target set for the prediction of drug efficacies and therefore capturing the synergistic combination effects that cannot be revealed by inhibiting any of the targets individually.
Let denote the model prediction error for a drug or drug combination in a testing set. In its most basic form, the prediction error is calculated as the absolute difference between the predicted and the actual treatment efficacy:(1)where refers to the predicted efficacy for drug by a model that takes and as training data. We take here a formal model-based strategy to estimate by formulating a predictive modeling framework for any training data ; the model construction and model selection algorithms will be proposed in the sequel.
In an earlier work by Pal and Berlow [30], two fundamental set theoretic rules were exploited for predicting the drug efficacy according to its target profiles:
Construction of a TIMMA model for predicting drug efficacy requires a selection of cancer-specific target set as the model parameter. Usually is a priori unknown and need to be inferred from the potential target set . In our model-based learning framework, the likelihood of a proposed target set being composed of cancer-specific targets can be evaluated using the prediction accuracy of the corresponding TIMMA model that takes as its parameter. More formally, we consider an objective function for model selection as the average leave-one-out (LOO) TIMMA prediction error:(8)where the leave-one-out prediction error for drug is given by Eq. 1 and Eq. 3–7. Given that the combinatorial space for is huge for even a modest number of targets, it is not possible to calculate the objective function for all the possible target subsets using exhaustive enumeration. We consider a Sequential forward floating search (SFFS) algorithm modified from [34] for minimizing Eq. 8 in a computationally efficient manner. The modified SFFS algorithm learns the optimal cancer-specific target set by aggregating and subtracting targets in at different steps, as defined in the following, with the aim of minimizing the prediction error , where is the cardinality of set ,i.e.:
We have further improved the scalability of the TIMMA algorithm to large and complex data in MATLAB by exploiting its matrix computation architecture. Briefly, the TIMMA model was represented as a 3- dimensional array, where each drug's contribution to the estimate of is calculated independently of each other. This multi-dimensional data structure has enhanced the computation efficiency significantly as most of the iteration loops can be avoided. Meanwhile, independent computing enables parallel distribution of the model prediction on separate processors, e.g. one processor for one drug, which will further decrease the computation time. For the SFFS target selection, the multi-dimensional data structure also facilitates the aggregation and comparison of prediction error at the Inclusion step when the target is added to , as can be incrementally derived based on that has been obtained in the previous iterations. The TIMMA implementation code is freely available at http://timma.googlecode.com/.
For the optimal target set selected by the SFFS algorithm, the result of the TIMMA model prediction is summarized in the predicted efficacy matrix, which enumerates the treatment efficacy for each of the combinatorial target inhibition in (Figure 1C). Here, we considered the predictions for the single and pairwise target inhibitions only, and derived a synergy score for the target pair (A, B) based on the multiplicative null model:(9)where and denote the predicted efficacies for the target pair and its individual targets, respectively. The multiplicative model is widely being used in the gene knock-out studies in model organisms to score quantitative genetic interactions between gene deletions [35], [36]. It has also been recently applied to investigate genetic interactions in human cancer cells using combinatorial RNAi screening [37], as well as to characterize drug synergy effects according to the Bliss independence model [38], [39].
Using the model predictions, we can calculate the synergy score also for those drug pairs (d1, d2) whose targets are included in . If one or both of the drugs are inhibiting multiple targets, e.g. and , then we assign a drug synergy score for the drug pair using the mean of its corresponding target pair synergy scores defined by the multiplicative model (Eq. 9), i.e.(10)In the given example, A deviation of from zero provides evidence for a non-additive interaction between the two drugs, where indicates synergy and indicates antagonism.
When the target set size is fixed at two, the TIMMA model construction algorithm evaluates the pairwise target inhibitions without considering any higher-order interactions. This enables the TIMMA modeling strategy to systematically predict target pairs with synthetic lethality effect. By definition, synthetic lethality among a target pair states that: (i) inhibition of either of the single targets will result in incomplete cancer killing; and (ii) inhibition of both of the targets simultaneously will block the complete cancer survival sub-network. Therefore, the target inhibition network for the synthetic lethal target pair can ideally be represented as two nodes in parallel, similar to the topology of or shown in Figure 1E. In comparison, there are two competing models: one with no links connecting the target nodes (referred to as a singleton model), and the other with two nodes linked in a sequence (referred to as a series model). Under the series model, no synthetic lethality effect is expected since the inhibition of a single target is already sufficient to block the cancer survival pathway. Therefore, from the model fitness perspective, we are expecting higher prediction accuracy for a synthetic lethal target pair under the parallel model, compared to both the series model as well as the singleton model. To evaluate the likelihood of a parallel model against the competing models for a given target pair (A, B), we defined a synthetic lethality score as the ratio of the fitness function of these two models, given by the total sum of squares (TSS) of the predictions:(11)The synthetic lethality score is conceptually different from the multiplicative synergy score as they are addressing different questions. The synthetic lethality score evaluates the pairwise target interactions by comparing the likelihood of three competing model structures, whereas the synergy score is derived based on the model averaging by combining all the possible models. Synthetic lethality corresponds to a special case of synergy, which requires minimal individual effects that are not considered explicitly in the multiplicative synergy score. Further, the higher-order target interactions, which are evaluated during the sequential forward search for the TIMMA model, are not considered when calculating the pairwise synthetic lethality score.
To evaluate the relative efficiency and accuracy of TIMMA, we initially compared the TIMMA and PKIM algorithms on the simulated data and on the CanOS1224 canine osteosarcoma cell line. In the more practical case studies, we then applied the optimized TIMMA model to infer effective drug targets in the context of MCF7 breast cancer and BxPC3 pancreatic cancer cell lines, where kinome-wide siRNA knockdown data are publicly available for experimental validation. Finally, we evaluated the synergistic effects of the predicted drug target combinations in the MDA-MB-231 breast cancer cells by carrying out pairwise siRNA silencing screens for the TIMMA-selected kinase targets.
We started by evaluating the relative performance of TIMMA and PKIM in terms of their accuracy in predicting the treatment efficacies for new drugs on the simulated dataset. It was found out that TIMMA systematically improved the average leave-one-out (LOO) prediction accuracy, compared to PKIM, at each predefined drug-target threshold (Figure 2A, paired t-test, p = 5.0024×10−10). Since TIMMA combines the information from a drug's subsets and supersets simultaneously, its predictions are more robust to data noise and other technical factors that are inconsistent with the model assumptions, compared to PKIM, which does not consider model averaging. In particular, TIMMA gains on average 22.4% increase in the prediction accuracy especially for affinity thresholds lower than 0.8, which correspond to the promiscuous cases with, on average, more than two targets per drug (Figure 2B). These results demonstrate the importance of the improvements provided by the TIMMA algorithm, which make it applicable also to more challenging and practical cases, where target promiscuity is common and knowledge about all the cellular targets of drugs is rarely available.
Another important consideration in the large-scale drug screens is the computational complexity of the prediction algorithms. The computation times for TIMMA and PKIM model construction algorithms, SFFS and greedy search, respectively, were compared on a standard 2.6 GHz desktop computer. In contrast to the exponentially increasing time that is needed for the PKIM model construction, TIMMA takes approximately linear increase in time with the number of targets (Figure 3A). Even though the SFFS is computationally more demanding than greedy search in model selection, TIMMA achieved marked speed-up due to the optimization techniques using multi-dimensional matrix computations (Figure 3B). Notably, with 20 targets and 10 drugs, for example, the greedy search will take 10 days, while the TIMMA takes on average 30 minutes to complete, and thus saves up 99% of the computation time. The enhancement in the computation speed facilitates the analyses of larger and more complex datasets with increasing number of drugs and their target information.
We next tested whether TIMMA can lead to improvements in the real dataset used in the PKIM work [25], first by fixing the threshold at 0.9. From the set of 317 kinases, the PKIM model identified 8 kinases with a mean LOO error of 0.1314, while TIMMA identified a different set of 8 kinases with a decreased LOO error of 0.0574 (Dataset S2). When varying the threshold, the average LOO prediction accuracy of TIMMA was significantly better than that of PKIM (Figure 4A, paired t-test, p = 1.3910×10−5). Similarly as in simulated dataset (Figure 2A), the improvement in the prediction accuracy varied with the selected cut-off threshold (Figure 4A). As expected, when the threshold is close to 1, the two models performed equally well, as the drug-target information is too few to make any reliable predictions; while TIMMA again systematically outperformed PKIM at the smaller thresholds.
As revealed in many kinome-wide drug binding assays, most drugs, albeit considered previously specific to single or double targets, have shown a relatively wide range of binding affinities to multiple off-target kinases [47]. Our model can also make use of such promiscuous drug-target interactions that are informative for predicting drug cancer killing efficacies. This was further investigated in a receiver operating characteristics (ROC) analysis of the prediction performance, where the problem was to distinguish the 12 most sensitive drugs with positive efficacy values (Figure 4B). In this analysis, the area under the ROC curve (AUC) for TIMMA was 0.9679 and for PKIM 0.7144, further demonstrating the improved predictive power of the TIMMA model.
To test whether the SFFS model selection algorithm can find solutions close to the global optimal target sets, we performed an exact analysis for maximally 12 kinases, where exhaustive search can be performed at a reasonable running time. More specifically, kinases from the full set of 317 kinases were randomly selected, where , and an exhaustive search was run to determine the optimal subsets of the kinases. We applied here a fixed cut-off threshold of , which equals to the average of all the values over the drug-target pairs. The optimal sets determined by the SFFS algorithm in TIMMA and by the greedy search algorithm in PKIM were compared with the global optimum in terms of prediction accuracy. The SFFS algorithm gave significantly better results than the greedy search for (Figure 5, paired t-test, p = 3.3397×10−6). This demonstrates that the computationally efficient SFFS algorithm can find solutions that are not too far from the globally optimal solution.
After confirming the appropriate performance of the TIMMA model, we applied it to two practical case studies. In the first one, we systematically evaluated the predictions of the TIMMA MCF-7 model against the experimental results from an independent kinome-wide siRNA study in the MCF-7 breast cancer cells [48]. The knock-down data were generated using a Methylene blue assays to assess cancer cell density in order to evaluate the quality of their siRNA screen (Figure S2 and Table S2 in [48]). The siRNA screen was designed to target 712 kinases in the human kinome, with three distinct siRNAs per kinase. The data was analyzed using the R package cellHTS2 [49], where a mean Z-score scaled by the per-plate median of the intensities of the negative controls was calculated for each kinase. A large positive Z-score indicates a strong inhibition effect and thus indicates high essentiality of the kinase for the cancer cell survival.
Here, we tested the essentiality of the kinases in the cancer-specific target set predicted by TIMMA using the 15 drugs targeting a total of 384 kinases. In other words, we asked the question: are the kinases selected by TIMMA as the most predictive of anticancer efficacy also highly essential individually for the cancer cell viability? The optimal target set found by TIMMA included 12 kinases {ZAK, CSF1R, GAK, MEK5, ABL2/EPHA8, ALK/LTK/PLK4/ROS1 and MEK1/MEK2}, with a mean LOO prediction error of 0.1392 (Dataset S5). The/symbol stands for the targets that are inhibited by the same set of drugs in the data and thus are indistinguishable by the model. The mean Z-score for these 12 kinases was 0.926, which is significantly higher than the average Z-score for random sets of 9 kinases selected from the 712 kinases (Figure 6A, permutation test, p = 0.0015). This shows that TIMMA tends to choose, in general, such kinases that are also individually more effective in blocking cancer cell growth. Among these kinases, ALK had the highest predicted single-kinase efficacy. ALK was also identified in the independent siRNA screen as the top essential kinase. However, our model does not assume that all the kinases in the optimal target set are essential individually. For instance, GAK and ROS1 had a relatively low Z-score, but still these were considered to have an important role in the cancer survival and/or proliferation process when combined with the other selected kinases (Figure 6B).
On the basis of the predicted efficacy matrix based on the selected kinase targets (Dataset S5), we derived the multiplicative synergy score (Eq. 9) for the drug pairs that are pairwise inhibiting the selected targets (Supplementary Table S1). We found that the top synergistic drug pairs are mainly GAK and ALK/LTK/PLK4/ROS1 inhibitors, some of which have been reported in the recent literature. For example, crizotinib combined with erlotinib has recently been shown to cause a complete and genotype-specific inhibition of tumor growth in non-small cell lung cancer (NSCLC) adenocarcinoma patient-derived pre-clinical treatment models in vivo [50]. Crizotinib-erlotinib combination was also ranked as the top one among the 12 drugs that are available in the MCF-7 model analysis, indicating that such a combination might also be effective for the treatment of specific resistive subtypes of breast cancer. Similarly, TAE-684, a potent ALK inhibitor has been found to provide selective activity against those mutations that conferred crizotinib resistance in cancer patients [51], suggesting a mechanistic insights into the crizotinib-TAE-684 combination, which was ranked as the second most synergistic pair by our model predictions. In general, the top-predicted synergistic drug pairs are not necessarily the individually most sensitive drugs, as their individual efficacies do not correlate with the multiplicative synergy score (Supplementary Table S1).
To visualize the combinatorial effect of the selected kinase targets, a target inhibition network was constructed by applying a threshold of 0.318 to binarize the predicted efficacy (Figure 7, Dataset S5). The threshold 0.318 was the scaled drug efficacy for crizotinib that inhibits ALK, which is the most essential kinase according to the siRNA screen and thus considered as effective in treating MCF7 cancer cells. The target inhibition network suggested that two parallel MEK1/2-dependent pathways as most important for the MCF-7 cancer cell survival. For example, simultaneous targeting of CSF1R and ALK/LTK/PLK4/ROS1 was predicted to enable blocking the two redundant pathways and result in a complete inhibition of the MEK1/2-dependent cell proliferation. Notably, CSF1R has been shown to act upstream of MEK1 and to induce Cyclin D2 expression via the Ras/Raf/MAPK pathway [52]. Similarly, ALK has been suggested to directly activate MEK1/2, independent of c-Raf [53]. Also, LTK has been implicated in cell growth via MAPK signaling [54]. Taken together, these findings support the idea that inhibition of both CSF1R and ALK/LTK/PLK4/ROS1 should have a synergistic effect on the cell survival. Indeed, the combination of sorafenib and crizotinib, inhibitors of CSF1R and ALK/LTK, respectively, has been considered for a clinical trial for treating advanced solid tumors (Pfizer, ClinicalTrials.gov, Identifier: NCT01441388).
To further show the applicability of TIMMA to such cases where combinatorial effects of kinase inhibition are considered, we utilized the results from a kinome-wide drug sensitization screen, in which the kinase siRNA-silencing was combined with the treatment of Aurora kinase inhibitors in BxPC-3 pancreatic cancer cell line [55]. Aurora kinases (Aurora A, Aurora B, and Aurora C) are serine/threonine kinases that are frequently overexpressed in many tumors. Accordingly, Aurora kinase inhibition has been proposed as potential cancer therapy to disrupt cancer cell division. The purpose of the study was to identify those kinases that when silenced would sensitize pancreatic cancer cells to the Aurora kinase inhibitor treatments. The RNAi screen was done using the Human Validated Kinase Set (HVKS) siRNA library from Qiagen, with two siRNAs per kinase. A total of 17 kinases were identified and confirmed in a validation screen to have at least 2 out of 4 siRNA sequences showing greater than 1.5-fold decreases in EC50 or EC30 values of the Aurora kinase inhibitor AKI-1 in dose-response curves [55].
We wanted to evaluate here the TIMMA model performance in predicting the experimental results in [56], especially the kinases that would sensitize the pancreatic cancer cells to the AKI-1 treatment. This question can be addressed in TIMMA by determining the synthetic lethality score for such kinases paired with the targets of AKI-1. The synthetic lethality score (Eq. 11) was calculated for the kinase pairs using the data of 15 drugs and 384 kinases and the drug efficacy in BxPC-3 cells [31]. The higher the score, the stronger the synthetic lethality effect for the kinase pair. Of these 15 drugs, 3 drugs (CHIR-265/RAF-265, nilotinib and PD0332991) were not tested for BxPC-3 and thus were removed (Dataset S3). Since none of the 12 compounds effectively targeted the two Aurora kinases, Aurora A and Aurora C, we considered here the Aurora B kinase as the only effective target of AKI-1. The TIMMA model was therefore tested on all those kinase pairs which contain Aurora B, and those kinase pairs whose synthetic lethality scores were higher than that of {Aurora B, Aurora B} pair were considered as synthetic lethal partners of Aurora B.
The TIMMA analysis based on Eq. 11 identified 19 kinases (multiple kinases are ranked the same as they are targeted by the same drug set), which showed stronger synthetic lethality interactions with Aurora B than with itself (Figure 8). Two (MET, PDGFRA) out of the three targets (MET, PDGFRA and PYK2) were experimentally validated as sensitizing targets of AKI-1 in the pancreatic cancer, representing a highly significant enrichment (hypergeometric test, p = 0.0046) (Figure S4 in [55]). In addition, the model predicted that PDGFRB might also be a potential sensitizer of AKI-1 treatment. Similar to the result in the MCF-7 cells, ZAK (ranked 3rd), MEK5 (ranked 7th) and GAK (ranked 9 th) were again found in the cancer-specific target set for BxPC-3 cells, suggesting that the synergy patterns of these kinases is common across these cancer types. In contrast, the model predicted that the combination of MEK1/MEK2 and AURKB inhibitors has least synthetic lethal capacity (Dataset S6), because individual essentiality of these two factors favors the series connection model rather than the parallel model in the synthetic lethality score [56], [57].
The final application case study was the human triple-negative breast carcinoma, where we experimentally validated the TIMMA target combination predictions using single and pairwise siRNA knock-downs on the MDA-MB-231 cells. The TIMMA model selected 20 optimal kinase targets {PLK1, AURKB, CDKL2, ZAK, ERBB4, TEK, TXT/BMX/CSK/EPHA5/EPHB1/EPHB4, CAMKK1/MAK/VRK2/TNNI3K/CDC2L6/DYRK1B/DYRK1A/TYK2} with an average LOO error of 0.11 (Dataset S7). These kinases and their functional interactions were mapped to the target inhibition network, which contained a total of 8 target nodes (Figure 9). The kinases belonging to the same node are inhibited by a common set of drugs, and therefore these drug targets are indistinguishable in terms of drug inhibition and their predicted efficacy values. Two of the selected kinase targets, PLK1 and AURKB, are known to be essential for cell growth, therefore serving here as positive controls for the model target predictions. However, due to their known role in cell growth, we excluded these two kinases from the experimental evaluation, and focused on the synergistic combinations between the remaining 18 kinases targets among the 6 target nodes.
In general, there were significant differences between the TIMMA-selected kinase targets, when these were silenced either individually or in combination in the siRNA screens, especially after their ranking according to the predicted efficacy (Figure 10A, Kruskall-Wallis rank sum test, p<10−15). Even after excluding the two essential kinases (PLK1 and AURKB), the 18 TIMMA-selected kinases showed higher cancer cell growth inhibition power in the single knock-down experiments (22% increase in cell inhibition), compared to the inhibition observed in the kinome-wide single-siRNA screen (Wilcoxon rank sum test, p = 0.28, Supplementary Table S2). Importantly, the 153 TIMMA-selected kinase pairs resulted in highly significant cancer cell killing improvement in the pairwise knock-down experiments (38% increase), compared to their single kinase inhibition efficacy (p = 0.0089, Bonferroni adjustment), indicating that TIMMA could select such kinase targets that, in general, are important for cancer cell survival, and especially when combined. Notably, when categorizing the selected target pairs as High and Low efficacy groups, according to their predicted treatment efficacies above or below the average of 0.6, there was a significant increase in the cancer cell growth inhibition percentages (23%, 48% and 80%), when comparing the High efficacy group to either the Low efficacy group, the single selected kinases or the kinome-wide background (p = 0.031, p = 0.013, p<10−15, Bonferroni adjustment, Supplementary Table S2). Taken together, these results indicate that the TIMMA model can effectively select and prioritize among the massive number of possible combinations those target combinations that are most potential for experimental testing or eventual clinical translation.
To investigate whether the model can select also such drug target combinations that individually show relatively low drugs efficacies, but will lead to increased drug synergy when combined, we focused on the set of 15 kinase pairs among the 6 target nodes ({CDKL2, ZAK, ERBB4, TEK, TXT/BMX/CSK/EPHA5/EPHB1/EPHB4, CAMKK1/MAK/VRK2/TNNI3K/CDC2L6/DYRK1B/DYRK1A/TYK2}, Figure 9) that are unique in terms of their drug profiles and thus distinguishable based on their TIMMA-predicted efficacy. We took an average of the synergy scores for those kinas pairs that are represented by the same target node pair. The synergy score calculated on the basis of the TIMMA-predictions correlated significantly with the synergy calculated on the basis of the single and pairwise siRNA measurements (Kendall correlation 0.39, p = 0.0463). When mapping the selected kinase target pairs to the available kinase inhibitor pairs, i.e. using Eq. 10, the correlation between the predicted and measured synergies improved further (Figure 10B, p = 0.0002). In particular, when using a cut-off predicted synergy of 0.36 (the dotted vertical line), the likelihood of obtaining a high measured synergy increased significantly (Wilcoxon rank sum test, p<5.9−7, Bonferroni adjustment). Among these top-20 most synergistic drug combinations for the MDA-MB-231 cells, there were a number of examples, such as the two top pairs, where the efficacy of one of the drugs in the combination was relatively low, or even zero, yet the predicted and measured synergy for the drug combination was high (Table 1). This demonstrates that our model is able to predict not only those pairs that are essential either individually or in combination, but also a number of synergistic combinations, where the predicted efficacy cannot be explained by the efficacy of the two single compounds when used alone (Supplementary Figure S4).
In this study, we utilized the principles of polypharmacological target inhibition modeling as a generic framework for pinpointing cancer-specific targets and predicting the effect of putative drug combinations. The main contribution of the present work was to introduce a novel model construction model, called TIMMA, and to demonstrate its feasibility in systematic investigation of the model predictions using kinome-wide single and pairwise siRNA knock-down experiments. We also showed that our enhanced model construction algorithm resulted in significantly better predictive accuracy and computational efficiency, compared with an existing algorithmic solution. With such improvements, the number of targets that can be included in the minimal set can go up to 20, which corresponds to maximally 20 drugs in a combination. In the three case studies, where we combined large-scale drug sensitivity screening and comprehensive drug-target data, we were able to identify a number of potential drug combinations for breast and pancreatic cancers. In more general terms, the optimized experimental-computational approach, empowered by the target inhibition network, allowed us to systematically explore how the kinase inhibitors and their cellular targets interact to modulate cancer growth phenotype on a global network-level, with the aim to identify molecular pathways behind drug action, as well as to suggest combinatorial treatment strategies that can block the cancer escape pathways and therefore tackle the resistance problem of the many current treatments approaches.
Network-based strategies, such as the one developed in the current work, provide a principled approach to systematically identify the key set of druggable vulnerabilities of cancer networks. Such efforts create a solid foundation towards implementing the emerging paradigm in drug discovery, the so-called ‘network pharmacology’ [3], which provides a more global understanding of the mechanism behind drug action and resistance by considering drugs and targets in their context of cellular networks and pathways. The current work also support the detection of synthetic lethal interactions, which is another conceptual framework recently proposed toward developing more effective therapeutic strategies [12], [13], [15]–[19]. More specifically, targeted perturbation or inhibition of a gene that has a synthetic lethal relationship with a driving cancer mutation holds great promise for being a highly specific and selective means to kill cancer cells without severe side-effects to normal cells. Compared to the conventional cytotoxic drugs, that affect both normal and cancerous cells, synthetic lethality can therefore address the fundamental challenges of anticancer therapy by optimally targeting differential features in each cancer type while sparing normal cells. However, despite the advances in siRNA and compound screening, synthetic lethal interactions between genes and/or drugs have remained extremely difficult to predict on a global scale [13], [18]. Network-based methods provide a convenient platform to finding functional interactions in disease networks, toward enabling identification of such effective drug targets and their combinations that tailored for more effective and personalized cancer medicine.
We focused here on the kinase targets because of their importance in many multi-target cancer treatment developments. This is also why we experimentally validated the model predictions using kinome-wide single siRNA and TIMMA-predicted pairwise siRNA screens, where the selected kinase targets were knocked down individually or in pairs in the given cell type to experimentally evaluate their essentiality either alone or in combination for the cancer cell survival. However, the same modeling principles could be applied also to other target families, such as enzymes or G protein coupled receptor (GPCR) targets, provided there will be enough target and drug promiscuity to allow for construction of the target inhibition networks. Moreover, while the siRNA silencing screens are convenient for the drug target investigation, the perturbation effects from the siRNAs cannot fully mimic the phenotypic effects of drug treatments. RNAi has also potential limitations due to potential off-target silencing effects and variable reagent efficacy, which may also partly explain the observed discrepancies between the drug treatment-based model predictions and their siRNA-based experimental validations. Therefore, one of our future aims is to apply the TIMMA model predictions to designing potential drug combination treatments, initially in various cancer cell models in vitro, and later also in primary samples from cancer patients ex-vivo. The drug treatments are also closer to the eventual translation of the model predictions in a clinical setup, at least until the RNAi-mediated target silencing has become safe and efficacious enough for clinical applications.
In an effective combinatorial setting, one needs to modulate a set of targets to achieve maximal efficacy, while avoiding others to reduce the risk of side effects. The current TIMMA algorithm addressed the first challenge: the optimal efficacy by multi-target modulation. The different model parameters and thresholds lead to a multiple candidate target inhibition networks for combinatorial treatments. From those candidate models, clinician could then ideally choose the combination that is most feasible and results in less known adverse effects, based on prior knowledge. Although there are information sources on drug side effects scattered around in databases, such as SIDER [58], ChEMBL [32], and PROMISCUOUS [59], we chose not to try to incorporate the side effect information in the current model building, because such information is still missing for many targeted drugs and the initial aim was to find effective target combinations. However, incorporating known side effect or toxicity information of drugs and their targets will be an important topic of future research. Possible approaches for such modifications include, for instance, usage of metabolic networks and pathways that are targeted by drugs [60], or combining multiple databases that contain a collection of drug features, such as medical indications, molecular targets, toxicity profiles or anatomical therapeutic and chemical classifications [61]. Further, rather than using a single response readout for drug efficacy, such as IC50, AA or DSS, the gene expression or metabolomic changes after a treatment could also be included as part of the drug response profiles, perhaps leading to be more comprehensive drug-disease networks in the future.
|
10.1371/journal.pgen.1003466 | Identification of a Tissue-Selective Heat Shock Response Regulatory Network | The heat shock response (HSR) is essential to survive acute proteotoxic stress and has been studied extensively in unicellular organisms and tissue culture cells, but to a lesser extent in intact metazoan animals. To identify the regulatory pathways that control the HSR in Caenorhabditis elegans, we performed a genome-wide RNAi screen and identified 59 genes corresponding to 7 positive activators required for the HSR and 52 negative regulators whose knockdown leads to constitutive activation of the HSR. These modifiers function in specific steps of gene expression, protein synthesis, protein folding, trafficking, and protein clearance, and comprise the metazoan heat shock regulatory network (HSN). Whereas the positive regulators function in all tissues of C. elegans, nearly all of the negative regulators exhibited tissue-selective effects. Knockdown of the subunits of the proteasome strongly induces HS reporter expression only in the intestine and spermatheca but not in muscle cells, while knockdown of subunits of the TRiC/CCT chaperonin induces HS reporter expression only in muscle cells. Yet, both the proteasome and TRiC/CCT chaperonin are ubiquitously expressed and are required for clearance and folding in all tissues. We propose that the HSN identifies a key subset of the proteostasis machinery that regulates the HSR according to the unique functional requirements of each tissue.
| The heat shock response (HSR) is an essential stress response that functions to maintain protein folding homeostasis, or proteostasis, and whose critical role in human diseases is recently becoming apparent. Previously, most of our understanding of the HSR has come from cultured cells and unicellular organisms. Here we present the identification of the heat shock regulatory network (HSN) in Caenorhabditis elegans, an intact, multicellular organism, using genome-wide RNAi screening. We identify 59 positive and negative regulators of the HSR, all of which have a previously established role in proteostasis, linking the function of the HSR to its regulation. Some HSN genes were previously established in other systems, many were indirectly linked to HSR, and others are novel. Unexpectedly, almost all negative regulators of the HSR act in distinct, tissue-selective patterns, despite their broad expression and universal cellular requirements. Therefore, our data indicate that the HSN consists of a specific subset of the proteostasis machinery that functions to link the proteostasis network to HSR regulation in a tissue-selective manner.
| The heat shock response (HSR) has been studied extensively as a cellular response to acute stress such as elevated temperature [1]. The master regulator of the HSR is Heat Shock Factor 1 (HSF1), a stress responsive transcription factor that regulates the inducible transcription of a family of genes encoding heat shock proteins (HSPs), many of which are molecular chaperones. In the absence of a stress signal, HSF1 is inhibited by a negative feedback loop mediated by the molecular chaperones HSP70 and HSP90 [2]–[7]. Upon heat shock, HSF1 is activated as the equilibrium of chaperones shifts toward association with metastable polypeptides.
Many key aspects of the HSR have been well established at a cellular level in cultured cells and unicellular organisms, yet the HSR has additional features that are only apparent in multicellular organisms. Heat shock inducible promoters contain multiple cis elements and can be differentially expressed across tissues [8]–[18]. The HSR is intimately associated with numerous tissue-specific and age-dependent human diseases and regulated cell non-autonomously by neuronal control [19], [20]. Finally, HSF1 has important roles during development and longevity, and activation of the HSR is attenuated during aging [13], [21]–[24]. However, despite the importance of the HSR in organismal physiology, relatively little is known about its regulation in multicellular organisms and the extent of differential regulation across distinct tissues is unexplored.
A comprehensive genetic analysis of the HSR regulatory pathways has not previously been possible in any system, in part because traditional forward genetic screens are inadequately suited to the identification of genes that regulate the HSR. These approaches depend on the introduction of mutations, which can destabilize the folding of the corresponding proteins, resulting in indirect induction of the HSR due to the expression of misfolded species. Indeed, a forward genetic screen in Drosophila described such mutations in a muscle-specific actin [25], [26]. RNAi based genetic screening resolves the limitations associated with traditional genetic screens associated with the HSR and has been used to gain important insights into many regulatory networks including those associated with models of aggregation-prone proteins, longevity, and stress responses [27]–[34].
In this study, we have used genome-wide RNAi screening to identify factors important for the positive and negative regulation of the HSR in the metazoan Caenorhabditis elegans in order to establish a comprehensive understanding of its regulation on an organismal level. Further, we used a fluorescent reporter to allow for the analysis of regulation in different tissues. This approach reveals a complex network of positive and negative HSR regulators with critical roles in maintenance of proteostasis that confer differential tissue-selective patterns of heat shock gene expression.
The genetic network upstream of HSF1 and the HSR was identified using a genome-wide RNAi screen in transgenic C. elegans expressing the heat shock (HS)-inducible fluorescent reporter phsp70::gfp constructed from the promoter of the C12C8.1 gene [13]. Expression of this reporter is not detected under ambient growth conditions of development and adulthood (Figure 1A) and is induced strongly by HS (Figure 1B). The threshold sensitivity of the screen was established using RNAi knockdown of hsf-1 to suppress HS-induction of the reporter as a reference control for positive regulators (Figure 1C), and RNAi knockdown of hsp-1, a member of the HSP70 family that negatively regulates the HSR, resulting in constitutive expression of the reporter as a reference control for negative regulators (Figure 1D).
Genetic modifiers of the HSR were identified by visual scoring of the phsp70::gfp reporter upon RNAi-mediated knockdown. A representative subset from each functional class was validated by analysis of endogenous hsp70 gene expression using qRT-PCR (Figure 1E and 1F). We also extended our analysis to another heat shock gene, hsp-16.2, a member of the small HSP family. Consistent with the HSR reporter results, HS-dependent induction of hsp70 and hsp-16.2 were reduced upon hsf-1 knockdown. Likewise, the basal expression of hsp70 and hsp-16.2 were increased upon hsp-1 knockdown. These experiments established the utility of the phsp70::gfp reporter and RNAi as a methodology for the identification of both positive and negative regulators of the HSR in C. elegans.
Having established the criteria for two classes of HSR genetic modifiers, we performed a genome-wide RNAi screen for genes whose knockdown blocked HS-dependent reporter induction, and for genes whose knockdown resulted in constitutive expression of the reporter. These screens were performed by RNAi feeding using a library containing RNAi constructs targeted against approximately 86% of genes in C. elegans [35].
The screen for positive regulators of the HSR identified genes with properties similar to hsf-1, whose knockdown suppressed induction of the HSR reporter. To ensure that decreased fluorescence of the reporter did not arise from indirect effects, such as transgene silencing, we performed a counter-screen against suppression of a phsp-4::gfp reporter, an ER stress-inducible gene that is not dependent on HSF1 [36], [37]. This led to the identification of seven positive regulators that are conserved to humans and function in chromatin remodeling, RNA processing, and protein synthesis (Figure 1E, Table 1, Table S1). None of these genes has been previously linked to HSR regulation, however each has been either associated with HS or implicated in the HSR. For example, dcp-66 is a subunit of the NuRD complex, of which other subunits in this complex have been shown to interact with human HSF1 [38]. Our data provide evidence that the HSF1-NuRD interaction has functional consequences on the regulation of the HSR. Likewise, Mi-2, a subunit of several complexes including NuRD, has been shown to affect the levels of HS genes in Drosophila [39]. Among the other positive regulators are genes associated with mRNA splicing and translation, biosynthetic processes that are highly sensitive to HS stress. F09D1.1 is a homologue of USP39, which has been implicated in recycling of the triple-snRNP complex, a step of splicing that is particularly sensitive to temperature. phi-8 and phi-11 are subunits of Splicing Factor 3, which has been shown to regulate alternative splicing, snr-3 is an sm protein which is expected to have a general role in mRNA splicing, and eftu-2 is an elongation factor 2-like protein predicted to have a general role in translational elongation. Finally, as expected, hsf-1 was identified in the screen as predicted for its central role in the HSR.
The screen for negative regulators of the HSR identified genes whose reduced expression resulted in the constitutive expression of the phsp70::gfp reporter. To ensure that these regulators activated the HSR in an HSF1-dependent manner, we employed a subsequent counter-screen using a hypomorphic hsf-1 mutant [40]. Candidate negative regulators were also tested for their ability to constitutively activate endogenous heat shock genes by qRT-PCR (Figure 1F). This strategy led to the identification of fifty-two genes that have the functional properties of negative regulators of the HSR (Table 2, Table S1).
Each of these negative regulators of the HSR function in specific steps of proteostasis and affect either gene expression, protein folding, trafficking, and clearance, and are conserved to humans. Among the regulators that affect protein folding are three prominent molecular chaperone machines corresponding to HSP70 (hsp-1), HSP90 (daf-21) and TRiC/CCT (cct-1, cct-2, cct-3, cct-4, cct-5, cct-6, cct-7, and cct-8) and three cochaperones (sgt-1, unc-45, and cyn-11). HSP70 and HSP90 are predicted from previous studies that identified them as negative regulators of HSF1 and the HSR. Likewise, a role for the TRiC/CCT chaperonin in the regulation of the HSR has been suggested from studies on a small molecule that interacts with TRiC/CCT and induces human HSF1 [41]. Regulation of the HSR by chaperonins is functionally conserved in bacteria, as downregulation of the prokaryotic chaperonin GroEL induces the HSR in E. coli [41], [42]. The selectivity of these genes representing three chaperone machines and three cochaperones as regulators of the HSN is unexpected given that C. elegans expresses nearly 200 chaperone genes, which suggests a high degree of selectivity for chaperone regulation of the HSR.
Other negative regulators of the HSR correspond to components of trafficking including subunits of the signal recognition particle (SRP) and other secretory pathway genes (F55C5.8, F08D12.1, F25G6.8, F38A1.8, R186.3, F38E11.5 and T14G10.5, hsp-3, T24H7.2, let-607) and mitochondrial import (hsp-6, T09B4.9). Consistent with this, knockdown of SRP subunits in yeast and E. coli has been shown to induce the HSR [43], [44] and our study now extends these observations to metazoans. Clearance components include ubiquitin associated (phi-32, uba-1, C53A5.6) and proteasomal subunits (pas-4, pas-5, pas-6, pbs-2, pbs-3, pbs-4, pbs-5, pbs-6, pbs-7, rpt-1, rpt-3, rpt-4, rpt-5, rpt-6, rpn-1, rpn-2, rpn-6, rpn-7, rpn-8, and rpn-11). Inhibition of the proteasome by small molecules has previously been shown to induce the HSR [45], [46]. It is intriguing that only the proteasome, and not autophagy or other proteases, functions as a regulator of the HSR, given the large number of components involved in protein clearance. The final class of regulators are involved in protein synthesis (dars-1) and gene expression (W04A4.5, pyp-1, and mdt-15). Microarray results confirm the induction of HSR genes upon mdt-15 knockdown [47]. While the pyp-1 subunit of the NuRF chromatin remodeler has not been previously linked to the HSR, other subunits of NuRF have been suggested to positively affect HSR gene expression [48]. Because we identified only one of 171 predicted E3-ligases (C53A5.6) and only one of 33 predicted tRNA synthetases (dars-1) in the C. elegans genome, we rescreened all members of these gene families and found no additional HSR regulators [49], [50].
A striking feature of the negative regulators of the HSR is that the HSR reporter in not uniformly induced across all tissues, but rather displays tissue-selective expression patterns in the intestine, muscle, and spermatheca (Figure 2, Table 2). Of the negative regulators, only knockdown of HSP70 and HSP90 induced expression of the reporter in all three tissues. By comparison, knockdown of the three cochaperones and all eight subunits of the TRiC/CCT chaperone machine induced the reporter only in the muscle. In contrast, downregulation of subunits of the proteasome and many secretory pathway genes induced the reporter only in the intestine and spermatheca, but not in the muscle. Knockdown of the remaining genes induced the reporter only in the intestine. These patterns are unlikely to be due to RNAi artifacts because the tissue-selective patterns of reporter induction were similar for all subunits within specific complexes, yet non-overlapping between different complexes (i.e., all proteasomal subunits induced in intestine and spermatheca, and all TRiC/CCT subunits induced in the muscle).
Tissue-selective expression of the HSR reporter was unexpected as nearly all of the negative regulators are ubiquitously expressed components of essential cellular machines. For example, even though proteasomal subunits do not induce the HSR in muscle, it has been shown in C. elegans that most, if not all, proteasomal subunits are expressed in muscle and that RNAi knockdown of proteasomal subunits yields muscle specific phenotypes such as stabilization of a ubiquitin-GFP reporter in the muscle and early onset aggregation of a polyQ disease model expressed only in the muscle [28], [51]. These results also suggest that depletion of subunits from complexes such as the proteasome does not induce the HSR simply by misfolding other subunits in that complex since these effects would not be expected to have tissue selectivity.
To further investigate the tissue-selective patterns of HSR regulation, we examined the expression of two additional reporters that are inducible by heat shock and dependent on HSF1. The phsp-16.2::gfp reporter is inducible in the intestine, muscle and excretory system, and is dependent on HSF1 and DAF-16 (Figure S1) [52], [53]. The pckb-2::gfp reporter is inducible only in the intestine and is also activated by the unfolded protein response, an ER stress response [54]. Knockdown of HSR negative regulators revealed highly overlapping patterns of tissue-specific induction with all three reporters (Table 3). In the muscle, there was a highly consistent pattern of induction between the C12C8.1 and hsp16.2 reporters, with two genes inducing both, seven inducing neither, and only a single gene showing differential induction. In the intestine, HSR negative regulators gave identical patterns of induction for the C12C8.1 and the ckb-2 reporters, with nine out of ten causing induction, yet only a smaller subset, three out of ten, also induced the hsp16.2 reporter. Given the differences in the regulation and function of the three genes, the three reporters demonstrate remarkably consistent patterns of tissue-selective HSR induction.
We next validated the tissue-selective effects using pharmacological inhibitors and mutants. We found that incubation of L4-staged worms with MG132, a pharmacological inhibitor of the proteasome, caused induction of the phsp70::gfp reporter in the intestine and the spermatheca, but not in the muscle tissue (Figure 3A). This pattern matches that seen with RNAi knockdown of proteasomal subunits, providing further support that the tissue-selective effects are unlikely to be an RNAi artifact. Most of the negative regulators are essential, however we were able to test the effects of mutations in T24H7.2, an ER localized HSP70, and the cochaperones unc-45, sgt-1, and cyn-11, and found using qRT-PCR that the expression of endogenous HSR genes was induced (Figure 3B). We further demonstrated that this induction was tissue-selective using qRT-PCR analysis on dissected intestinal cells. We found that T24H7.2 mutant animals, but not unc-45 mutant animals, induced endogenous hsp70 in the intestine (Figure 3C). The tissue-selective induction of endogenous genes in the intestine by these mutations matched the induction of the phsp70::gfp reporter by RNAi knockdown, thus providing a validation of both the use of RNAi and the fluorescent reporter.
The genes that we identified form a genetic regulatory network of the HSR in C. elegans. To characterize the relationship between these regulators, we utilized an interaction network from previous physical, genetic, and predicted interaction data [55]. A network representation of the interaction data, in which HSR regulators are nodes and interactions between them are edges, reveals that HSR negative regulators are enriched in interactions with other HSR negative regulators: 39 of 52 negative regulator genes are connected in a single interaction network (Figure 4). We next applied a community detection algorithm to determine the structure of this interaction network [56], [57]. This analysis shows that the network is composed of three distinct modules, indicated by the shapes of the nodes. The modular structure of this network is unlikely to have arisen by chance since it does not appear in randomized networks containing the same number of nodes and connections (p<10−4). The three modules are primarily composed of clearance, cytoplasmic protein folding, and gene expression and protein synthesis components, respectively. While it is unsurprising that proteasome or protein folding subunits cluster together into distinct modules, the existence of the third module is entirely unexpected. The modules identified using the interaction data (node shapes) correspond closely with the observed tissue patterns of reporter induction (node colors) thus providing additional validation of both the specificity of tissue expression and network structure. These results further suggest that the underlying functional modules give rise to the tissue-specific patterns of HSR induction.
To further probe the genetic properties of the HSN, we investigated the relationships between the positive and negative regulators to provide a systems-level pathway analysis. We tested whether depletion of positive regulators (that decrease reporter induction by heat shock) would suppress reporter induction mediated by depletion of negative regulators. We found that knockdown of each positive regulator prior to knockdown of the negative regulator hsp-1 (HSP70) decreased induction of the reporter (Figure 5A). This indicates that the positive HSR regulators are epistatic to HSP70. These data are consistent with a model in which the positive regulators of the HSR act at or downstream of chaperone-mediated regulation of the HSR. Similar results were obtained for other negative regulators including daf-21 (HSP90), pas-4 (proteasome), C53A5.6 (E3 ubiquitin ligase), let-607 (ER transcription factor), F38A1.8 (SRP), hsp-6 (mitochondrial HSP70), and dars-1 (Asp tRNA synthetases).
These results, in addition to the tissue-independent nature of the positive regulators and their association with biosynthetic processes, further distinguish the roles of the positive regulators from the negative regulators. RNAi knockdown is not equivalent to genetic ablation, so these relationships correspond to sensitivities rather than absolute dependencies. Therefore, we tested the effects of depletion of the positive regulators in a strain containing a deletion in the negative regulator T24H7.2, an ER localized HSP70. Each of the positive regulators decreased induction of the HSR upon mutation of T24H7.2, thereby confirming the results with double RNAi (Figure 5B). Together, our data indicate that the positive regulators are epistatic to the negative regulators and either function downstream or at the same step in the pathway. We favor the latter model and propose that the positive and negative regulators function together in an integrated HSR regulatory network (Figure 6).
Our findings indicate that the heat shock regulatory network (HSN) enables the HSR to sense and respond to a wide range of disruptions in proteostasis, thus providing a direct link between the function of the HSR and its regulation. The four functional clusters of the HSN each identify a small subset of the entire proteostasis machinery that functions in gene expression and protein synthesis, folding, trafficking, and clearance. The negative regulators fall within each of these functional categories, whereas the positive HSR regulators are more restrictive and cluster only to gene regulation and protein synthesis. Previous studies on the mitochondrial stress response have revealed that depletion of specific subunits of electron transport chain complexes leads to induction of the mitochondrial stress response; however in our screen we did not identify subunits of macromolecular complexes as regulators of the HSR, thus revealing differences in how these two compartments detect and respond to a proteostatic imbalance [58].
The negative regulators of the HSN displayed a surprising extent of tissue-selective effects on HS gene expression, which may arise from differences in the expression levels or activity of the regulators between tissues. One clear example of differential tissue specific expression of HSR regulators occurs in activated B cells, that rely heavily rely upon the secretory pathway and exhibit high expression levels of secretory pathway components (for a review, see [59], [60]). But as described in the results section, nearly all components of the HSN are broadly expressed. For example, TRiC/CCT is required for the folding of actin and tubulin, which are expressed in every cell [61]. However, the specialized function of muscle tissue could necessitate a stronger dependency for actin and myosin, which in turn explains the functional requirement for TRiC/CCT and account for the enhanced sensitivity of muscle to TRiC/CCT depletion. In addition to differential sensitivity to the regulators, our results indicate that each tissue exhibits distinct profiles of HS-inducible genes, which likely arises from tissue-specific factors that influence HS gene inducibility. Together, these data indicate that in addition to its unique proteome and specialized function, each tissue may contain a distinct complement of the proteostasis machinery, a differential sensitivity to disruption of proteostasis networks, and a distinct response to proteostasis disruption.
Induction of the HSR has been shown to be protective in multiple models for diseases of protein conformation; therefore, knockdown of the negative regulators and induction of the HSR would be predicted to suppress protein aggregation. Instead, there is substantial overlap (twenty out of fifty-two genes) between the negative regulators of the HSR and a separate genome-wide screen for enhancement (early onset) of polyQ aggregation in muscle (Table S1). The common gene set includes the TRiC/CCT chaperonin (6), HSP70, mitochondrial HSP70, the proteasome (10), ubiquitin, and the E1 ubiquitin ligase. Therefore, it is likely that knockdown of these genes leads to both a disruption of proteostasis and activation of the HSR. Consistent with this model, there is almost no overlap with another genome-wide screen for suppression of polyQ aggregation in muscle. In contrast, knockdown of positive regulators, which suppresses the HSR, would be predicted to cause early-onset polyQ aggregation, and indeed, five of the seven positive HSR regulators have this phenotype and none suppress polyQ aggregation.
The paradigm for HSR regulation has previously focused on HSF1 and the negative feedback loops consisting of the HSP70 and HSP90 chaperones. Our results reveal that the network that regulates the HSR is much larger and corresponds to at least fifty-nine genes of this newly defined HSN. Many of these genes have been previously linked to HSR regulation in other systems, including prokaryotes, suggesting that this regulatory network is likely conserved through evolution. The precise mechanistic links between many of these genes and the HSR and other components of the HSN remains to be defined, and it will be important to investigate whether the tissue-selective regulation of the HSR is also conserved. Nevertheless, the identification of these genes in a comprehensive genetic screen for HSR regulators not only validates their functional properties but also reinforces the evolutionary conservation of the HSR. In summary, the systems-level identification and characterization of the HSR regulatory network described in this paper provides several important insights into regulation of the HSR during stress and provides a basis for future analysis of HSR regulation during development, ageing, and human disease.
Nematodes were handled and analyzed using standard laboratory techniques and cultured at 20°C [62]. Worms were synchronized by bleaching with hypochlorite (NaOCl) and hatching overnight in M9. Where indicated, intestines were dissected from living animals in M9 media. All nematode strains were derived from the N2 Bristol wild-type strain. The following strains were used: 1) AM446 rmIs223[phsp70::gfp; pRF4(rol-6(su1006))]; 2) SJ4005 zcIs[phsp-4::gfp]V; 3) CL2070 dvIs70 [phsp-16.2::gfp; rol-6(su1006)]; 4) BC14636 dpy-5(e907) I; sIs13872[rCesB0285.9::gfp+pCeh361]; 5) PS3551 hsf-1(sy441)I; 6) AM658 hsf-1(sy441)I; rmIs223[phsp70::gfp; rol-6 (su1006)]; 7) RB1694 T24H7.2(ok2107) II; 8) RB703 unc-45(ok468) III; 9) RB1053 R05F9.10(ok1000) II; and 10) VC1372, rab-21&cyn-11(ok1879) II [13], [37], [40], [52], [63].
Genome-wide RNAi screening was performed using a bacterial feeding approach with a library targeting approximately 86% of the C. elegans genome (MRC Geneservice, Cambridge, U.K.). Bacterial cultures were grown overnight in LB with 5 µg/ml tetracycline and 50 µg/ml ampicillin and induced with 1 mM IPTG for four hours. To avoid L1-stage developmental arrest associated with essential genes, L1 larvae were allowed to develop for 19 hours on plates containing OP50 bacteria prior to exposure to RNAi.
The genome-wide screen was performed in 96-well liquid cultures containing approximately 10 animals, 50 µl M9, 5 µg/ml cholesterol, 5 µg/ml tetracycline, 50 µg/ml ampicillin, 0.4 mM IPTG, 0.1 µg/ml fungizone, and 75 µl of RNAi bacterial suspension and grown at 20°C for 60 hours in a temperature-controlled shaker. For the heat shock screens, the animals were sensitized by exposure for two hours at 24°C, 24 hours before screening for reporter induction. The heat shock conditions are at 31.5°C for two hours followed by 24 hours of recovery at 20°C prior to screening for stress-induced fluorescence. Screening was performed using Leica MZ16-FA fluorescence microscope equipped with a GFP2 filter.
Validation and analysis of the regulators from the primary screen were done using solid RNAi plates containing nematode growth medium (NGM) agar with 5 µg/ml tetracycline, 50 µg/ml ampicillin and 1 mM IPTG and seeded with RNAi bacteria. Synchronized worms grown on OP50 bacteria for 19 hours were incubated on RNAi plates for 48 hours before analysis of induction (negative regulators) or wrapped in parafilm and heat shocked in a water bath at 33°C for 1 hour and then recovered for 24 hours prior to analysis (positive regulators). Worms were immobilized in levamisole and imaged using either a BD Pathway 435 High-content Bioimager (BD Biosciences) or a Zeiss Axiovert 200 fluorescent microscope. A gene was scored as positive only if >20% of animals demonstrated induction. Epistasis analysis was performed by knockdown of each positive regulator as before followed by double RNAi of the positive and negative regulators together.
Each RNAi construct was validated by sequencing. Functional information on the identified genes was collected using WormBase [64].
Pharmacological inhibition of the proteasome was conducted using transgenic animals carrying the phsp70::gfp reporter grown on standard NGM plates seeded with OP50 bacteria. L4 larval stage animals were incubated with 100 µM MG132 (AG Scientific) in 0.5% DMSO or 0.5% DMSO alone for 2–3 hours and returned to plates. Fluorescence was scored the next day in young adult animals.
Transgenic animals carrying the fluorescent reporter were mounted on 3% agarose pads, immobilized with 2 mM levamisole and viewed using the Zeiss Axiovert 200. Animals were imaged using 10X/0.25 A-Plan and 100X/1.4 oil DIC Plan-APOCHROMAT objectives. Images were captured using a Hamamatsu digital camera (C4742-98) with Axiovision Release 4.7 software. Tissue-identification was based on nematode anatomy and tissue morphology using images from the C. elegans atlas [65]. A tissue was scored as positive only if >20% of animals demonstrated induction.
RNA was isolated from whole animals lysed by vortexing for twenty minutes after addition of TRIzol (Invitrogen) and DNA was removed using a DNA-free Kit (Ambion) according to standard protocols. cDNA synthesis was performed using an iScript cDNA Synthesis Kit (BioRad) and qRT-PCR was performed using an iQ SYBR Green Supermix Kit (BioRad) using provided protocols and run on a BioRad iCycler. 18S RNA was used as a normalization control.
We utilized a graph partitioning scheme that separates the network into groups of nodes, which collectively maximizes the density of within-partition edges in the network [56], [57]. The significance of the number of interactions between the negative regulators was tested by comparing their density to the density of interactions predicted genome-wide in C. elegans. The significance of the modularity of the HSR negative regulator network was tested by sampling Monte Carlo realizations in which we exchanged pairs of edges, maintaining the degree distribution of the network.
|
10.1371/journal.pcbi.1000871 | Probability Matching as a Computational Strategy Used in Perception | The question of which strategy is employed in human decision making has been studied extensively in the context of cognitive tasks; however, this question has not been investigated systematically in the context of perceptual tasks. The goal of this study was to gain insight into the decision-making strategy used by human observers in a low-level perceptual task. Data from more than 100 individuals who participated in an auditory-visual spatial localization task was evaluated to examine which of three plausible strategies could account for each observer's behavior the best. This task is very suitable for exploring this question because it involves an implicit inference about whether the auditory and visual stimuli were caused by the same object or independent objects, and provides different strategies of how using the inference about causes can lead to distinctly different spatial estimates and response patterns. For example, employing the commonly used cost function of minimizing the mean squared error of spatial estimates would result in a weighted averaging of estimates corresponding to different causal structures. A strategy that would minimize the error in the inferred causal structure would result in the selection of the most likely causal structure and sticking with it in the subsequent inference of location—“model selection.” A third strategy is one that selects a causal structure in proportion to its probability, thus attempting to match the probability of the inferred causal structure. This type of probability matching strategy has been reported to be used by participants predominantly in cognitive tasks. Comparing these three strategies, the behavior of the vast majority of observers in this perceptual task was most consistent with probability matching. While this appears to be a suboptimal strategy and hence a surprising choice for the perceptual system to adopt, we discuss potential advantages of such a strategy for perception.
| For any task, the utility function specifies the goal to be achieved. For example, in taking a multiple-choice test, the utility is the total number of correct answers. An optimal decision strategy for a task is one that maximizes the utility. Because the utility functions and decision strategies used in perception have not been empirically investigated, it remains unclear what decision-making strategy is used, and whether the choice of strategy is uniform across individuals and tasks. In this study, we computationally characterize a decision-making strategy for each individual participant in an auditory-visual spatial localization task, where participants need to make implicit inferences about whether or not the auditory and visual stimuli were caused by the same or independent objects. Our results suggest that a) there is variability across individuals in decision strategy, and b) the majority of participants appear to adopt a probability matching strategy that chooses a value according to the inferred probability of that value. These results are surprising, because perception is believed to be highly optimized by evolution, and the probability matching strategy is considered “suboptimal” under the commonly assumed utility functions. However, we note that this strategy is preferred (or may be even optimal) under utility functions that value learning.
| Human performance in perceptual tasks is often benchmarked by optimal strategies. An optimal strategy is one that performs best with respect to its objectives or maximizes expected reward or equivalently, minimizes a cost function [1], [2]. Previous studies have investigated whether performance in perceptual tasks is consistent with normative models that use maximum likelihood estimation (MLE) [3]–[5], Bayesian inference [6]–[8], signal detection theory [9]–[11], or other computational frameworks. These previous studies either implicitly or explicitly assume a certain cost/utility function that defines the optimal decision. In contrast, the question of which utility/cost function is used by the nervous system for perceptual tasks has not been systematically investigated [but see 12], [13].
In this study we aim to computationally characterize human perceptual decision making strategies. As different strategies may be used across individuals, we characterize the strategy used by each individual observer instead of modeling the behavior of an “average observer”. We used a spatial localization task, as it is simple and fundamental to perceptual processing. While spatial localization has been studied extensively, it has not been investigated in the context of decision making strategies. In nature, at any given moment, we are typically exposed to both visual and auditory stimuli, and scene perception and analysis requires simultaneous inference about the location of auditory and visual stimuli (as well as other sensory stimuli such as tactile, and olfactory). Therefore, multisensory spatial localization represents a task that the perceptual system is implicitly engaged in at all times. This task is particularly useful for probing decision-making strategies because it involves an automatic causal inference about the sources of stimuli, and distinct patterns of behavior corresponding to different strategies. For each observer we examined which of three plausible decision making strategies best accounts for their performance. We use a Bayesian causal inference model of multisensory perception [14] to quantify subjects' responses as one of three strategies as well as compare them to qualitative predictions of such strategies.
One strategy tested was the objective of minimizing the mean squared error. This is a commonly used loss function in normative models of human behavior [3], [4], [7], [15], [16]. It assumes that the nervous system tries to minimize the squared error on average. This utility function in the context of our task implies model averaging, i.e., weighted averaging of the estimate derived from two different causal structures [14]: a common cause hypothesis and an independent causal hypothesis, each weighted by their respective probability (see Figure 1c).
Another strategy we tested was to minimize the error in the inferred causal structure as well as the error in the spatial estimate. This strategy in the context of our task translates into model selection [17], [18]. This strategy also maximizes the consistency in the inference process [13]. In our task, model selection maximizes consistency between the causal structure chosen and the estimate of location. In this strategy, the estimate of location is purely based on the causal structure that is deemed to be most likely (see Figure 1d).
The third strategy tested is probability matching [19]–[22]. This strategy has been reported to be used by humans in a variety of cognitive tasks. In these tasks, probability matching refers to the phenomenon in which observer's probability of a given response matches the probability of appearance of the given target. For example, if the task is to predict which one of two colored lights will be presented in each trial, in an experiment in which each color is presented with a certain probability, then the participant's frequency of predicting each color will be consistent with the relative frequency of the presentation of the color. For a situation where a green light is presented 70% of the time, and a blue light 30% of the time, probability matching behavior would predict the green light on approximately 70% of trials. This strategy is considered to be sub-optimal in terms of economic and utility theory because once it is known that the green light is presented more often, observers should predict the green light on all trials to maximize their utility or gain (.70 proportion correct vs. .70×.70+.30×.30 = .58 proportion correct). Therefore, probability matching has not received much attention in the study of perception—which is generally considered to be highly optimized by evolution [but see 23]–[25 for evidence in visual selective attention]. Nonetheless, because of its implication in the decision making literature, we included this strategy in our analysis. In our task, this strategy would translate into choosing a causal structure according to the probability of the underlying causal structure. Thus, this strategy is one step removed from matching the probability of response outcomes but rather matches the probability of the implicit causal structure (see Figure 1e).
This study was conducted according to the principles expressed in the Declaration of Helsinki. All participants in the experiment provided written informed consent in approval with the UCLA Institutional Review Board.
One hundred and forty six subjects participated in the experiment. We used a large sample because we wanted to be able to detect even small subpopulations (e.g., a small percentage of observers) who may adopt a different strategy. Participants sat at a desk in a dimly lit room with their chins positioned on a chin-rest 52 cm from a projection screen. The screen was a black acoustically transparent cloth subtending much of the visual field (134° width×60° height). Behind the screen were 5 free-field speakers (5×8 cm, extended range paper cone), positioned along azimuth 6.5° apart, 7° below fixation. The middle speaker was positioned below the fixation point, and two speakers were position to the right and two to the left of the fixation. The visual stimuli were presented overhead from a ceiling mounted projector set to a resolution of 1280×1024 pixels with a refresh rate of 75 Hz.
The visual stimulus was a white-noise disk (.41 cd/m2) with a Gaussian envelope of 1.5° FWHM, presented 7° below the fixation point on a black background (.07 cd/m2), for 35 ms. The visual stimuli were always presented so that their location overlapped the center of one of the five speakers behind the screen positioned at −13°, −6.5°, 0°, 6.5° 13°. Auditory stimuli were ramped white noise bursts of 35 ms measuring 69 dB(A) sound pressure level at a distance of 52 cm. The speaker locations were unknown to the participants.
In order to familiarize participants with the task, each session started with a practice period of 10 randomly interleaved trials in which only an auditory stimulus was presented at a variable location, and subjects were asked to report the location of the auditory stimulus.
Practice was followed by 525 test trials that took about 45 minutes to complete. 15 repetitions of 35 stimulus conditions were presented in pseudorandom order. The stimulus conditions included 5 unisensory auditory locations, 5 unisensory visual locations, and all 25 combinations of auditory and visual locations (bisensory conditions). On bisensory trials, subjects were asked to report both the location of auditory stimulus and the location of visual stimulus in sequential order. The order of these two responses was consistent throughout the session, and was counter-balanced across subjects. Subjects were told that “the sound and light could come from the same location, or they could come from different locations.” As a reminder, a blue ‘S’ or green ‘L’ was placed inside the cursor to remind subjects to respond to the sound or light respectively.
Each trial started with fixation cross, followed after 750–1100 ms by the presentation of the stimuli. After 450 ms, fixation was removed and a cursor appeared on the screen vertically just above the horizontal line where the stimuli were presented and at a random horizontal location in order to minimize response bias. The cursor was controlled by a trackball mouse placed in front of the subject, and could only be moved in the horizontal direction. Participants were instructed to “move the cursor as quickly and accurately as possible to the exact location of the stimulus and click the mouse”. This enabled the capture of continuous responses with a resolution of 0.1 degree/pixel.
We used a Bayesian causal inference model of multisensory perception augmented with one of the three decision strategies described above to classify the decision making strategy used by each participant. Details of the model can be found elsewhere [14], but in summary, each stimulus or event, , in the world causes an underlying noisy sensory estimate, , of the event (where i is indexed over sensory channels). Similar to [14], the sensory estimate for our task is the perceived location of the auditory and visual stimuli. We use a generative model to simulate experimental trials and subject responses by performing 10,000 Monte Carlo simulations per condition. Each individual sensation is modeled using the likelihood function . Trial-to-trial variability is introduced by sampling the likelihood from a normal distribution around the true sensory location, analogous to having auditory and visual sensory channels corrupted by independent Gaussian noise with parameters σA and σV respectively. We assume there is a prior bias for the central location, as modeled by a Gaussian distribution centered at 0°. The strength of this bias, σP, is a free parameter. The causal inference model infers the underlying causal structure, C, of the environment based on the available sensory evidence and prior knowledge using Bayes' rule shown in Equation 1.(1)Figure 1 shows a schematic example for a bimodal stimulus presentation. The competing causal structures are shown in Figure 1-B, where either the sensations could arise from a common cause (C = 1, Figure 1-B left), or from independent causes (C = 2, Figure 1-B right). The optimal estimates for the visual and auditory locations are given in Equation 2 for the common cause model, and Equation 3 for the independent model.(2)(3)The difference in our modeling compared to [14] is in producing the final spatial location estimates. The goal of the nervous system is to come up with the best estimates of stimulus locations, and . If the objective is to minimize mean squared error of the spatial estimates, then the optimal estimates are obtained by model averaging:(4)where is the optimal estimate of auditory location given there is a single cause (Eq. 2), and is the optimal estimate of auditory location given there are independent causes (Eq. 3) (see Figure 1-A). The final estimate is a weighted average of the two estimates each weighted by the posterior probability of the respective causal structure (Figure 1-C). is computed likewise.
In model selection strategy (Figure 1-D), on each trial, the location estimate is based purely on the causal structure that is more probable given the sensory evidence and prior bias about the two causal structures:(5)For probability matching (Figure 1-E), location estimates are based on selecting a causal structure based on the inferred posterior probability of the structure. In other words, the selection criterion is stochastic and no longer deterministic. This can be achieved by using a variable selection criteria, , that is sampled from a uniform distribution on each trial.(6)All three models described above have four free parameters: the standard deviation of the auditory and visual likelihoods σA and σV, the standard deviation of the prior over space, σP, and the prior probability of a common cause, p(C = 1) = pcommon. We fit subject responses to the causal inference model for each of the three strategies and determine the best strategy based on the maximum likelihood fit for each subject (see Supplementary Text S1 for a detailed description of the fitting procedure).
The three decision strategies produce distinct patterns of responses across trials and stimulus conditions. Figure 2 shows response distributions for each of the three strategies generated by Monte Carlo simulations for a few stimulus conditions. For these simulations, we used parameter values that are typically found when fitting human observers data. Because vision has a much higher precision in this task than hearing, visual estimates are not affected much by sound. Therefore, we focus our attention on the auditory responses shown in blue. In general, the model averaging strategy mostly has unimodal response distributions, and in conditions with moderate conflict between the visual and auditory stimuli, the auditory responses are shifted in the direction of the visual stimulus (Figure 2-A). In contrast, for the model selection strategy, the auditory responses are mostly bimodal and consistent with either unisensory auditory responses, or complete fusion of the stimuli (Figure 2-B). The probability mass around each peak varies consistently with the expected probability of each causal structure. In other words, for conditions in which the discrepancy between visual and auditory locations is large, and thus the probability of a common cause is low, there is a large probability mass at the auditory location, and in conditions where the conflict is small, and thus the probability of common cause is high, there is a much larger probability mass around the visual capture location (i.e., location shifted towards visual stimulus). The fixed selection criterion results in distinct separation between the two auditory response distribution modes. For any probability of a common cause greater 0.5, the auditory response will be fused with the visual response. Similarly, the probability matching strategy also shows bimodal auditory response distributions (Figure 2-C). However, in contrast with model selection, the modes are not as distinct. Due to the variable model selection criteria (ξ), even when the probability of a common cause is high, there is a small probability of providing an auditory response consistent with the independent causal structure. Due to the high uncertainty in the auditory signals (i.e., large variance of auditory likelihood), this can even be observed when the stimulus locations are identical (left column).
For each participant, each of the decision strategy models was fitted to the data, and the observer was classified by the strategy that explained the data best. In order to be highly confident in the classifications, for an observer to be included in the sample we required that the log-likelihood difference between the best-fitting model and the second best-fitting model exceed a value of 3—which is considered substantial evidence for the support of one model vs. another [26]. In Table 1, we report the results from 110 participants whose data met this criterion. Among these participants, on average the log-likelihood difference between the top two best-fitting models was 24.6 (median 17.6), which is in the range considered as decisive evidence for a model relative to another model [26]. On average, the best fitting model accounted for 83% of the variance in the individual participant's data (generalized coefficient of determination [27]: R2 = 0.83±0.11). The model fits for the probability-matching group data is also shown for all stimulus conditions in Supplementary Figure S1. Therefore, the best-fitting model fitted the data very well, and the classifications were highly reliable.
The number of participants classified as utilizing the matching, selection, or averaging strategy is provided in Table 1. Probability matching is the decision strategy used by the vast majority of observers (82/110). Model averaging was second followed by model selection. The proportion of males and females is not significantly different for each strategy (two-sample test for equality of proportions, p>0.05). The difference in distribution of ages among the three groups was also statistically insignificant (two-sample Kolmogorov-Smirnov test, p>0.05). It should be pointed out that these results are not sensitive to the subject exclusion criterion described above. The results remain qualitatively the same even if we do not exclude any participants at all: N = 146, matching = 64%; selection = 18%; averaging = 18%, or if we use other exclusion criteria (e.g., margin of 10 instead of 3: N = 82, matching = 79%; selection = 5%; averaging = 16%).
We also tested whether the model selection strategy could explain the data better than the other two strategies if we allow a bias in choosing a model (i.e., if the criterion can take on any value as a free parameter, rather than fixed at .5 as in Equation 5). Despite the additional free parameter for this model, we find similar proportions of categorization: N = 110, matching = 72%; selection = 13%; averaging = 15% – and after applying Bayesian Information Criteria regularization for the additional free parameter: matching = 72%; selection = 11%; averaging = 17%.
We aimed to gain insight into the decision making strategy used in a perceptual task, by comparing three strategies and testing which one accounts best for the observers' data. Our computational modeling tools allow us to perform this type of analysis for each individual observer. Perceptual functions, in particular the basic ones that are shared across species (and arguably key to the survival of the organism) such as spatial localization, are often thought to be optimal. Perceptual functions are evolutionarily old and thus, it is argued that there has been sufficient amount of time for them to have been optimized by evolution [28], and indeed several studies have shown a variety of perceptual tasks to be “statistically optimal.” For the same reason, it is also expected that the optimized and automated perceptual processes to be largely uniform across individuals.
We examined the decision strategies in an auditory-visual spatial localization task on a large sample of observers consisting of 110 individuals. First, we found that not all observers appear to utilize the same strategy. This variability across individuals suggests that this localization process is not predestined or hard-wired in the nervous system. More importantly, the vast majority of participants seem to use a probability matching strategy. This finding is surprising because this strategy is not statistically optimal in the conventional sense.
Why should the majority of individuals use a “suboptimal” strategy in this basic task? To address this question, it is best to step back and re-examine the notion of optimality. While a probability matching strategy may not be optimal in a static environment, it may be optimal or close to optimal in a dynamic one [29], and especially useful in exploring patterns in the environment. Humans instinctively have the tendency to search for regularities in random patterns [30]–[33], and it has been suggested that probability matching results from the addition of an “informatic” utility that considers learning and exploring an important component in survival and ecological rationality [34]. Thus, while probability-matchers might look irrational in the absence of predictable patterns, they would have a higher chance of finding patterns if they exist [19]. In the context of our experiment, although the stimuli were uniformly random, perhaps the matchers subconsciously explore for patterns within the stimuli.
The observed probability matching behavior suggests that the nervous system samples from a distribution over model hypotheses on each trial. Sampling-based representational coding has been proposed to account for neurophysiological phenomena such as spontaneous neural activity [35] and variability in neural responses [36], as well as other stochastic perceptual phenomena such as bistability [37], [38]. Alternatively, it is conceivable that a case-based selection rule [39] that, on each trial, chooses the most appropriate model from an earlier experience (not necessarily from the current experiment) resembling the current sensations, would produce this behavior.
While probability matching was the modal response strategy found in the current study, we are not claiming that probability matching is used in all perceptual tasks, or even in all spatial tasks. Optimal performance in perceptual tasks has been reported by some previous studies. A recent study found observers' behavior to be consistent with the expected loss function in a visual discrimination task [40], however, the results are ambiguous with respect to the specific decision making strategy utilized (averaging, selection, and probability matching) as they would make similar predictions. Knill, in a study of perception of slant from compression and binocular disparity cues [41], reported optimal performance. In this study, which used an almost identical structure inference model to the one used here, observers' responses were explained well by model averaging. However, probability matching was not considered, and regarding model selection vs. averaging, the author points out that determining exactly which strategy was used by the participants is difficult. Perhaps most relevant to our current findings is a previous study of auditory-visual spatial localization in which the observers' performance was found to be consistent with model averaging [14]. Although model selection and probability matching were not tested, the response profiles were unimodal and thus, inconsistent with these strategies. The sample size was relatively small in this study (N = 19), yet together with the aforementioned studies, these findings raise the question of what are the factors that influence the decision-making strategy adopted by observers. It is likely that the specific strategy used by participants depends on the context, instruction, prior experience, and many other factors [42]. Landy et al. [43] found that stimulus variability and unpredictability from trial to trial can result in adoption of a variety of suboptimal strategies by participants in a texture orientation perception task. Even for a given context, stimuli, and instruction (as in this experiment), some subjects' construal of the task may affect their utility/cost function. The specific computational constraints such as criteria of speed and accuracy could also favor the use of one strategy over another. Also in our study, subjects had to make sequential reports of both modalities requiring responses to be held in working memory, which has been suggested to have a role in the decision process [19], [32], [44]. The specific factors influencing perceptual decision making strategies is an open question for future studies.
Probability matching has been shown to happen when people's response probability matches the relative frequency of the presented stimuli. Here we show that the nervous system can even match the probability of a more abstract construct such as the probability of causal structure of the stimuli which is one step removed from the stimuli themselves. This finding suggests that probability matching may be a general decision-making strategy operating at multiple levels of processing. The results of this study altogether suggest that the nervous system does not necessarily use the commonly assumed least squared error cost function in perceptual tasks, and underscore the importance of considering alterative objectives when evaluating perceptual performance.
|
10.1371/journal.ppat.1007445 | Whole genome screen reveals a novel relationship between Wolbachia levels and Drosophila host translation | Wolbachia is an intracellular bacterium that infects a remarkable range of insect hosts. Insects such as mosquitos act as vectors for many devastating human viruses such as Dengue, West Nile, and Zika. Remarkably, Wolbachia infection provides insect hosts with resistance to many arboviruses thereby rendering the insects ineffective as vectors. To utilize Wolbachia effectively as a tool against vector-borne viruses a better understanding of the host-Wolbachia relationship is needed. To investigate Wolbachia-insect interactions we used the Wolbachia/Drosophila model that provides a genetically tractable system for studying host-pathogen interactions. We coupled genome-wide RNAi screening with a novel high-throughput fluorescence in situ hybridization (FISH) assay to detect changes in Wolbachia levels in a Wolbachia-infected Drosophila cell line JW18. 1117 genes altered Wolbachia levels when knocked down by RNAi of which 329 genes increased and 788 genes decreased the level of Wolbachia. Validation of hits included in depth secondary screening using in vitro RNAi, Drosophila mutants, and Wolbachia-detection by DNA qPCR. A diverse set of host gene networks was identified to regulate Wolbachia levels and unexpectedly revealed that perturbations of host translation components such as the ribosome and translation initiation factors results in increased Wolbachia levels both in vitro using RNAi and in vivo using mutants and a chemical-based translation inhibition assay. This work provides evidence for Wolbachia-host translation interaction and strengthens our general understanding of the Wolbachia-host intracellular relationship.
| Insects such as mosquitos act as vectors to spread devastating human diseases such as Dengue, West Nile, and Zika. It is critical to develop control strategies to prevent the transmission of these diseases to human populations. A novel strategy takes advantage of an endosymbiotic bacterium Wolbachia pipientis. The presence of this bacterium in insect vectors prevents successful transmission of RNA viruses. The degree to which viruses are blocked by Wolbachia is dependent on the levels of the bacteria present in the host such that higher Wolbachia levels induce a stronger antiviral effect. In order to use Wolbachia as a tool against vector-borne virus transmission a better understanding of host influences on Wolbachia levels is needed. Here we performed a genome-wide RNAi screen in a model host system Drosophila melanogaster infected with Wolbachia to identify host systems that affect Wolbachia levels. We found that host translation can influence Wolbachia levels in the host.
| Insects are common vectors for devastating human viruses such as Zika, Yellow Fever, and Dengue. A novel preventative strategy has emerged to combat vector-borne diseases that exploits the consequences of vector-insect infection with the bacteria Wolbachia pipientis [1–4]. Wolbachia is a vertically transmitted, gram-negative intracellular bacterium known to infect 40–70% of all insects [5, 6]. Wolbachia provides hosts with resistance to pathogens such as viruses [7–10]. Remarkably, Wolbachia infections can reduce host viral load enough to render insect hosts incapable of transmitting disease-causing viruses effectively [1, 2, 11–24]. The relationship between Wolbachia and a host is complex and dynamic. Understanding how bacterial levels can change is vital because it dictates how Wolbachia manipulates the host insect. For example, the antiviral protection provided by Wolbachia is strongest when Wolbachia levels within a host are high [10, 25–27]. On the other hand, Wolbachia can become deleterious to the host when Wolbachia population levels are too high leading to cellular damage and reduced lifespan[28–30]. To apply Wolbachia as an effective tool to combat vector-borne viruses we need a better understanding of host influences on Wolbachia levels.
Wolbachia infects a large host and tissue range suggesting interaction with various host systems and pathways for successful intracellular maintenance within a host [5, 31]. To date, reports suggest that Wolbachia levels may be influenced in various contexts by interaction with host cytoskeletal components [32–35], the host ubiquitin/proteasome [36], host autophagy [37], and by host miRNAs [16, 38]. A comprehensive analysis of host systems that influence Wolbachia levels has not been carried out and will further our knowledge of this symbiotic relationship and reveal molecular mechanisms that occur between Wolbachia and the host to maintain it.
Wolbachia-host interactions can be studied in the genetically tractable Drosophila melanogaster system which allows for the systematic dissection of host signaling pathways that interact with the bacteria using the wide array of genetic and genomic tools available. The Drosophila system enables rapid unbiased screening of host factors that impact Wolbachia at the cellular and organismal level. While some influences on the relationship, such as systemic effects, require studies in the whole organism, many aspects of molecular and cellular signaling can be studied in a Drosophila cell culture-based system. Drosophila cells are particularly amenable to genome-scale screens because of the ease and efficacy of RNAi in this system [39]-[40]. Previous cell culture-based RNAi screening has been a successful approach to study a wide range of intracellular bacteria-host interactions in Drosophila cell lines [41–44]. Thus, we reasoned that this was a feasible approach for studying Wolbachia-host interactions.
Here we performed a whole genome RNAi screen in a Wolbachia-infected Drosophila cell line, JW18, which was originally derived from Wolbachia-infected Drosophila embryos and has previously proven suitable for high-throughput assays [36, 45]. The goal was to determine in an unbiased and comprehensive manner which host systems affect intracellular Wolbachia levels. The primary screen identified 1117 host genes that robustly altered Wolbachia levels. Knock down of 329 of these genes resulted in increased Wolbachia levels whereas 788 genes led to decreased Wolbachia levels. To characterize these genes, we generated manually curated categories, performed Gene Ontology enrichment analysis, and identified enriched host networks using bioinformatic analysis tools. The effects on Wolbachia levels were validated in follow-up RNAi assays that confirmed Wolbachia changes visually by RNA FISH as well as quantitatively using a highly sensitive DNA qPCR assay. We uncovered an unexpected role of host translation components such as the ribosome and translation initiation factors in suppressing Wolbachia levels both in tissue culture using RNAi and in the fly using mutants and a chemical-based translation inhibition assay. Furthermore, we show a decrease in overall translation in Wolbachia-infected JW18 cells compared to Wolbachia-free JW18DOX cells and that an inverse trend exists between Wolbachia levels and host translation levels in JW18 cells. This work provides strong evidence for a relationship between Wolbachia and host translation and strengthens our general understanding of the Wolbachia-host intracellular relationship.
Wolbachia is an intracellular bacterium that resides within a wide range of insect hosts. To identify host factors that enhance or suppress intracellular Wolbachia levels, we performed a genome-wide RNAi screen in Wolbachia-infected JW18 Drosophila cells that were originally derived from Wolbachia-infected embryos [45]. In order to visually detect Wolbachia levels we established a specific and sensitive RNA Fluorescence In Situ Hybridization (FISH) method consisting of a set of 48 fluorescently labeled DNA oligos that collectively bind in series to the target Wolbachia 23s rRNA (Fig 1A). This enabled detection of infection levels ranging from as low as a single bacterium in a cell to a highly infected cell and could clearly distinguish Wolbachia-infected cells from Wolbachia-free cells (Fig 1B). Thus, we were able to assess Wolbachia infection levels in the JW18 cell population and found that under our culturing conditions we could stably maintain JW18 cells with a Wolbachia infection level of 14% of the JW18 cells (Fig 1C). Of the infected cells, 73% of the cells had a low Wolbachia infection (1–10 bacteria), 13.5% had a medium infection (11–30 bacteria), and 13.5% were highly infected (>30 bacteria). Though Wolbachia levels may change in different culturing conditions, the JW18 cell line maintained Wolbachia levels stably for the duration of the screen. These experiments confirmed the feasibility and sensitivity of RNA FISH to detect different levels of Wolbachia infection in Drosophila cells in a highly sensitive manner.
Prior to screening we characterized the JW18 cell line and its associated Wolbachia strain by generating a JW18 DNA library and sequencing it using DNAseq technology (S1 Fig). This allowed for phylogenetic analysis of the Wolbachia strain and revealed that it clustered most closely with the avirulent wMel strain which is well characterized for its antiviral effect on RNA-based viruses in Drosophila as well as in mosquitos (S1A Fig) [1, 2, 7, 8]. Further analysis included gene copy number variation of the Wolbachia genome and identified one deleted and one highly duplicated region (3–4 fold increased) (S1B Fig). The deleted region contained eight genes known as the “Octomom” region postulated to influence virulence [27, 46]. The loss of “Octomom” has also been reported in wMelPop-infected mosquito cell lines after extended passaging over 44 months [47]. This suggests that loss of this region happened independently in two cases and may be related to passage in cell culture. A highly duplicated region spans approximately from positions 91,800–127,100 and contains 38 full or partial genes, including those with unknown function as well as genes predicted to be involved in metabolite synthesis and transport, molecular chaperones, DNA polymerase III subunit, DNA gyrase subunit, and 50S ribosomal proteins. For analysis of gene copy number variation in the JW18 cell line, the DNA library was aligned to the Release 6 reference genome of D. melanogaster. This revealed that the cell line is of male origin with an X:A chromosomal ratio of 1:2 and tetraploid in copy number (S1C Fig). Bioinformatic analysis on genes of high or low copy number did not reveal an enrichment for any particular molecular or cellular functional class and the majority (72%) of genes in the JW18 cell line were at copy numbers expected for a tetraploid male genotype (4 copies on autosomes, 2 copies on X). This made the JW18 cell line suitable for RNAi screening.
As a first step to uncovering Wolbachia-host interactions, we asked whether gene expression changes occur in the host during stable Wolbachia infection. To do this, we used a control Wolbachia-free version of the JW18 cell line which was previously generated through doxycycline treatment (JW18DOX) (S2A Fig). A comparison of host gene expression changes in the presence and absence of Wolbachia through RNAseq analysis revealed 308 and 559 host genes that were up- or down-regulated respectively by two-fold or more (padj<0.05) (S2B Fig). Of these genes, 21 displayed major expression changes of log2 fold >4 (DptB, Wnt2, SP1173, bi, FASN3, CG5758, CP7Fb, beta-Man) or log2-fold <-4 (CG12693, CG13741, Tsp74F, esn, cac, CG4676, CG42827, CG18088, CG17839, 5-HT2A, CG43740, CG3036, aru) (also see S2C Fig). The presence of Wolbachia led to elevated gene expression of several components of the host immune response including the Gram-negative antimicrobial peptide Diptericin B (DptB), which was the most highly upregulated gene in the presence of Wolbachia (S2C Fig). Gene ontology (GO) analysis further confirmed a host immune response with enriched terms such as ‘response to other organism’ and ‘peptidoglycan binding’ that included genes for antimicrobial peptides (attA, AttB, AttC, DptB, LysB) and peptidoglycan receptors (PGRP-SA, -SD, -LB, -LF) as well as antioxidants such as Jafrac2, Prx2540-1, Prx2540-2, Pxn, GstS1 with ‘peroxiredoxin’ and ‘peroxidase activity’. Other expression changes included extracellular matrix components such as upregulation of collagen type IV (Col4a1 and vkg) and downregulation of genes for integral components of the plasma membrane including cell adhesion components (kek5, mew, Integrin, and tetraspanin 42Ed and 39D). Gene ontology analysis further identified a significant enrichment of ion transporters and channels that were downregulated as well as genes encoding several proteins such as myosin II, projectin and others associated with the muscle Z-disc that were downregulated. Finally, we observed an overall upregulation of host proteasome components at the RNA level in the presence of Wolbachia (S3 Fig), which is in line with proteomics data of proteasome upregulation in the presence of Wolbachia [48, 49]. In summary, these host factors may play an important role in the Wolbachia-host relationship however their specific roles in this interaction remain to be determined.
The screening approach combined the visual RNA FISH Wolbachia detection assay (Fig 1) with in vitro RNAi knockdown of host genes to ask which host genes influence Wolbachia levels (Fig 2A). Prior to screening, we tested whether RNAi was a feasible approach in JW18 cells. First, we confirmed that RNAi had no adverse effects such as cytotoxicity on the cells using a negative control dsRNA targeting LacZ which was not present in our system (Fig 2B and 2C). Second, we tested RNAi knockdown efficiency in the JW18 cell line. To do this a Jupiter-GFP transgene present in the cell line was targeted for knockdown using dsRNA to GFP. High knockdown efficiency was achieved using this RNAi protocol as seen by the efficient knockdown of the Jupiter-GFP transgene both visually by RNA FISH (90.2% reduction) (Fig 2D and 2E) and by protein levels as shown by Western blot (97.9% reduction) (Fig 2F) compared with either the ‘no dsRNA’ knockdown (Fig 2B) or ‘LacZ’ knockdown (Fig 2C) conditions. This confirmed the suitability of the JW18 cell line for an RNAi-based screening approach. For controls that alter Wolbachia levels, we identified a host ribosomal gene, RpL40, from a pilot screen that consistently led to increased Wolbachia levels when depleted by RNAi (Fig 2G) compared to cells that were not treated by RNAi (Fig 2B) or treated with lacZ dsRNA treatment (Fig 2C). We achieved 96.3% RNAi knockdown efficiency as confirmed by qPCR for RpL40 levels relative to a no knockdown control (Fig 2H). At the time of the screen we did not know of any host protein whose knockdown would decrease Wolbachia levels. Therefore, as a Wolbachia-decreasing control, cells were incubated with 5μM doxycycline for 5 days which successfully reduced the Wolbachia levels in the cells by 91.9% as measured by RNA FISH (Fig 2I and 2J). To quantify the effect of the controls on Wolbachia levels we isolated genomic DNA from each treated sample and used quantitative PCR DNA amplification to detect the number of Wolbachia genomes per cell by measuring Wolbachia wspB copy number relative to the Drosophila gene RpL11 (Fig 2K). Relative to control cells, the RNAi treatment with RpL40 resulted in a 3.4-fold increase in Wolbachia, doxycycline decreased Wolbachia levels 6.3-fold, whereas LacZ and GFP RNAi had no significant effect confirming that our controls allowed us to manipulate Wolbachia levels in the JW18 cell line and that this cell line with its relative low infection rate (Fig 1C) provided a sensitive tool for detecting dynamic changes in Wolbachia levels through an RNAi screening approach.
The layout of the whole genome screen is illustrated in Fig 2A. Briefly, Wolbachia-infected JW18 cells were incubated with the DRSC Drosophila Whole Genome RNAi Library version 2.0 which was pre-arrayed in 384 well tissue culture plates such that each well contained a specific dsRNA amplicon to target one host gene. The 5-day incubation period allowed for efficient host gene knockdown. Thereafter the cells were processed for RNA FISH detection of Wolbachia 23s rRNA. Total fluorescence signal was detected using automated microscopy and served as a readout for Wolbachia levels within each plate well. Host cells within each well were detected by DAPI staining. Finally, the Wolbachia fluorescence signal was divided by the total number of DAPI-stained host cells detected to provide an average Wolbachia per cell readout which was normalized to the plate average (represented as a robust Z score). The library was screened in triplicate. The raw screening data were subjected to several quality control steps (S4 Fig). Briefly, we realigned the DRSC Version 2.0 Whole Genome RNAi library dsRNA amplicons with Release 6 of the D. melanogaster genome using the bioinformatic tool UP-TORR [50]. This provided an accurate updated description of the gene target for each dsRNA amplicon. Initially the library included 24 036 unique dsRNA amplicons targeting 15 589 genes, however owing to updates in gene organization and annotation models of the reference genome since the initial release of the library we removed 1499 outdated amplicons from our subsequent analysis as they were no longer predicted to have gene targets (S1 Table). We also excluded 1481 amplicons that were annotated in UP-TORR to target multiple genes (S2 and S3 Tables). We further excluded 66 amplicons for a positional effect on the dsRNA library tissue culture plates at the A1 position (S5 Fig, S4 Table). Thus, we effectively screened 20 990 unique dsRNA amplicons targeting 14 024 genes (80% of D. melanogaster Release 6 genome). A further quality control step to reduce false positive hits was to cross-reference potential hits with RNAseq gene expression data for the JW18 cell line to exclude genes with undetectable expression in the cell line (S5 Table).
To identify and select for hits from the primary data, we first analyzed the screen-wide controls. A plot of all controls included in the whole genome screen revealed that RpL40 knockdown increased Wolbachia levels (median robust Z score of 2.2), conversely doxycycline treatment decreased Wolbachia levels throughout the screen (median robust Z of -3.5), whereas a standard control included in the whole genome library, Rho1, and GFP RNAi knockdown did not significantly affect Wolbachia levels (Fig 3A). We used this range as a guide to set robust Z limits for primary hits at ≥ 1.5 or ≤ -1.5. Every dsRNA amplicon was screened in triplicate. To be considered as a ‘hit’ amplicon at least 2 of the 3 replicates needed to satisfy the robust Z score limits (S4 Fig). To categorize the primary screen hits, each gene was assigned to a ‘High’, ‘Medium’, and “Low’ bin based on the confidence level (S4 Fig). This was determined based on the total number of different dsRNA amplicons representing a hit gene in the library and how many of these dsRNA amplicons had a significant effect on Wolbachia levels (S4 Fig). In this manner, we were able to stratify the primary screen hits to assist in follow up analysis.
The screen identified 1117 genes that when knocked down had a significant effect on the Wolbachia levels in JW18 cells (S6 Table). Knock down of 329 of the 1117 genes resulted in increased Wolbachia levels, suggesting that these genes normally restrict Wolbachia levels within the host cell (Fig 3B). Knockdown of 788 genes resulted in decreased Wolbachia levels, suggesting Wolbachia may be dependent on these host genes for survival within the host cell (Fig 3B). For each of the two hit categories, genes were classified by confidence level (described in S4 Fig, and Fig 3C). We found a higher proportion of low confidence hits (21%) in the category of genes that decreased Wolbachia levels compared to genes that led to Wolbachia level increases which only contained 12.5% low confidence hits. To analyze the expression of the 1117 genes, the hits were distributed into 9 bins based on their gene expression level from JW18 RNAseq data (S6A Fig). Hits displayed a wide range of expression and an enrichment of low expression for hits that decreased Wolbachia levels (S6B Fig). We did not observe any biases for variation in gene DNA copy number based on DNAseq data for the JW18 cell line (S6C Fig).
Next, we asked whether changes in Wolbachia levels could be explained by effects on host cell proliferation or were independent of effects on host cell proliferation. We measured cell proliferation using the raw screen data by normalizing the number of cells scanned per well (DAPI) to the number of fields of view required to capture the cells. This allowed us to generate a robust Z score measure of cell proliferation effects for the 1117 genes identified as hits. For genes that increased Wolbachia levels, 12% (41 genes) increased cell proliferation (robust Z>1), 45% (147 genes) decreased cell proliferation (robust Z<-1), and 43% (141 genes) had no effect on cell proliferation (Fig 3D, S7 Table). These data suggest that a significant number of gene knockdowns (45%) may indirectly lead to an increase in Wolbachia levels through slowed cell proliferation. Importantly, 43% of hits identified had no effect on cell proliferation whilst increasing Wolbachia levels. These results suggest that changes in Wolbachia levels are not strictly linked to host cell proliferation. For genes that decreased Wolbachia levels, the majority (82%, 644 genes) did not affect cell proliferation and 2% (19 genes) increased and 16% (125 genes) decreased cell proliferation (Fig 3D, S7 Table). To summarize, the screen identified 1117 host genes that act to support or suppress Wolbachia levels within the host Drosophila cell.
To classify the 1117 gene hits identified in the whole genome screen, we first manually curated the hits using gene annotation available on FlyBase (http://www.flybase.org) relating to each gene such as gene family, domains, molecular function, gene ontology (GO) information, gene summaries, interactions and pathways, orthologs, and related recent research papers. We identified distinct categories of genes that when knocked down by RNAi increased (Fig 4A) or decreased (Fig 4B) Wolbachia levels. The largest gene category that led to decreased Wolbachia levels by RNAi knockdown contained genes for host metabolism and transporters suggesting that Wolbachia strongly relies on this aspect of the host (Fig 4B). On the other hand, gene knockdowns that increased Wolbachia contained many components of the core ribosome network, translation factors, and the proteasome core and regulatory proteins network (Fig 4A). Six of the broad gene categories could be further sub-classified for processes that either enhanced or suppressed Wolbachia levels. First, RNAi knockdown of members in the category containing cytoskeleton, cell adhesion and extracellular matrix components decreased Wolbachia, these included cadherins, formins, spectrin and genes involved in microtubule organization, whereas knockdowns that resulted in increased Wolbachia were actin and tubulin-related. Second, Wolbachia levels may be sensitive to disturbances in membrane dynamics and trafficking. Specifically, knockdown of SNARE components, endosomal, lysosomal and ESCRT components decreased Wolbachia, whereas knockdown of components of COPI, endosome recycling, and several SNAP receptors increased Wolbachia levels. Third, disruptions in several cell cycle-related components decreased Wolbachia levels, while Wolbachia levels increased upon disruption of cytokinesis, the separase complex and the Anaphase Promoting Complex. Fourth, the knockdown of components related to RNA helicases and the exon junction complex decreased Wolbachia, while disruption of many spliceosome components increased Wolbachia. Fifth, epigenetic changes influenced Wolbachia levels: knockdown of members involved in heterochromatin silencing, Sin3 complex and coREST decreased Wolbachia levels, whereas knocking down members of the BRAHMA complex resulted in increased Wolbachia levels. Finally, Wolbachia levels were sensitive to changes in host transcription. We observed that disruption of components in the mediator complex and regulators of transcription from Polymerase II promoters decreased Wolbachia, whereas knockdown of the BRD4pTEFb complex involved in transcriptional pausing and other transcriptional elongation factors resulted in increased Wolbachia levels. Together, this manual curation revealed that this whole genome screen yielded host genes that suppress or enhance Wolbachia levels and that these primary hits could be classified into distinct gene categories.
Further GO term enrichment analysis using the online tool Panther (http://www.pantherdb.org/) suggested that the 329 genes resulting in Wolbachia increases formed a robust dataset as many of the enriched terms overlapped with our manual curation (S7 Fig). In contrast, there was a lack of enrichment for the 788 Wolbachia-decreasing genes even though manual curation had sorted many of these genes into categories. For this reason, further analysis focused on the 329 host genes that increased Wolbachia when knocked down by RNAi.
To assess whether specific host networks were enriched within the 329 host genes identified as potential suppressors of Wolbachia we used two bioinformatic tools namely the Kyoto Encyclopedia of Genes and Genomes (KEGG), and the protein complex enrichment analysis tool (COMPLEAT) with criteria for a network restricted to complexes with 3 or more components (p<0.05) [51]. This analysis revealed enrichment of several host networks among the 329 genes whose knockdown resulted in Wolbachia increases including a striking 67.5% of the core cytoplasmic ribosome (56/83 expressed ribosomal proteins) and 31.3% of all translation initiation components (10/32 expressed proteins) as well as 70.1% the core proteasome (24/34 expressed proteins) (Fig 5A, and S8 Fig). These findings strongly suggested that perturbations in host translation components could alter Wolbachia levels. For both networks, the majority of components did not significantly affect cell proliferation within the duration of the RNAi screen assay (circles), though some did have a negative impact (robust Z<-1) (square) (Fig 5A, see Fig 3D). Importantly, these data show that Wolbachia level fluctuations are independent of host cell proliferation changes because Wolbachia levels increased in RNAi knockdowns of network components regardless of the presence or absence of cell proliferation changes (Fig 5A, S9 Fig). We chose to validate and characterize the novel Wolbachia-host translation interaction identified in the whole genome RNAi screen.
We validated the influence of the ribosome, translation initiation complex, and proteasome on Wolbachia levels by knocking down representative members of each network using RNAi knockdown in JW18 cells (Fig 5B). Each gene was validated using two different dsRNA amplicons that were designed to target different parts of the gene. Effects on Wolbachia levels were assessed quantitatively by DNA qPCR measuring the number of Wolbachia genomes (wspB DNA copies) relative to the number of host cell nuclei (RpL11 DNA copies). Network validation is represented relative to untreated JW18 control cells (No RNAi) and the positive control RpL40 RNAi knockdown is included for reference. For the translation initiation network, we selected eIF-4a, eIF-2 subunit beta, eIF-3c, eIF-3i, and eIF-3ga. All ribosome components’ RNAi knockdown significantly increased Wolbachia levels by 5-fold or more (Fig 5B). For the ribosomal network, we selected RpL10, RpL36, RpLP1, RpS4, RpLP2, RpS3 and RpS26 for validation and each RNAi knockdown resulted in a significant increase of nearly 10-fold or higher Wolbachia levels relative to untreated JW18 control cells (Fig 5B). We also validated the proteasome network using RNAi knockdown of three selected genes (Rpn11, Rpt2, Rpn2) which resulted in significant Wolbachia increases (S10 Fig). To summarize, we were able to validate that RNAi knockdown of ribosomal, translation initiation, and proteasomal networks leads to striking increases in Wolbachia levels in JW18 cells.
To characterize the changes in Wolbachia levels in the JW18 cell line when ribosome or proteasome (S10 Fig) components are perturbed by RNAi, we visually classified the level of Wolbachia infection in cells using the Wolbachia-detecting 23s rRNA FISH probe combined with DAPI staining and the GFP-Jupiter transgene labelling microtubules to identify the cells. Each cell was classified according to its Wolbachia infection into the following categories: uninfected (no Wolbachia), low (1–10 Wolbachia), medium (11–30 Wolbachia), and high (>30 Wolbachia) infection. Similar to Fig 1C, in a control LacZ knockdown JW18 control population 14% of the total number of cells were infected. In contrast, RNAi knockdown of the ribosome component RpS3 resulted in an overall dramatic increase in the total number of infected cells (73%) (Fig 5C). Of the infected cells in the control population, 73% had a low level of infection whereas 13.5% had a medium level infection and 13.5% had a high level of infection (Fig 5D). A comparison of the extent of infection revealed a 1.6-fold increase in medium and highly infected cells after knockdown of network components compared to the control (Fig 5D). Similar results were obtained in proteasome RNAi knockdown cells showing an increase in Wolbachia-infected cells to 87% (S10 Fig). Together, our results show an increase in the total number of infected cells after ribosomal network knockdown and within this population a relative increase of medium to high infected cells, however the majority (57%) of cells maintained a low level of infection.
Next, we tested whether these networks could influence Wolbachia in the fly (Fig 6 and S10 Fig). In Drosophila, Wolbachia are found abundantly in the ovary. To test the effect of perturbing the ribosome, females from a Wolbachia-infected stock were crossed to available ribosomal mutant alleles for RpL27A and RpS3 at 25°C. Then, the Wolbachia infection level in the ovaries of 5 day-old Wolbachia-infected heterozygous mutant and wild-type siblings were compared by RNA FISH for Wolbachia 23s rRNA. We observed dramatic increases in Wolbachia levels in the ribosomal mutants compared to the control sibling ovaries at early stages of oogenesis in the germarium as well as in maturing egg chambers (Fig 6A and 6C). Quantification of the integrated density of the 23s rRNA Wolbachia FISH probe in Z-stack projections of germaria for ribosomal mutants confirmed a 2.94-fold (RpL27A) and 3-fold (RpS3) increase in the mutant compared to control siblings (Fig 6B) (Non-parametric Mann Whitney, RpL27A and RpS3 p<0.0001). Further, quantification of stage 10 egg chambers revealed a 1.6-fold (RpL27A) and 1.27-fold (RpS3) increase compared to their respective control siblings (Fig 6D) (Non-parametric Mann Whitney, RpL27A p = 0.0002, RpS3 p = 0.0089). The fecundity of both ribosomal mutant lines was assessed by counting eggs laid per female as well as assessing the embryo viability. We found no significant difference between ribosomal mutant and control Wolbachia-infected siblings nor between Wolbachia-infected and uninfected flies, suggesting that the rate of oogenesis and viability of offspring are not affected by reducing the levels of ribosomal proteins nor by the level of Wolbachia infection (S11 Fig). In conclusion, these results demonstrate that Wolbachia levels are sensitive to changes in the host ribosomal network in both early and late stages of Drosophila oogenesis and that under the conditions tested Wolbachia-infection does not impact fecundity of the animals. Similar results were obtained in proteasomal subunit Prosβ6 (DTS5) mutant flies (S10 Fig).
Apart from the Drosophila ovary, we tested the effect of ribosomal and proteasome mutations on Wolbachia levels in other tissues including larval imaginal discs, adult male testes, and in the whole fly. To do this we processed RpS3 mutant and control larval imaginal discs for Wolbachia RNA FISH visualization (S12 Fig). We found significantly increased levels of Wolbachia in haltere-, wing-, leg- discs (2.23-fold (p<0.0001), 1.15-fold (p = 0.0229), and 1.74-fold (p<0.0001) Non-parametric Mann Whitney) respectively. Similar results were obtained in proteasomal mutant (DTS5) flies (S12 Fig) showing increased Wolbachia in haltere-, wing-, and leg- discs (2.55-fold (p<0.0001), 1.91-fold (p = 0.0005), 2.0-fold (p<0.0001) Non-parametric Mann Whitney) as well as in the larval brain (2.26-fold (p = 0.0003). These data suggest that increases in Wolbachia levels in RpS3 and Prosβ6 (DTS5) mutants occur early in development and in a variety of tissue types (S12 Fig). In addition, we found significantly increased Wolbachia levels (2.43-fold (p = 0.0360) in the hub of adult RpL27A mutant testes compared to control siblings as well as a significant 2.8-fold increase in Wolbachia in proteasomal DTS5 mutant testes compared to sibling controls (p = 0.0093) (Non-parametric Mann Whitney) (S13 Fig). Finally, we assessed the Wolbachia level increase in whole flies using DNA qPCR and found increased Wolbachia levels in RpS3 mutants for males (1.22-fold) and females (1.56-fold) and RpL27A mutant females (2.74-fold) compared to control siblings (S14 Fig). Together these data suggest that Wolbachia level increases in ribosomal and proteasomal mutants occur in a wide range of tissue types and is not sex specific.
Next, we asked whether a direct relationship exists between Wolbachia and host translation. To do this we asked whether chemical inhibition of host translation by cycloheximide would alter Wolbachia levels in host Drosophila. Wolbachia-infected D. melanogaster were fed cycloheximide-containing food or control food for 7 days prior to genomic DNA extraction of whole flies. We tested the Wolbachia-levels in individual whole flies using DNA qPCR and found increased Wolbachia levels in flies fed on cycloheximide compared to control flies (Fig 6E). This suggested that Wolbachia levels are sensitive to host translation and that perturbation of host translation leads to increased Wolbachia levels.
Having observed that Wolbachia levels are sensitive to host translation, we wanted to observe the relationship between Wolbachia and host translation levels in an unperturbed manner in Drosophila JW18 cells and in the fly. To correlate levels of host translation with levels of Wolbachia, we combined Wolbachia RNA FISH detection with a visual fluorescent ‘click’ chemistry-based method to assess global protein synthesis levels in host cells (Fig 7). This assay is based on a sensitive, non-radioactive method that utilizes ‘click’ chemistry to detect nascent protein synthesis in cells (Fig 7A) [52]. Detection of protein synthesis was based on the incorporation of a specialized alkyne-modified methionine homopropargylglycine (HPG) or alkyne-modified puromycin (OPPuro) (S13A–S13C Fig) into newly synthesized proteins in JW18 cells (Fig 7) or Drosophila testes (S13 Fig) respectively. Labelled proteins were detected using a chemo-selective ligation or “click” reaction between the alkyne modified proteins and an azide-containing fluorescent dye which was added. This resulted in a fluorescent readout within each host cell correlating to the level of protein synthesis. Note that we assumed the majority of protein synthesis detected in this assay was host-related, however HPG can also be incorporated during bacterial protein synthesis, thus Wolbachia translation will have contributed to the overall fluorescent readout. We further processed the samples to detect Wolbachia by RNA FISH and then imaged cells using confocal microscopy (Fig 7A and 7B). Quantification of this fluorescent readout of protein synthesis in the JW18 population and Wolbachia-free JW18DOX revealed that the median translation level in the Wolbachia-infected JW18 cell line was reduced 23.6% compared to Wolbachia-free JW18DOX (Non-parametric Mann Whitney, p<0.0001) (Fig 7C). This observation is consistent with previous observations of a global reduction in host translation [53] and translation machinery [49, 53, 54] in the presence of Wolbachia infection. To extend this observation, we binned cells into categories based on the level of Wolbachia infection and found a decreasing trend in the level of translation in cells as Wolbachia infection level increased (Fig 7D). The median translation level decreased by 43% when comparing cells that did not contain Wolbachia to cells containing a medium-high level of infection (Fig 7D). Furthermore, we found a statistically significant negative correlation (r = -0.1344, p = 0.006, Pearson’s correlation) between translation levels and Wolbachia levels in JW18 cells (S15 Fig). Together, these results show a relationship between Wolbachia infection level and host translation in JW18 cells.
The recent applications of Wolbachia as a tool to lower the transmission of vector-borne viruses necessitates a comprehensive analysis of the relationship between Wolbachia and the vector host. In particular, the observation that increasing Wolbachia density leads to stronger antiviral effects in vectors [10, 25–27] argues for a thorough examination of how intracellular Wolbachia levels are controlled. Here we focused on understanding which host systems influence Wolbachia levels. We performed a comprehensive unbiased whole genome RNAi screen that adapted RNA FISH for a high throughput approach. Traditionally, visual cell culture-based screens that investigate host-pathogen interactions use immunofluorescent staining, luminescent readouts, or fluorescently-tagged pathogens. The lack of tools for Wolbachia such as a commercially available antibody or a fluorescently-tagged Wolbachia strain necessitated our RNA FISH approach as a visual assay. This screen confirmed the feasibility of an RNA FISH detection approach as 1117 host genes were identified that alter Wolbachia levels. This accounted for approximately 8% of all screened genes. Knock down of 329 of these genes resulted in increased Wolbachia levels whereas 788 genes resulted in decreased Wolbachia levels. In summary, the screen successfully identified a comprehensive array of host genes that influence intracellular Wolbachia levels.
Here we report that Wolbachia levels are sensitive to changes in host translation. When host translation components such as the ribosome or translation initiation complex are perturbed by RNAi we observe remarkable increases in Wolbachia levels (Fig 5). In support, Wolbachia levels increase in the Drosophila ovary (Fig 6), testis hub (S13 Fig), larval imaginal discs (S12 Fig), and in the whole fly (S14 Fig) for ribosomal mutants. Furthermore, Wolbachia levels increase upon global host translation inhibition when flies were fed with cycloheximide (Fig 6). Collectively, these results provide the first evidence that Wolbachia levels are sensitive to host translation level changes and suggests that host translation might normally play an inhibitory role in regulating intracellular Wolbachia levels.
In addition to the sensitivity that Wolbachia displays towards host translation levels it is possible that host translation is also directly affected by Wolbachia. Quantification of protein synthesis in individual JW18 cells compared to JW18DOX cells revealed that JW18 cells had overall significantly lower translation level compared to the Wolbachia-free JW18DOX cell line (Fig 7). When JW18 cells were classified according to Wolbachia infection level, higher Wolbachia levels in JW18 cells correlated with significantly lower levels of host translation as measured by global protein synthesis levels (Fig 7, S15 Fig). We did not observe changes in translation components at the RNA level (S3 Fig), however, recent proteomics studies revealed that over 100 host proteins with roles in host translation were suppressed in the presence of Wolbachia [49, 55]. The mechanisms underlying this remain to be determined. One possibility is that the host translation is dampened by a stress response to Wolbachia. However, our gene expression analysis did not suggest any major alterations in stress response-related genes at the RNA level in response to Wolbachia (S16 Fig). Yet, several significant changes in stress response were detected at the proteome level suggesting that stress could play a role in the Wolbachia-host intracellular relationship [49, 55, 56]. Host translation shutdown via metabolic stress pathways is a common mechanism employed by pathogens [57]. The other possible mechanism for dampening host translation is active manipulation of host translation machinery by Wolbachia perhaps at the post-translational level as our data do not suggest changes at the transcriptional level (S3 Fig). Wolbachia encodes and expresses a fully functional type IV secretion system and many potential effector proteins [58–60]. Although the majority of Wolbachia effector proteins remains to be characterized, it is possible that Wolbachia encodes effectors that can manipulate the host’s translation machinery at a post-translational level as is the case for other intracellular bacteria such as Legionella [61, 62].
Wolbachia interaction with host translation could be important in the context of positive-strand RNA virus infection in the host. Wolbachia-mediated suppression of viral replication in hosts is well described [7, 8, 11, 12, 63]. Multiple mechanisms may underlie this observation including interference with viral entry and very early stages of viral replication [38, 53, 56, 63–70]. All viruses depend on host translation machinery for replication of their genomes. One intriguing possibility is that the interaction between Wolbachia and host translation could impact viral replication [53]. Wolbachia infection inhibits positive-strand RNA viral replication at very early stages of viral replication in the virus lifecycle [53, 63]. Thus, changes in host translation could be one mechanism by which Wolbachia infection contributes towards viral replication interference. Future work to elucidate whether this is a contributing mechanism to the Wolbachia-mediated antiviral response in a wide range of Wolbachia-host-virus relationships may provide valuable field applications for combating vector-borne viruses.
Our whole genome screen yielded a diverse range of host systems and complexes that influenced Wolbachia levels. Manual curation and bioinformatic analyses such as GO term enrichment and network analysis identified host pathways such as translation initiation, ribosome, cell cycle, splicing, immune-related genes, proteasome complex, COPI vesicle coat, polarity proteins and the Brahma complex. The GO term enrichment analysis and COMPLEAT network analysis suggested that the 329 genes resulting in Wolbachia increases formed a more robust dataset than the larger 788 gene category resulting in Wolbachia decreases owing to a lack of enrichment for specific networks and processes in this category. For this reason, we focused on the host networks that increased Wolbachia in this report. Nevertheless, future follow-up analysis on genes that decreased Wolbachia levels especially the larger categories such as metabolism & transporters, cytoskeleton, cell adhesion & extracellular matrix, as well as membrane dynamics and vesicular trafficking may yield rewarding results. We already appreciate that Wolbachia relies on several aspects of these broad categories. For example, Wolbachia can alter host iron, carbohydrate and lipid metabolism [71–75]. Further, Wolbachia interacts with host cytoskeleton such as microtubules for transport and host actin [32, 34, 35, 76]. Finally, Wolbachia resides within a host-derived membrane niche, as such genes identified in the membrane dynamics and vesicular trafficking category would be of interest [77, 78]. Further investigation of these Wolbachia-decreasing categories may provide comprehensive insights into Wolbachia-host interactions.
Interestingly our study also revealed that knockdown of the core proteasome leads to increases in Wolbachia levels (Fig 4A, S8 Fig, S10 Fig). A previous report suggested that Wolbachia require high levels of proteolysis for optimal survival [36]. An explanation for this discrepancy might be that previous observations of decreased Wolbachia levels were based on RNAi experiments that knocked down ubiquitin-related components not the core proteasome. Ubiquitination is known to function in many diverse contexts and pathways such as autophagy, cell cycle, immune response, DNA damage response and regulation of endocytic machinery [79]. Our screen also identified several ubiquitin-related components whose knockdown resulted in decreased Wolbachia levels (Fig 4B). Perhaps Wolbachia relies on the host ubiquitination system for survival in an unknown but specific context, not simply for providing amino acids as nutrients from the degradation of proteins by the proteasome. Our gene expression data (S3 Fig) along with recent proteomics studies [49, 55] suggest that the host proteasome is upregulated in the presence of Wolbachia. We propose that the host proteasome plays an inhibitory role in Wolbachia level regulation. Perhaps Wolbachia levels are controlled by degradation of effector proteins in the cytosol, thereby preventing Wolbachia from utilizing the host cell in an optimal manner. The results both in the JW18 cell line as well as in the ovary (S10 Fig), testis hub (S13 Fig), and larval imaginal discs (S12 Fig) of D. melanogaster strongly suggest that the host core proteasome normally plays a restrictive role in Wolbachia-host interactions that is separate from observations of ubiquitin pathway perturbation.
In summary, here we presented a whole genome screen to identify host systems that influence Wolbachia levels. Our focus was on Wolbachia sensitivity to alterations of host translation-related components such as the ribosome and translation initiation factors. We report a novel relationship between Wolbachia and host translation and suggest a restrictive role for host translation on Wolbachia levels. Future work to identify whether Wolbachia is able to actively manipulate host translation will provide valuable insight into understanding this unique host-symbiont relationship.
A stable Wolbachia-infected Drosophila cell line (JW18) and a doxycycline-treated Wolbachia-free cell line (JW18DOX), where Wolbachia infection of the JW18 cell line was removed by treatment with doxycycline to generate a Wolbachia-free version were kindly provided by William Sullivan at UCSC. Cell lines were maintained in Sang and Shield media (Sigma) supplemented with 10% heat inactivated One Shot Fetal Bovine Serum (Life Technologies). The doubling time of the JW18 cells was calculated from a growth curve using the formula (t2-t1)/3.32 x (logn2-logn1) where ‘t’ is time and ‘n’ is cell number. The JW18DOX cell line was originally treated with doxycycline in late 2010 and thereafter maintained in normal culturing media without doxycycline.
D. melanogaster infected with the wMel Wolbachia strain was a gift from Luis Teixeira (Instituto Gulbenkian de Ciência). To generate a Wolbachia-infected double balancer line, Wolbachia-infected virgins were crossed to Sp/CyO; MKRS/TM6B males. In the next generation Wolbachia-infected +/CyO;+/TM6B female virgins were crossed to males of the original double balancer stock. In the final generation, a stock of Wolbachia-infected Sp/CyO; MKRS/Tm6B double balancers was established.
Ribosomal mutant fly stocks were ordered from the Bloomington Drosophila Stock Center. The following haploinsufficient lines were used: RpS32/TM2 (stock no. 1696) and RpL27A1/CyO (stock no. 5697). Males from each line were crossed at 25°C to the Wolbachia-infected double balancer line described above. Siblings from each cross were matured for 5 days before tissues were dissected and stained for Wolbachia using the 23s rRNA Wolbachia-specific FISH probe. Tissues were imaged in Z-stacks using confocal microscopy. Quantification of the integrated density of the 23s rRNA Wolbachia FISH probe in tissue Z stacks were done using Fiji Image Processing software as described in the FISH section.
A dominant temperature-sensitive (DTS) lethal mutant for proteasome component Pros26, known as DTS5, was a gift from John Belote, Syracuse University. Heterozygotes die as pupae when raised at 29°C, but are viable and fertile at 25°C. The mutant contains a missense mutation in the gene encoding the β6 subunit of the 20S proteasome. Males were crossed to Wolbachia-infected female double balancers at the permissive temperature of 25°C. Hatched offspring from the cross were matured at the non-permissive 29°C for 5 days prior to dissection of the tissues. Imaging and analysis of tissues of siblings were done as described above for the ribosomal mutant crosses.
For fecundity testing we performed the following crosses: In the first generation, we crossed Wolbachia-infected Sp/CyO; MKRS/TM6B double balancer virgin females to males from RpS32/TM2, RpL27A1/CyO, and DTS5 stocks. For control fecundity experiments that were Wolbachia-free, we set up the reciprocal crosses using virgin females from RpS32/TM2, RpL27A1/CyO, and DTS5 stocks and males from the Wolbachia-infected Sp/CyO; MKRS/TM6B double balancer stock. In the next generation we collected the following virgin females for fecundity testing: RpS32/MKRS and TM2/MKRS (control sibling), RpL27A1/CyO and Sp/CyO (control sibling), and DTS5/TM6B and MKRS/TM6B (control sibling). For fecundity testing, all virgins were 2–4 days old. These virgins were mated with Oregon R males for one day prior to setting up the cages for fecundity testing. For testing, 3 females and 1 OregonR male were allowed to mate together and lay eggs on agar plates for 6 hours each day for 3 days. For each genotype between 6–17 ‘3x1’-matings were set up. 24 hours later eggs were counted and scored for hatching.
Genomic DNA was extracted from cells or Drosophila tissues using a DNeasy Blood & Tissue Kit (Qiagen) following manufacturer’s instructions. To quantify the level of Wolbachia in the sample a DNA qPCR assay was performed using SYBR Green I Master 2x (Roche), using a Roche LightCycler 480 machine. Primer sets included a primer set to detect wspB which is a gene encoding a Wolbachia surface antigen (F: 5’ ACA ACA GCT ATA GGG CTG AAT TGG AA 3’, R: 5’ TCA GGA TCC TCA CCA GTC TCC TTT AG 3’), as well as a primer set to detect the Drosophila gene RpL32 (also known as RpL49) (F: 5’ CGA GGG ATA CCT GTG AGC AGC TT 3’, R: 5’GTC ACT TCT TGT GCT GCC ATC GT 3’). Wolbachia levels were normalized by the host nuclear marker for each sample.
RNA was extracted and DNase-treated using a RNeasy kit (Qiagen) according to manufacturer’s instructions. Total RNA was reverse transcribed using an RNA to cDNA EcoDry Premix (OligodT) or EcoDry Premix (Random hexamer) (Clontech) according to manufacturer’s instructions. Quantitative PCR was performed on 1/200 of the RT reaction using LightCycler 480 SYBR Green I Master 2x (Roche) and a Roche LightCycler 480 machine. Results were normalized to the housekeeping gene Rp49. Primer sets used to validate RNAi knockdown were designed to amplify areas outside of the dsRNA amplicon. Gene knockdown was represented relative to expression levels in LacZ dsRNA-treated cells.
Genomic DNA was extracted from JW18 cells in duplicate. Samples were quantified using a Qubit fluorometer (Thermo Fisher Scientific). DNA libraries were prepared using a Nextera DNA Library Prep kit (Illumina) according to manufacturer’s instructions. DNA libraries were sequenced on an Illumina HiSeq2500 Sequencing platform in two lanes as paired-end reads 100 cycle lanes.
Total RNA was extracted in triplicate from JW18 and JW18TET cells and DNase-treated. RNA was quantified by Nanodrop and 5μg of each sample was subjected to two rounds of rRNA depletion using a Ribo-Zero rRNA Removal Magnetic kit (Epicentre, Illumina) or NEBNext rRNA Depletion Kit (Human/Mouse/Rat) (New England BioLabs, E6310L). After rRNA depletion libraries were prepared according to manufacturer’s instructions using the NEBNext Ultra Directional RNA Library Prep Kit for Illumina (New England BioLabs, E7420L) and NEBNext Multiplex Oligos for Illumina Index Primers Set I (Illumina, E7335). After adaptor ligation, the libraries were amplified by qPCR using the KAPA Real-time amplification kit (KAPA Biosystems). Finally, libraries were purified using Agencourt AMPure XP beads (Beckman Coulter) as described in the NEBNext Ultra Directional RNA Library Prep Kit for Illumina (New England BioLabs, E7420L) protocol. Quality and quantity was assessed using a Bioanalyzer (Agilent) and a Qubit fluorometer (Thermo Fisher Scientific). Libraries were sequenced on an Illumina HiSeq2500 Sequencing platform in single read 50 cycle lanes.
Differential gene expression analysis was performed from one lane of high output, single end reads 50, Illumina HiSeq run. The experiment consisted of 3 technical replicates each for JW18 and JW18TET cells. The alignment program, Tophat (version 2.0.9) (https://ccb.jhu.edu/software/tophat/index.shtml) was used for reads mapping with two mismatches allowed. Featurecounts (http://bioinf.wehi.edu.au/featureCounts/) was used to find the read counts for annotated genomic features. For the differential gene statistical analysis, DESeq2 R/Bioconductor package in the R statistical programming environment was used (http://www.bioconductor.org/packages/release/bioc/html/DESeq2.html).
We mapped short reads generated from DNA-Seq with Bowtie2 version 2.2.9 [80]. We used default parameters and mapped to combined sequences of Drosophila genome release 6 [81] and Wolbachia pipientis wMel ([58], GenBank accession ID AE017196.1). We determined basal ploidy level of JW18 cells by clustering normalized DNA-Seq read densities as in [82]. In doing that, we identified different copy number segments whose normalized read densities are between zero (no DNA content) to the mean density (basal ploidy level). Clusters of such read densities indicate the minimum ploidy. From the determined basal ploidy, we called copy numbers of JW18 cell line genome using Control-FREEC version 5.7 [83] at 1 kb levels. We called copy numbers in an identical way to [82] but with this exception; we performed calling twice and combined the results. Control-FREEC performs GC contents-based normalization of DNA-Seq reads. Therefore, we set the minimum expected GC contents to be 0.30 for robust copy number calling of the cell line genome first. Then we underwent our analysis again with the minimum expected GC contents of 0.25 to increase sensitivity against the bacterial genome. In our reports, we combined copy number calls from the former for JW18 cells, and from the latter for Wolbachia. In S1 Fig, we used DNA-Seq results from [83] to call copy number calls on S2R+ and Kc167 cells. We re-analyzed the original data after mapping to the release 6 genome as above.
The method described in [59] was used to analyze the genotype of the Wolbachia strain in the JW18 cell strain. Briefly, fastq sequences were mapped against a “holo-genome” consisting of the Release 5 version of the D. melanogaster genome (Ensembl Genomes Release 24, Drosophila_melanogaster.BDGP5.24.dna.toplevel.fa) and the Wolbachia wMel reference genome (Ensembl Genomes Release 24, Wolbachia_ endosymbiont_of_drosophila_melanogaster.GCA_000008025.1.24) [84, 85]. Holo-genome reference mapping was performed using bwa mem v0.7.5a with default parameters in paired-end mode. Mapped reads for all runs from the same sample were merged, sorted and converted to BAM format using samtools v0.1.19 [86]. BAM files were then used to create BCF and fastq consensus sequence files using samtools mpileup v0.1.19 (options -d 100000). Fastq consensus sequence files were converted to fasta using seqtk v1.0-r76-dirty and concatenated with consensus sequences of Wolbachia-type strains from [27]. Maximum-likelihood phylogenetic analysis on resulting multiple alignments was performed using raxmlHPC-PTHREADS v8.1.16 (options -T 12 -f a -x 12345 -p 12345 -N 100 -m GTRGAMMA) [87]. Copy number variants were detected by visual inspection of read depth across the wMel genome.
GO enrichment analysis were performed using PANTHER Version 12.0 (release 2017-07-10) (http://www.pantherdb.org/). The entire set of screened genes was used as the experimental background. Protein complex enrichment analysis was performed using COMPLEAT (http://www.flyrnai.org/compleat/). As the experimental background we used the entire set of screened genes. Complex size was limited to ≥3, with a p value filter of p<0.05.
Outdated amplicons from the Drosophila RNAi Screening Center (DRSC) whole genome library 2.0 were identified using Updated Targets of RNAi Reagents (UP-TORR) (http://www.flyrnai.org/up-torr/). For amplicons that could be transferred to Release 6, we followed the dsRNA in vitro synthesis protocol as described by the DRSC. The DNA templates were generated by PCR on genomic DNA extracted from the JW18 cells, genomic DNA from wild-type flies or pBlueScript SK (+) plasmid DNA (in the case of LacZ). All gene specific primer sequences were selected by the DRSC and the T7 promoter sequence (TAATACGACTCACTATAGGG) was added to the 5’ ends of all primer pairs. Gradient PCR reactions were performed with Choice Taq Mastermix (Denville Scientific Inc.) using 5ng of genomic DNA, 0.1ng of plasmid DNA, or 1:3 diluted PCR template DNA. PCR products were verified by electrophoresis on a 0.7% (w/v) agarose gel with the 1kb PLUS ladder (Invitrogen) and only products with a clear single band were selected for IVT. IVT was performed according to manufacturer’s instructions for the MEGAscript T7 Transcription Kit (Ambion) using 8μl of amplified T7-flanked PCR product per reaction. dsRNA products were DNase-treated using Turbo DNase (Ambion) and purified with Qiagen RNeasy Mini spin columns (Qiagen) according to manufacturer’s protocols. Quality of purified dsRNA was assessed by electrophoresis on a 0.7% agarose gel, and concentration was determined by Nanodrop (Thermo Fisher Scientific) [40].
RNAi in JW18 cells was done using a bathing method described by the DRSC [40]. dsRNA aliquots were prepared in serum-free Sang and Shield media (Sigma). dsRNA was added to wells to yield a final concentration of 25nM. For the whole genome screen the pre-arrayed DRSC Drosophila Whole Genome Library Version 2.0 was used. For each RNAi experiment, sub-confluent JW18 cells were scraped, pelleted (1000rpm for 5–15 minutes), and re-suspended in serum-free Sang and Shield media (Sigma) to seed 40 000 cells in 384 well format. Cells and dsRNA were incubated together at room temperature for 30 minutes in serum-free conditions. Thereafter Sang and Shield media (Sigma) supplemented with 10% heat inactivated One Shot Fetal Bovine Serum (Life Technologies) was added to each well and incubated at 25°C for 5 days before analysis.
Cells were plated in Poly-L-lysine-coated chambered cover-glass wells (Thermo Scientific) and allowed to settle. Medium was aspirated and cells were washed with 1xPBS before fixing with 4% PFA in 1xPBS (Electron Microscopy Sciences). Cells were washed twice in 1xPBS followed by two washes in 100% methanol (Fisher) before finally adding 100% methanol to each chamber and sealing it with Parafilm M Film (Sigma) for storage at -20°C overnight or up to 1 month. Samples were rehydrated using the following washes: MeOH: PBT (1xPBS, 0.1% Tween-20) (3:1), MeOH:PBS (1:1), MeOH: PBS (1:3), and a final wash in 1xPBS. Samples were then post-fixed for 10 minutes in 4% PFA at room temperature. In a pre-hybridization step, samples were incubated in 10% deionized formamide and 2x SCC for 10 minutes at room temperature. Pre-hybridization buffer was then removed, and a hybridization solution containing a Wolbachia-specific FISH probe was added and incubated overnight at 37°C. For each sample, the volume of hybridization buffer added was dependent on the type of well used, but enough should be added to cover the sample. Typically, 60μl of hybridization buffer comprised 10% Hi-Di deionized formamide (Applied Biosystems Life Technologies), 1μl of competitor (5mg ml-1 E. coli tRNA (Sigma) and 5 mg ml-1 salmon sperm ssDNA (Ambion)), 10mM vanadyl ribonucleoside complex (New England Biolabs), 2xSSC (Ambion), 50μg nuclease-free BSA (Sigma), 10ng Wolbachia-specific FISH probe, made up to 60μl with DEPC-treated water. The Wolbachia specific probe was designed to the Wolbachia 23s rRNA and labeled with Quasar670 (Stellaris). After overnight hybridization samples were washed twice in pre-warmed pre-hybridization buffer for 15 minutes at 37°C. Followed by two washes in 1x PBS for 30 minutes each. Finally, samples were stained for 5 minutes in 1:500 DAPI:1xPBS followed by two washes in 1x PBS. Unless otherwise stated, samples were imaged as Z-stacks on a Zeiss LSM 780 confocal at 63x. For Wolbachia detection in dissected Drosophila ovaries and testes, the same protocol was followed from fixation onwards.
Quantification of Wolbachia levels based on the intensity of the Wolbachia 23s rRNA probe for tissues was done in Fiji Image processing software. Z-stacks capturing entire tissues were projected as ‘sum slices’. Each tissue was manually outlined using the Freehand tool. The measurements tool was set to capture the integrated density within the outlined tissue of the FISH probe channel stack as well as provide an area measurement of the outlined tissue. We normalized the integrated density reading for each by its total area.
Large-scale RNAi screening was done using the DRSC Drosophila Whole Genome Library Version 2.0 that was seeded in Corning clear bottom, black 384 well plates with 0.25μg dsRNA pre-arrayed per well. This concentration of dsRNA was appropriate for the bathing method of RNAi [40]. JW18 cells were re-suspended in serum-free media at 4x106 cells/ml and an automated Matrix Wellmate dispenser (Thermo Fisher Scientific) was used to dispense 40 000 cells into each well of the 384 well plates in a sterile tissue culture hood. The cells were incubated with the dsRNA in serum-free media for 30 minutes before automatic dispensing of Sang and Shield media (Sigma) supplemented with 10% heat inactivated One Shot Fetal Bovine Serum (Life Technologies) into each well. Plates were incubated at 25°C for 5 days in a humidity chamber. After 5 days plates were drained followed by automated dispensing of 4% paraformaldehyde (Electron Microscopy Sciences) and incubation at room temperature for 10 minutes and automatically aspirated thereafter. An automated BioTek EL406 liquid handler (BioTek) was used throughout the protocol for all aspiration steps and a Matrix Wellmate (Thermo Fisher Scientific) was used for all dispensing steps. Next, plates were washed once with 1xPBS followed by three washes with 100% methanol (Fisher), sealed with Parafilm M film, and stored overnight at -20°C (or up to 1 month). All subsequent rehydration, post-fixation, pre-hybridization and hybridization steps of the RNA FISH protocol described above were carried out in an automated manner. After overnight hybridization at 37°C the plates were washed twice with pre-warmed pre-hybridization buffer followed by incubation at room temperature for 30 minutes with 1xPBS/DAPI. Finally, plates were washed once with 1xPBS and 40μl 1xPBS was dispensed into all wells and plates were sealed with aluminum foil and stored at 4°C.
Plates were imaged with a 20x objective lens using an Arrayscan VTI Microscope (Cellomics) coupled with the automated image analysis software HCS Studio Cellomics Scan Version 6.6.0 (Thermo Fisher Scientific). Image acquisition involved identification of DAPI stained cell nuclei as primary objects, followed by application of a ring mask around the primary objects to identify Wolbachia associated with each cell as secondary objects. Segmentation of the objects was optimized to exclude any areas containing cell clumps. For each well, 1500 primary objects (DAPI cell nuclei) were acquired. RNAi screen primary data analysis and criteria for hit selection is summarized in S4 Fig.
Protein synthesis levels in JW18 cells were detected using a Click-iT HPG Alexa Fluor 594 Protein Synthesis Assay Kit (Molecular probes, C10429). Regular Sang and Shield media (containing methionine) (Sigma) was removed from JW18 cells. Cells were washed once in 1xPBS. A working solution of Click-iT HPG was prepared according to manufacturer’s instructions using methionine-free Grace’s Insect Medium (Thermo Scientific). Cells were incubated for 30 minutes in 50μM Click-iT HPG working solution. After incubation, cells were washed once in 1xPBS followed by fixation in 5% formaldehyde. To combine the protocol with RNA FISH detection of Wolbachia in the cells, we next proceeded to wash the cells twice in methanol followed by storing the sample in 100% methanol at -20°C overnight. From this point, we followed the RNA FISH protocol described in the FISH section. After FISH hybridization and post-hybridization washes, we incubated cells with 0.5% Triton X-100 in 1xPBS for 20 minutes at room temperature. Cells were washed twice with 3% BSA in 1x PBS. The Click-iT Reaction Cocktail was added to the samples for 30 minutes at room temperature protected from light. Thereafter, samples were washed once with Click-iT Reaction Rinse Buffer before staining with 1x HCS NuclearMask Blue Stain working solution as per manufacturer’s instructions. Samples were imaged on confocal at 63x magnification. Controls included in the assay were as follows: incubation of cells with cycloheximide at a final concentration of 100μg/ml for 1 hour prior to the start of the experiment as well as during the 30-minute incubation with HPG; and a negative control sample that was not incubated with HPG.
Quantification of the Click-iT fluorescent intensity signal within each cell was done in a similar manner as described for FISH signal in egg chambers in the FISH section. Briefly, projected Z-stacks were manually outlined in Fiji and integrated density and area measurements were captured for each cell using the measurements tool. This allowed for a normalized integrated density measurement for individual cells. These data could then be paired with the Wolbachia level within individual cells as measured by RNA FISH.
Protein synthesis in the Drosophila testis was detected by the Click-iT Plus OPP Alexa Fluor 594 protein synthesis assay kit (Molecular Probes) as previously described [88]. Samples were incubated for 30 minutes in 1:400 Click-iT OPP reagent in fresh Shields and Sang M3 Insect medium (Sigma).
|
10.1371/journal.pmed.1002727 | Association between severe drought and HIV prevention and care behaviors in Lesotho: A population-based survey 2016–2017 | A previous analysis of the impact of drought in Africa on HIV demonstrated an 11% greater prevalence in HIV-endemic rural areas attributable to local rainfall shocks. The Lesotho Population-Based HIV Impact Assessment (LePHIA) was conducted after the severe drought of 2014–2016, allowing for reevaluation of this relationship in a setting of expanded antiretroviral coverage.
LePHIA selected a nationally representative sample between November 2016 and May 2017. All adults aged 15–59 years in randomly selected households were invited to complete an interview and HIV testing, with one woman per household eligible to answer questions on their experience of sexual violence. Deviations in rainfall for May 2014–June 2016 were estimated using precipitation data from Climate Hazards Group InfraRed Precipitation with Station Data (CHIRPS), with drought defined as <15% of the average rainfall from 1981 to 2016. The association between drought and risk behaviors as well as HIV-related outcomes was assessed using logistic regression, incorporating complex survey weights. Analyses were stratified by age, sex, and geography (urban versus rural). All of Lesotho suffered from reduced rainfall, with regions receiving 1%–36% of their historical rainfall. Of the 12,887 interviewed participants, 93.5% (12,052) lived in areas that experienced drought, with the majority in rural areas (7,281 versus 4,771 in urban areas). Of the 835 adults living in areas without drought, 520 were in rural areas and 315 in urban. Among females 15–19 years old, living in a rural drought area was associated with early sexual debut (odds ratio [OR] 3.11, 95% confidence interval [CI] 1.43–6.74, p = 0.004), and higher HIV prevalence (OR 2.77, 95% CI 1.19–6.47, p = 0.02). It was also associated with lower educational attainment in rural females ages 15–24 years (OR 0.44, 95% CI 0.25–0.78, p = 0.005). Multivariable analysis adjusting for household wealth and sexual behavior showed that experiencing drought increased the odds of HIV infection among females 15–24 years old (adjusted OR [aOR] 1.80, 95% CI 0.96–3.39, p = 0.07), although this was not statistically significant. Migration was associated with 2-fold higher odds of HIV infection in young people (aOR 2.06, 95% CI 1.25–3.40, p = 0.006). The study was limited by the extensiveness of the drought and the small number of participants in the comparison group.
Drought in Lesotho was associated with higher HIV prevalence in girls 15–19 years old in rural areas and with lower educational attainment and riskier sexual behavior in rural females 15–24 years old. Policy-makers may consider adopting potential mechanisms to mitigate the impact of income shock from natural disasters on populations vulnerable to HIV transmission.
| Periods of climate extremes have been shown to lead to increases in high-risk behaviors, particularly in agricultural communities dependent on rainfall for their livelihoods.
Prior studies have linked these increases in riskier sexual behaviors, such as extramarital partnerships and transactional sex, to increases in HIV acquisition.
The Lesotho Population-Based HIV Impact Assessment, a national HIV survey conducted from November 2016 to May 2017 following a 2-year severe drought in southern Africa, allowed us to reevaluate this relationship in the setting of expanded antiretroviral use.
We paired geospatial data on accumulated rainfall from 2014 to 2016 with data from the survey to determine if there were any associations between drought and HIV outcomes.
A total of 12,887 adults ages 15–59 years completed a detailed questionnaire, and 11,682 underwent an HIV test.
Adolescent girls and young women ages 15–24 years in rural areas of drought had higher rates of high-risk behaviors, such as early sexual debut and transactional sex, and had lower educational attainment.
Living in a drought area appeared to be associated with greater HIV prevalence in young females and was associated with a lower HIV prevalence in young males.
However, external migration, commonly seen during these periods, was associated with a greater prevalence of HIV in men and women.
Future policy on mitigation of climate change in southern Africa may consider including HIV prevention interventions in populations at high risk. This could include preexposure prophylaxis (PrEP) for migrants and young women in areas of severe food insecurity.
Likewise, HIV programs that provide social and economic support to young women as part of an HIV prevention strategy should consider targeting areas affected by drought.
Further studies should be done in other settings to investigate the external validity of these findings, particularly in terms of the impact on HIV in young women.
| The impact of climate change on human health is becoming increasingly evident. Aside from changes in infectious disease transmission directly related to disturbances favoring multiplication of disease vectors, periods of climate extremes are often associated with changes in behavior as people struggle to survive in the face of loss of agricultural production [1,2]. As people, particularly women, address their food insecurity, they may be less likely to take steps to protect themselves from HIV infection [3,4], and studies have documented increases in HIV infections during drought-related famine periods in Africa [5,6]. A previous study of 21 Demographic and Health Surveys (DHS) across 19 countries in sub-Saharan Africa from 2003 to 2009 demonstrated that recurrent rainfall shocks were likely responsible for approximately 11% of HIV infections because of negative income shocks, particularly in high-prevalence countries and predominantly agrarian societies [6]. There is also concern that food insecurity could lead to decreased access to antiretrovirals (ARVs) because of economic constraints or decreased adherence or absorption of ARVs [7–9], with a subsequent increase in community viral load, drug resistance, and HIV transmission [10–13]. Despite the increased frequency and severity of droughts in the Sahel and southern Africa, few countries’ climate change adaptation policies currently include any intensified efforts for HIV prevention during climate-related events [14].
Southern Africa experienced 2 years of an El Nino–induced regional drought during the growing seasons of 2014–2015, including during key stages of crop development, leading to food shortages in 2016 and increased food costs for almost 40 million people in the region [15]. In Lesotho, where 55% of the population grow their own food, most people survive on rain-fed subsistence farming [16]. The drought led to a 67% reduction in maize production, with almost 25% of the population requiring emergency food assistance by August 2016 [16]. Lesotho also has a long tradition of labor migration to South Africa, and during periods of drought, migration increases, particularly from predominantly rural districts [16–18]. Furthermore, Lesotho is a country with a hyperendemic HIV epidemic, with prevalence above 25% in the adult population, and therefore is at greater risk of disruption to any improvements in epidemic control [3,19].
The Lesotho Population-Based HIV Impact Assessment (LePHIA) was a national survey conducted from November 2016 to May 2017 in collaboration with the Ministry of Health and the Centers for Disease Control and Prevention, with funding from the President’s Emergency Plan for AIDS Relief (PEPFAR). LePHIA examined the status of the HIV epidemic in Lesotho by measuring HIV prevalence, incidence, and viral load suppression (VLS). This study used LePHIA data to assess whether people living in areas most severely affected by the drought had higher HIV prevalence or changes in risk behaviors and whether there was any difference in the continuum of care among people living with HIV (PLHIV). The analysis additionally disaggregates by age band to examine the outcomes in youth and older people [19,20].
LePHIA employed a two-stage sampling design to select a nationally representative sample of adults and children aged 0–59 years in 418 enumeration areas (EAs) across all 10 districts. The sample size was powered on a relative standard error of 30% for incidence of HIV in adolescent girls and young women (AGYW). Consenting heads of household completed a household questionnaire, including a roster of all household members who resided in or had slept in the household the previous night. These individuals then consented to a questionnaire on sociodemographic and behavioral factors (S1 Text) and to home-based HIV testing. A guardian or parent provided permission to approach 10-to-17-year-olds who were then asked for assent for all procedures. The adult questionnaire was administered to participants aged 15–59. Written informed consent was documented at each stage via electronic signature. All participants provided written consent. The LePHIA protocol and data collection tools were approved by the Lesotho Research and Ethics Committee, the institutional review boards at Columbia University Medical Center (#AAAQ8537), and the United States Centers for Disease Control and Prevention.
Survey staff administered the household and the adult questionnaires during a face-to-face interview with participants using Google Nexus 9 tablets. The household questionnaire collected data on household assets and access to food. The adult questionnaire included questions on lifetime and recent sexual behaviors, as well as questions on the HIV continuum of care for those who reported being HIV positive. Only one female participant aged 15 years or older in each household was randomly selected to answer questions about experiences with sexual violence, to mask the nature of the questions to other members of the household. Any female younger than 18 who reported being sexually exploited was referred to support services for counseling and further management.
Rapid HIV testing was conducted using point-of-care (POC) tests—Determine HIV-1/2 Rapid Test (Alere)—and confirmed with a Uni-Gold HIV Test (Trinity Biotech), as per the national algorithm. Laboratory verification of all HIV-positive results was done using the Geenius HIV-1/2 Supplemental Assay (Bio-Rad). Viral load testing was done preferentially on plasma, or on dried blood spots (DBSs) if necessary, at a central lab on an automated platform.
Drought was quantified using precipitation estimates from the Climate Hazards Group InfraRed Precipitation with Station Data (CHIRPS), which blends a variety of satellite imagery with interpolated weather station data to create gridded rainfall estimates at dekadal time-step and 0.05° resolution, or approximately 30 km2 [21]. To capture the multiseason impact of El Nino, the 2-year total rainfall from June 2014 to May 2016 was summed and then ranked among all 2-year rainfall amounts within the 1981–2016 period and converted to an empirical percentile; therefore, a value of 1% signifies the driest 2-year period in the 35-year record, 100% signifies the wettest, and 50% signifies close to the median value. By using the rainfall accumulation over a 24-month period, the accumulation of data decreased the relative importance of errors in rainfall measurement. Use of these biennial deviations also ensured that any observed negative income shock for that period was a reflection of deviations from the norm, rather than a continuation of underlying farmers’ long-term adjustments to declining precipitation levels [21]. This gridded dataset was prepared by the Vulnerability Analysis and Mapping (VAM) Geospatial Analysis Team at the Analysis and Trends Service of the World Food Programme (WFP), using scripts developed in-house.
Latitude and longitude data from the centroid of each LePHIA-sampled EA were overlaid on the gridded rainfall dataset, with all individuals within each gridded area assigned the same level of rainfall, using ArcGIS Pro 2.0.1 (ESRI). Drought was defined at a percentile of 15% or lower of the historical record in order to generate a binary variable that approximates the level below which rainfall deficits are particularly harmful to gross domestic product (GDP) and maize yields [6]. This translates into the 2-year period being one of the five driest periods during the 35-year historical record.
The effect of the drought was examined on recent behavior (over the past 12 months), HIV prevalence, and the continuum of HIV care at the individual level, using weighted data for all analyses, per an a priori analysis plan (S1 Analysis Plan). Commercial and forced sex were lifetime measures because of the smaller number of respondents. Design weights were calculated based on sampling design, including probability of household selection, and adjusted for nonresponse at the household, individual, and biomarker level using the Chi-Squared Automatic Interactor Detector (SI-CHAID) software (Statistical Innovations); this was stratified by urban or rural residence, age group, and region, with peri-urban populations grouped with rural as per the Lesotho Bureau of Statistics. Poststratification weights were calculated to reflect the age distribution of the 2016 Lesotho census. Additional weights were used for the subsample of women replying to the sexual violence questions. A household wealth quintile was generated using principal components analysis of household assets to generate a wealth score, based on the previously described methodology used in the DHS [22]. Poverty was defined as a household living in the lowest two quintiles. All analyses were done in Stata version 15.1 using weighted data, with Jackknife replicate weights used for variance estimation.
We used logistic regression to assess the association of drought with individual-level likelihood of infection as a binary variable, stratified by region and gender, based on prior data indicating that income shock differentially impacts males versus females [20,23], as does urban versus rural location [6]. The analysis was stratified by age to highlight the 15-to-24-year-olds and, where numbers allowed, into adolescents (aged 15–19 years) versus young adults (aged 20–24 years) to identify those most likely to have been recently infected. We examined the association with behaviors associated with HIV acquisition, such as recent condom use, commercial sex [24], and migration [25]; in young people for whom the drought would be most likely to affect recent transitions, we also examined associations with educational attainment [24], transactional and intergenerational sex (sexual partner who was 10 or more years older), and early sex and marriage in 15-to-19-year-olds [26,27]. Transactional sex was defined as nonmarital, noncommercial sex entered into on the assumption of material benefits [28]. Educational attainment was defined as whether they had attended secondary school or greater and did not require completion. Food insecurity was defined as any 24-hour period without food to eat because of lack of resources in the past 4 weeks. To assess drought and the treatment cascade, we assessed the association of drought with awareness of HIV status, reported ARV therapy (ART) use, and VLS, defined as HIV RNA < 1,000 copies/ml in PLHIV. Note that the planned investigation of the association of drought with recent infection was not performed because of the small number of recent cases identified in LePHIA.
A multivariable logistic regression model of the association of drought with HIV prevalence was constructed according to a UNAIDS hierarchical framework linking environmental disasters to population displacement, poverty, and behavioral change [3] and included variables known to be associated with HIV infection—namely, age and household economic status—or to be independently associated with drought in univariable analysis with a p-value < 0.10 and plausibly associated with HIV infection. Because there appeared to be opposite patterns of association between drought and HIV prevalence by sex, we fitted the model with an interactive variable combining drought and sex, with participants living without drought coded as 0, males living in drought coded as 1, and females living in drought coded as 2.
Because of the multiple comparisons included in this study, we used the Benjamini-Hochberg method to adjust the probabilities for the chance of a false positive [29]. For this calculation, we obtained the false discovery rate (FDR)-corrected p-value for each hypothesis and for each strata of analysis—for instance, including all univariable analyses conducted to assess the associations between drought and behavior in rural females. We used an acceptable level of false positives of 5% and report significance based on the corrected p-value, with a corrected p-value of 0.10 considered weak evidence of an association. For the multivariable analysis, we used a significance threshold of 0.025 based on the Bonferroni correction of adjusting the normal p-value of 0.05 to reflect the two models.
Of 9,403 selected households, 8,824 (94%) completed a household interview; 9% of households in rural areas were vacant, compared to 6% of urban households. Of 7,893 eligible women and 6,135 eligible men, 12,887 (92%) completed an interview, and 11,682 of those (91%) were tested for HIV. The vast majority of households (94.8% of urban and 93.9% of rural) were in areas of drought (Fig 1). There was no difference in the sex ratio in drought versus nondrought areas, nor was there a difference in age distribution (Table 1). Compared to all urban residents and those living in rural nondrought areas, a greater proportion of participants in rural drought-affected areas were in impoverished (55.3%) and food-insecure households (39.6%).
In univariable analysis, both males and females residing in rural drought-affected areas were more likely to be poor, with the greatest effect on females, who had 4-fold higher odds of being members of impoverished households (odds ratio [OR] 4.22, 95% confidence interval [CI] 2.34–7.62, p < 0.001; Table 2). Among females residing in urban areas, drought was associated with an almost 5-fold increase in the odds of selling sex (OR 4.86, 95% CI 2.20–10.72, p < 0.001) and a 3-fold increase in the odds of having experienced forced sex (OR 3.11, 95% CI 1.42–6.85, p = 0.005). Among rural females, drought was associated with a reduction in condom use (OR 0.70, 95% CI 0.54–0.92, p = 0.01). Among males, there was weak evidence that living in rural drought-affected areas increased the odds of migration (OR 1.76, 95% CI 0.94–3.30, p = 0.08); this was not significant after FDR correction.
For young females living in a rural drought-affected area, there was an association with lower proportions attending secondary education (OR 0.44, 95% CI 0.25–0.78, p = 0.005), as well as a 3-fold increase in transactional sex (OR 3.26, 95% CI 1.78–5.98, p < 0.001). Among 15-to-19-year-olds, girls had increased odds of early sexual debut (OR 3.11, 95% CI 1.43–6.74, p = 0.004).
HIV prevalence was 26.9% (95% CI 25.4–33.4%) in urban areas and 24.7% (95% CI 23.5–25.8%) in rural areas and was higher in females (Table 3). Although the association was only significant at p < 0.10 after FDR correction, female adolescents living in drought-affected areas had almost 3-fold higher odds of HIV infection (OR 2.77, 95% CI 1.19–6.47, p = 0.02) compared to their counterparts in rural areas without drought. This association was weaker among girls living in urban drought areas (OR 1.84, 95% CI 0.94–3.62, p = 0.08), again compared to their counterparts in urban areas without drought.
In terms of the treatment cascade (Table 3), drought was not associated with awareness of status, reported ART use, or VLS among all PLHIV, aside from in young women in urban settings, for whom there was weak evidence of lower awareness of HIV-positive status (OR 0.23, 95% CI 0.08–0.70, p = 0.01).
In the multivariable analysis, after adjusting for age and economic, marital, and educational status as well as recent sexual behavior, there was a protective effect of drought noted for young males in terms of HIV infection (adjusted OR [aOR] 0.35, 95% CI 0.17–0.72, p = 0.006) and weak evidence of higher odds of HIV infection in young females (aOR 1.80, 95% CI 0.96–3.39, p = 0.07; Table 4). Recent migration, marital status, and intergenerational sex in the past year had the strongest associations with HIV in young people. In older people, there was no association between drought and HIV infection, but HIV infection was associated with food insecurity; being married; and reporting intergenerational, transactional, or commercial sex in the past year, whereas condom use at last sex was not protective. Higher educational attainment was protective in all ages.
To our knowledge, this is the first study of the association of drought with HIV infection in the era of expanded use of ART [19]. In times of drought, it is expected that families will adopt extraordinary measures to ensure that they can secure food, potentially including such non-agricultural sources of income as casual labor, domestic service, and alternative forms of sex work. The LePHIA survey provided an opportunity to assess coping mechanisms that might have been adopted in the 12 months prior to the survey, during times of crop failure and hunger.
The findings from our study suggest a drought response pattern including alteration of behaviors increasing risk [6]. The observed associations were stronger in rural areas, where food shortages and income shocks would be most pronounced because of limited diversification of economic activity. They are consistent with other results from LePHIA, which found some association between food insecurity and incident infection in AGYW. Moreover, we found that there was an increase in the constellation of risk behaviors that were independently associated with HIV infection, including transactional and commercial sex, suggesting that some women may indeed be relying on sexual favors to survive drought, if less openly in rural areas than in urban [20,30]. The increase in early sexual debut and reduced educational attainment in girls in rural areas is consistent with adolescent girls being removed from school for transactional partnerships or marriage so that families can benefit from the bride price [31]. The strength of the association between marriage and HIV infection suggests that there may be relatively high rates of transmission between married couples, either through infections developed prior to marriage or through infections acquired from extramarital partners.
For women in urban areas, despite an attenuated effect of the drought on household poverty, being in a drought-affected region was associated with substantially higher reporting of commercial and coercive sex, which supports results from a recent United Nations Populations Fund survey in Lesotho linking drought to an increase in gender-based violence [32]. This is indicative of increased vulnerability, potentially reflective of internal migration from rural areas seeking employment, and of global disruption of the economy due to the severity of drought and impact on other sectors [33,34]. As 48% of women reporting a history of commercial sex work in Lesotho are HIV positive, and forced sex also increases the risk of HIV [19], the circular migration back to rural homesteads poses significant risk to partners and communities [35,36]. Furthermore, as education was strongly protective against HIV infection in our multivariable model, the lower school enrollment seen in young rural females in drought areas could have far-reaching consequences, in terms of both HIV acquisition and entrenchment of poverty. Drought clearly has both immediate and long-term consequences and requires different targeted policies.
Concerning the HIV continuum of care, the results are reassuring in terms of broader epidemic control, as there were no observed associations in terms of VLS. There were various interventions in the form of food support and household rations, including the Super Cereal from Global Fund and Ready-to-Use Therapeutic Food from the WFP, targeting malnourished PLHIV on ART or with tuberculosis, and these efforts may have been successful in incentivizing PLHIV to stay on treatment and in mitigating the worst hunger-associated effects of the drought. However, the highly migratory nature of the Lesotho population makes it difficult to interpret the completeness of the data; as those on ART were preferentially provided with food support [13], stable patients might have been more likely to remain in the country, with nonadherent or undiagnosed PLHIV not being captured by the survey because of out-migration.
Study limitations include the ecological nature of the study, which did not measure indicators of the experience of drought at the individual level. However, we were able to demonstrate that households in rural drought areas were significantly more likely to be impoverished, which we did not observe in urban areas, supporting the hypothesized impact of drought on agrarian households. Second, the timing of the survey, which was conducted after the drought, and its cross-sectional design limit the ability to determine causality for observed associations; however, the focus on behavior changes in the past year makes it more likely that observed behaviors occurred during the drought. Social desirability bias may have influenced reporting of sexual behaviors, although underreporting would be more likely to bias the results toward the null hypothesis. The severity of the drought also meant that few people were unaffected, limiting our ability to assess certain outcomes. Furthermore, the drought data were based on combined satellite and ground station measurement of rainfall, and there were only two ground stations in Lesotho, which had intermittent functionality during the drought period. This means that the drought grid was derived primarily from statistical downscaling of satellite data, impacting the accuracy of estimates at high resolution and allowing for some misclassification of drought severity. However, the use of 2 years of data and using relative rank rather than an absolute value for our indicator of drought should ensure that our results are reasonably robust. It should also be noted that the survey only included persons residing in Lesotho; those who had left and not returned as of the survey date may have influenced HIV transmission, but we could only observe the association between drought, migration, and HIV infection among returned migrants. Although there was only weak evidence of an association between migration and drought for rural males, the multivariable model identified a strong association between migration and HIV infection in young people. Migration has played a key role in the epidemic in the region, as male migrants have been reported to engage in commercial sex and have less access to HIV care [25,37,38]. The lower prevalence of HIV in young men suggests that, in the absence of migration, drought and its attendant financial constraints might protect men, as it may reduce both incentives and resources to pay for commercial or transactional sex [20]. Finally, because of the breadth of this analysis, the multiple comparisons increase the risk of type I errors, which we have tried to limit by using FDR. Some of the multivariable findings that were marginally significant before the correction become nonsignificant with the application of FDR. In part, this is due to the smaller number of cases once the data are disaggregated, limiting the statistical power to test multiple associations at once. Future work with larger samples of young people may enable more fine analyses of causality and the associations between risk factors. However, the positive and statistically significant associations between drought and risk behaviors among young women—and between those behaviors and HIV infection—remain, presenting a plausible and concerning pattern of vulnerability.
In conclusion, this study provides further evidence for the need for a coordinated policy and strategy to attenuate the effects of drought on HIV infection in southern Africa. Young women appear particularly susceptible to the negative income shocks of drought, whereas men seem predominantly affected in terms of labor migration and its potential for increased long-term risk [39]. Potential interventions should minimize these shocks by targeting the myriad factors contributing to vulnerability and could include cash transfers to encourage families to keep children in school and avoid early marriage, provided to rural families in times of food shortage [40,41], and expanded programs for AGYW, sex workers, and migrants, including preexposure prophylaxis (PrEP). In light of the anticipated acceleration of such climatic extremes, more research is urgently needed on improving resilience of crops to drought to mitigate the severity of impact on household incomes and public health.
|
10.1371/journal.pcbi.1002302 | Predictability of Evolutionary Trajectories in Fitness Landscapes | Experimental studies on enzyme evolution show that only a small fraction of all possible mutation trajectories are accessible to evolution. However, these experiments deal with individual enzymes and explore a tiny part of the fitness landscape. We report an exhaustive analysis of fitness landscapes constructed with an off-lattice model of protein folding where fitness is equated with robustness to misfolding. This model mimics the essential features of the interactions between amino acids, is consistent with the key paradigms of protein folding and reproduces the universal distribution of evolutionary rates among orthologous proteins. We introduce mean path divergence as a quantitative measure of the degree to which the starting and ending points determine the path of evolution in fitness landscapes. Global measures of landscape roughness are good predictors of path divergence in all studied landscapes: the mean path divergence is greater in smooth landscapes than in rough ones. The model-derived and experimental landscapes are significantly smoother than random landscapes and resemble additive landscapes perturbed with moderate amounts of noise; thus, these landscapes are substantially robust to mutation. The model landscapes show a deficit of suboptimal peaks even compared with noisy additive landscapes with similar overall roughness. We suggest that smoothness and the substantial deficit of peaks in the fitness landscapes of protein evolution are fundamental consequences of the physics of protein folding.
| Is evolution deterministic, hence predictable, or stochastic, that is unpredictable? What would happen if one could “replay the tape of evolution”: will the outcomes of evolution be completely different or is evolution so constrained that history will be repeated? Arguably, these questions are among the most intriguing and most difficult in evolutionary biology. In other words, the predictability of evolution depends on the fraction of the trajectories on fitness landscapes that are accessible for evolutionary exploration. Because direct experimental investigation of fitness landscapes is technically challenging, the available studies only explore a minuscule portion of the landscape for individual enzymes. We therefore sought to investigate the topography of fitness landscapes within the framework of a previously developed model of protein folding and evolution where fitness is equated with robustness to misfolding. We show that model-derived and experimental landscapes are significantly smoother than random landscapes and resemble moderately perturbed additive landscapes; thus, these landscapes are substantially robust to mutation. The model landscapes show a deficit of suboptimal peaks even compared with noisy additive landscapes with similar overall roughness. Thus, the smoothness and substantial deficit of peaks in fitness landscapes of protein evolution could be fundamental consequences of the physics of protein folding.
| One of the most intriguing questions in evolutionary biology is: to what extent evolution is deterministic and to what extent it is stochastic and hence unpredictable? In other words, what happens if “the tape of evolution is replayed:” are we going to see completely different outcomes or the constraints are so strong that history will be repeated [1]–[4]? If evolution is envisaged as movement of a population across a fitness landscape, the question can be reworded more specifically: among the numerous trajectories connecting any two points on the landscape, what fraction is accessible to evolution? Until recently, these remained purely theoretical questions as experimental study of fitness landscapes in the actual sequence space was impractical, due both to the technical difficulty of producing and assaying numerous expressed sequence variants and to the more fundamental problem of defining an adequate quantitative measure of fitness. However, recent experimental studies of fitness landscapes could potentially shed light on the problem of evolutionary path predictability.
The most thoroughly characterized feature of empirical fitness landscapes is the structure near a peak. In experiments that examine the peak structure, a high fitness sequence is typically subjected to either random mutations or an exhaustive set of mutations at a small number of important sites. The resulting library of mutants is then assayed to measure a proxy of fitness [5]–[9]. Significant sign epistasis (a situation in which the fitness effect of a particular mutation can be either positive or negative depending on the genetic context) has been observed. Deviations from the additive fitness model have been found to be independent of the genetic context and purely random [10]–[13]. Because these studies characterize only a small region of the landscape, they cannot be used to address the question of path predictability.
Another broad class of experiments probes the evolutionary trajectories from low to high fitness. Usually, in such experiments, a random peptide is subjected to repeated rounds of random mutagenesis and purifying selection [8], [14]–[17]. During this process fitness grows with each generation and eventually stagnates at a suboptimal plateau. The characteristics of the fitness growth as well as the dependence of the plateau height on the library size can be used to classify landscapes [18]. A quantitative comparison to the model of random epistatic landscapes ( is the number of sites in an evolving sequence and is the number of sites that affect the fitness contribution of a particular site through epistatic interactions) can even yield quantitative estimates of and [19], [20]. The directed evolution studies explore the evolutionarily accessible portion of the landscape and could in principle be used to shed light on the question of path predictability. However, the inaccessible regions of the landscape remain unexplored and the volume of data at this point is insufficient to obtain quantitative conclusions regarding path predictability.
A different type of landscapes has been explored in various microarray experiments where protein-DNA(RNA) binding affinity serves as the proxy for fitness [21], [22]. These experiments produce vast, densely sampled landscapes. A comparison with a sophisticated Landscape State Machine model of a correlated fitness landscapes yields estimates of the model parameters [23], [24]. The DNA binding landscapes, in principle, contain the information required for the analysis of path statistics, and could be a valuable resource for advancing the understanding of evolutionary path predictability.
Empirical studies that exhaustively sample a region of the fitness landscape allow one to actually assess the accessibility of the entire set of theoretically possible evolutionary trajectories in a particular (small) area of the fitness landscape. For example, all mutational paths between two states of an enzyme, e.g., the transition from an antibiotic-sensitive to an antibiotic resistant form of -lactamase [25]–[27] or the transition between different specificities of sesquiterpene synthase [28] have been explored. The results of these experiments, which out of necessity explore only short mutational paths of amino acid replacements, suggest that there is a substantial deterministic component to protein evolution: only a small fraction of the possible paths are accessible for evolution [25], [29]–[31].
Recent analyses of fitness data have revealed dense networks of genetic and molecular interactions responsible for the substantial ruggedness and sign epistasis of empirical fitness landscapes [13], [32]. The emerging quantitative analysis of fitness landscapes can shed light on some of the most fundamental aspects of evolution but the interpretation of the currently available experimental results requires utmost caution as only a minuscule part of the sequence space can be explored, and that only for a few more or less arbitrarily selected experimental systems.
Here we focus on the question of the predictability of mutational paths which is intimately tied to the ruggedness/smoothness of the fitness landscapes. The study of random landscapes of low dimensionality revealed an intuitively plausible negative correlation between the roughness of a landscape and the availability of pathways of monotonic fitness [33]. In the same study, Carneiro and Hartl showed that experimentally characterized landscapes are significantly smoother than their permuted counterparts and exhibit greater peak accessibility [33].
To gain insights into the structure of the fitness landscapes of protein evolution and in particular the accessibility of mutational paths we used a previously developed simple model of protein folding and evolution [34]. The key assumption of this model, which is based on the concept of misfolding-driven evolution of proteins [35]–[37], is that the fitness of model proteins is determined solely by the number of misfolded copies that are produced before the required abundance of the correctly folded protein is reached. We have previously shown that this model accurately reproduces the shape of the universal distribution of the evolutionary rates among orthologous protein-coding genes along with the dependencies of the evolutionary rate on protein abundance and effective population size [34]. These results appear to suggest that our folding model (described in detail the Methods section) is sufficiently rich to reproduce some of the salient aspects of evolution. The model is also simple enough to allow exhaustive exploration of the fitness landscapes, which prompted us to directly address the problem of evolutionary path predictability.
We build on the efforts of Carneiro and Hartl [33] who examined the statistics of evolutionary trajectories. Although counting monotonic fitness paths reveals important features of the landscapes, we argue that reliable retrodiction of the evolutionary past is possible (i.e., evolution is quasi-deterministic) only when the available monotonic paths are similar to each other in a quantifiable way. We therefore propose a measure of path divergence to quantify the difference between the available monotonic paths. Our aims are to investigate the structure of the fitness landscapes of protein evolution and to elucidate the connection between the roughness of landscapes and the predictability of mutational trajectories. We analyze three classes of fitness landscapes: landscapes in which fitness is derived from the folding robustness of model polymers; additive random landscapes perturbed by noise; and experimental landscapes derived from the combinatorial mutation analysis of drug resistance and enzymatic activity. We show that all three classes of landscapes are markedly smoother than their randomly permuted counterparts and all exhibit a similar qualitative connection between roughness and path predictability. However, at the same level of path predictability, the folding landscapes have substantially fewer fitness peaks. Equivalently, mutation paths are more predictable than one would expect based on the number of peaks if the landscapes were uncorrelated. Given that the statistical properties of the model landscapes can be directly traced to the constraints imposed by the energetics and kinetics of a folding heteropolymer, we hypothesize that the relative smoothness and the suppression of suboptimal peaks in fitness landscapes of protein evolution are fundamental consequences of protein folding physics.
Carneiro and Hartl compared small random landscapes to several empirical fitness landscapes using deviation from additivity as a measure of roughness [33]. They found that empirical landscapes were significantly smoother than their random counterparts and that the degree of smoothness was correlated with the number of monotonic paths to the main summit. Deviation from additivity of a landscape is computed by fitting an additive model in which the fitness of each sequence is different from the peak fitness by the sum of contributions of the substitutions that differentiate it from the peak sequence. The negative fitness contributions of the substitutions to the peak fitness are adjusted to minimize the sum of squares of the differences between the actual fitnesses in the landscape and the fitnesses predicted by the additive model. Deviation from additivity is defined as , where is the number of points in the landscape.
Because roughness of a multidimensional landscape with variable degree connectivity is not an intuitive concept, we introduce three additional quantitative measures to probe alternative facets of the concept of roughness. First, local roughness is the root mean squared difference between the fitness of a point and its neighbors, averaged over the entire landscape. As defined, local roughness conflates the measures of roughness and “steepness.” For example, a globally smooth landscape, in which fitness depends only on the distance from the peak, will have a non-zero local roughness. However, because there is a large number of directions that change the distance from the peak by one, the local roughness of a globally smooth landscape will be vanishingly small. In addition, our landscapes tend to be globally flat–so that the average decrease in fitness due to a single mutation step away from the main peak is much smaller than the local fitness variability–everywhere except a small region around the main peak (see Fig. 1). Therefore, the landscape-average local roughness in our case is a true measure of the local fitness variability.
Second, the fraction of peaks is the number of points with no fitter neighbors divided by the total number of points in the landscape. A strictly additive landscape has a single peak [30] whereas the peak fraction in landscapes derived from the folding model as well as the corresponding randomized landscapes depends on the method of landscape construction, alphabet size and sequence length.
Third, the roughness of a landscape can be assessed by identifying its tree component. The tree component is the set of all nodes with no more than one neighbor of higher fitness. Thus, the tree component includes peaks and plateaus. Monotonic fitness paths along the tree component form a single or several disjoint tree structures without loops. In the limit of high selection pressure, a mutational trajectory that finds itself on the tree component has a single path to the nearest peak or plateau, i.e. evolution on the tree component is completely deterministic. We use the mean distance to the tree component, i.e. the distance to the tree component averaged over the landscape, as a measure of roughness. In a fully additive landscape, only the peak sequence and its immediate neighbors belong to the tree component and therefore the mean distance to the tree component is a measure of the diameter of an additive landscape (which, for example, could be defined as the maximum pairwise distance between points on the landscape). Kauffman and Levin have shown that in a large class of correlated random landscapes, the mean distance to the tree component grows only logarithmically with the number of points in the landscape [19].
We utilize two quantitative measures of the predictability of evolutionary trajectories. First is fraction of monotonic paths to the main peak which is computed by counting the number of simple (without reverse substitutions or multiple substitutions at the same site) monotonic paths to the main peak from each point on the landscape, dividing it by the total number of simple paths (where is the Hamming distance from point to the peak), and averaging over the landscape via(1)where is the number of points in the landscape and the sum excludes the main peak. The monotonic path fraction measures the scarcity of accessible evolutionary paths when selection is strong. When the monotonic path fraction is small, evolution is more constrained.
Second, the mean path divergence, is a fine-grained measure of evolutionary (un)predictability. We first define the divergence of a pair of paths and , as the average of the shortest Hamming distances from each point on one path to the other path. Suppose that we have a way of generating stochastic evolutionary paths. The outcome of a large number of evolutionary dynamics simulations is a collection of paths with their associated probabilities of occurrence. In general, the probability of occurrence of an evolutionary path is proportional to the product of fixation probabilities of its constituent mutation steps. Given a bundle of paths with the same starting and ending points, we define its mean path divergence to be(2)where is the probability of occurrence of path in the ensemble. In other words, if two paths were drawn from the bundle at random with probabilities proportional to , their expected divergence would be . Alternatively, if we were to fix one path to be the most likely path in the bundle and to select the second path at random with probability proportional to , the divergence would be proportional to as well.
The six quantitative characteristics of fitness landscapes are summarized in Table 1.
In an additive landscape, the mutational trajectory is maximally ambiguous. As every substitution that brings the sequence closer to the peak increases fitness, substitutions can occur in any order and all shortest mutational trajectories to the peak–without reverse substitutions or multiple substitutions at the same site–are monotonic in fitness. In the strong selection limit of our model defined below, all monotonic trajectories have roughly the same probability of occurrence, so the mutational path cannot be predicted.
The mean path divergence is a better measure of the predictability of evolutionary trajectories than the number or fraction of accessible paths. Even when only a small fraction of paths are monotonic in fitness, these paths could potentially be quite different, perhaps randomly scattered over the landscape. In such a case, prediction of the evolutionary trajectory would be inaccurate despite the scarcity of accessible paths which will be reflected in a high value of path divergence.
Equation (2) introduces the mean path divergence of a bundle of paths with the same starting and ending points. The landscape-wide mean path divergence is measured by constructing representative path bundles with all possible [start, peak] pairs including suboptimal peaks as trajectory termination points. Path divergence is averaged over all bundles with the starting and ending points separated by the same Hamming distance. To construct the path bundles, we employed a low mutation rate model in which the attempted substitutions are either eliminated or fixed in the population before the next mutation attempt occurs.
We invoke the misfolding-cost hypothesis to assign a fitness to a sequence that folds with probability to a particular structure. To produce an abundance of correctly folded copies, an average of of misfolded copies are produced. The “fitness” of a sequence should be a monotonically decreasing function of the cost incurred by the misfolded proteins. Previously we showed that qualitative conclusions drawn from the average population dynamics on the fitness landscape did not depend on the precise functional relationship between the number of misfolded copies and fitness [34]. We use simply the negative of the number of misfolded copies and assign a fitness , to a sequence whose probability of folding to the reference structure is . Because the exact population dynamics model is not important, we use diploid population dynamics in the low mutation rate limit. Therefore, the probability of fixation of a mutant in the background of is given by(3)where is the effective population size [38] which in all simulations was fixed at . The required abundance is a measure of the strength of selection. In the limit of large , the probability of fixation of a beneficial mutation is unity whereas deleterious mutations are never fixed. Since the effective population size is large in our simulations, neutral mutations are almost never fixed either. Because uphill steps in the fitness landscape are equally likely, all monotonic uphill trajectories have equal evolutionary significance.
In the analysis that follows, we study the association between landscape roughness and path predictability for the folding landscapes and their randomized (also referred to as permuted or scrambled) versions. In the scrambled landscapes, the topology (i.e. connectivity) of the landscape is preserved but the fitness values are randomly shuffled. We also compare the roughness and path predictability characteristics of the model and the experimental landscapes for -lactamase [25] and sesquiterpene synthase [28] to those for noisy additive landscapes with a continuously tunable amount of roughness.
Here we examined the fraction of monotonic paths and introduced mean path divergence as quantitative measures of the degree to which the starting and ending points determine the path of evolution on fitness landscapes. The lower the mean path divergence value, the more deterministic (and predictable) evolution is. Global measures of landscape roughness correlate with path divergence in the three analyzed classes of fitness landscapes: additive landscapes perturbed by noise, landscapes derived from our protein folding model and two small empirical landscapes. The folding landscapes are substantially smoother than their permuted counterparts. As a result, although in all analyzed landscapes only a small fraction of the theoretically possible evolutionary trajectories is accessible, this fraction is much greater in the folding and experimental landscapes than it is in randomized landscapes. In addition, the mean path divergence in the randomized landscapes is significantly smaller than in the original landscapes. Thus, the model and empirical landscapes possess similar global architectures with many more diverged monotonic paths to the high peaks than uncorrelated landscapes with the same distribution of fitness values. Consequently, evolution in fitness landscapes is substantially more robust to random mutations and less deterministic (less predictable) than expected by chance. These findings are compatible with the concept that might appear counter-intuitive but is buttressed by results of population genetic modeling, namely, that robustness of evolving biological systems promotes their evolvability [39]–[41]. Additionally, the folding landscapes exhibit a substantial deficit of peaks compared to perturbed additive landscapes and experimental landscapes, a property that translates into a substantially greater fraction of paths leading to the main peak.
When it comes to the interpretation of the properties of fitness landscapes described here, an inevitable and important question is whether the folding model employed here is sufficiently complex and realistic to yield biologically relevant information. In selecting the complexity of our folding model, we attempted to construct the simplest model which exhibits 1) a rich spectrum of low energy conformations across the sequence space, and 2) a non-trivial distribution of substitutions effects on the low energy conformations. An important choice is whether the location of monomers is confined to a lattice or can be varied continuously. When the configuration space is continuous, the distribution of energy barriers between energetically optimal conformations can extend to zero. Therefore, the subtlety of distinctions between conformations can lead to a richer structure of the fitness landscape. We chose not increase the complexity of the model further and treated monomers as point-like particles in a chain where the distance between nearest neighbors is fixed but the angle between successive links in the chain in unrestricted. Our level of abstraction is therefore somewhere between lattice models and all-atom descriptions of proteins [42]–[51].
Another important choice is the number of the model monomer types. Again, we opted for an intermediate level of abstraction and chose four types of monomers: hydrophobic, hydrophilic, and positively and negatively charged. This choice drastically reduces the size of the sequence space while retaining some of the substitution complexity whereby hydrophilic and charged monomers can be swapped under some conditions without radically altering the native state. The intermediate level of abstraction in our approach has its pros and cons. Although the model reproduces key features of protein folding such as the existence of the hydrophobic folding nucleus and two-stage folding kinetics [52], [53], compact conformations certainly do not represent proteins. Rather, we might think of our monomers as representing structurally grouped regions several (perhaps up to a dozen) amino-acids in length. Compact conformations in the model might therefore be analogous to tertiary structures of proteins. Representing sequence space with only four monomer types and treating mutations without reference to the underlying DNA or genetic code does not accurately reflect the natural mutation process. However, our goal was to isolate the features of fitness landscapes which could be traced directly to the constraints imposed by the heteropolymer folding kinetics and energetics. We therefore used a simple sequence space and a homogeneous mutation model to avoid compounding the fitness landscape structure by the complexity derived from the mutation process.
Most importantly, our folding model has been shown to reproduce the observed universal distribution of the evolutionary rates of protein-coding genes as well as the dependencies of the evolutionary rate on protein abundance and effective population sizes [34]. Therefore, despite its simplicity, the behavior of this model might reflect important aspects of protein evolution. In particular, the conclusions drawn from the analysis of the model landscapes exhaustively explored here could also apply to the fitness landscapes of protein evolution. In the previous work, we concluded that the universal distribution of evolutionary rates and other features of protein evolution follow from the fundamental physics of protein folding [34]. The results presented here suggest that the (relative) smoothness and a substantial deficit of peaks in the fitness landscapes of protein evolution that lead to mutational robustness and the ensuing evolvability could similarly follow from the fact that proteins are heteropolymers that have to fold in three dimensions to perform their functions.
The experimental landscapes considered here are decidedly incomplete. Due to experimental limitations, only the analysis of binary substitutions at a handful of sites is feasible at this time. The incompleteness of the empirical landscapes analyzed in this work could be the cause of the observed lack of peak suppression. This proposition will be put to test by the study of larger parts of experimental landscapes that are becoming increasingly available.
The goal of this study is to explore the relationship between roughness and path divergence in realistic fitness landscapes. Our polymer folding model provides a simple way of constructing such landscapes. The model has been described in detail previously [34].
In brief, the model polymer is a flexible chain of monomers in which the nearest neighbors interact via a stiff harmonic spring potential with rest length . The angles between the successive links in the chain are unrestricted. There are four types of monomers: hydrophobic H, hydrophilic P, and charged + and −. Next nearest neighbors and in the chain and beyond interact via a pairwise potential(4)where is the distance between monomers and , is the monomer's charge, is the Debye-Hückel screening length, and and depend on the pair in question. The interaction parameters are chosen to mimic the essential features of the amino-acid interactions. To emulate the effects of solvent, we assign a stronger attraction to the HH pair than to the PP, ++, and −− pairs. There is also a long range repulsion between H and P and even stronger repulsion between H and the charged monomers. The values of the parameters are , Debye-Hückel screening length . The Lennard-Jones coefficients and are(5)Note that a can be substituted by a in the subscripts and the coefficients are symmetric with respect to the interchange of the indices.
The energy of the chain is(6)where the first term is the sum of the pairwise energies given by Eq. (4) over non-nearest neighbor pairs, and the second term reflects the springs connecting nearest neighbors. The spring constant is proportional to temperature . The parameters are fixed for all simulation runs at , and the quench temperature . To mimic the observed tendency of the and termini to be in close proximity, we fixed the endpoint monomers of the model sequences to be of and types.
Dynamics of folding are simulated via over-damped Brownian kinetics which are appropriate when inertial and hydrodynamic effects are not important. Units are chosen so that each component of the 'th monomer's coordinates is updated according to(7)where is the time step and is a random variable with zero mean, variance , uncorrelated with for other times, monomers and spatial directions.
The “native structure” of a particular sequence is represented by an equilibrium ensemble of conformations. The ensemble is constructed by identifying the typical folded conformation and measuring the characteristic RMSD due to thermal fluctuations in the folded state. Three thousand quenches are then performed and the resulting folded conformations are accumulated. The equilibrium ensemble that represents the native structure is defined as the largest cluster of quenched conformations within RMSD distance from each other. Thus, each conformation in the ensemble differs from any other by an amount comparable to the differences introduced by thermal fluctuations alone.
The concept of the native structure ensemble allows us to compute the probability that a sequence folds to a particular structure in a natural, physically plausible fashion. Given a native structure ensemble we assess its conformation space density by computing the distance between each member of the ensemble and its closest neighbor. Given the set of these shortest distances we compute the median and the median absolute deviation (MAD) . A new conformation is deemed to belong to the ensemble if the shortest distance from this conformation to the members of the ensemble is smaller than .
Given a native structure ensemble of some sequence we compute the probability that sequence (which could be itself) folds to the this structure by accumulating equilibrated quenched conformations of and using the above criterion to determine the fraction that belong to the native structure ensemble of . Because sample conformations are computed, the smallest measurable is . The sample size used to measure dictated by the computational demands of the model, introduces a random component to the model fitness landscapes. As we report below, model landscapes turn out to be substantially smoother than random. Therefore the underlying global structure of the model landscapes appears to survive the modest amount of randomness introduced by the relatively small sample size used for measuring .
Robust folders (sequences with a high probability of correct folding) tend to have large linear regions stretched by repulsive Coulomb interactions. Because the linear regions have no contacts with other monomers, we focused our attention on compact conformations with a high monomer contact density. Substitutions in these higher complexity conformations were more likely to exhibit non-trivial effects. To find compact robust folders in the vast available sequence space of -mers (the sequences are of length but the endpoint monomer types are fixed) with monomer types, we implemented a simulated annealing search which optimized the correct folding probability divided by the cube of the native conformation's radius of gyration. The search produced over 800 sequences with and at least two distinct regions of the polymer in mutual contact.
We examined each single substitution mutant of a robustly folding sequence and computed the folding probability to the structure of the original sequence. All mutants with were added to the landscape and if their mutants were also examined. This process is repeated until all mutants of the last sequence under consideration have .
From our study of complete landscapes we estimate that on average for each sequence with which is included into the landscape, roughly 6 others with need to be examined. Since each quench and equilibration takes about 2–4 seconds, landscape construction takes roughly 30 minutes to an hour per included sequence. Thus landscapes larger than 10,000 sequences take months to compile.
At the time of submission, 39 complete landscapes have been constructed, the largest comprising 12969 sequences.
The organization of the folding fitness landscapes and experimental landscapes were compared with perfectly additive landscapes perturbed by noise constructed as follows. Each substitution to the peak fitness sequence was assigned a negative fitness differential drawn at random from an exponential distribution with parameter . The sum over the fitness differentials of a particular set of substitution was modified by either additive of multiplicative noise [54]. Additive noise is drawn from a Gaussian distribution with zero mean and standard deviation which was varied between and The multiplicative perturbation is achieved by multiplying the fitness by a number drawn from a uniform distribution raised to a positive power varied between and When is small, multiplicative factors are close to unity and the perturbation is small as well. If the perturbed fitness was positive, the mutant was included into the landscape. The noise amplitude was varied to obtain a family of landscapes of continuously varying roughness. Only the data for the additive landscapes with multiplicative noise were included in this manuscript. Landscapes perturbed by other types of noise exhibited essentially the same qualitative behavior.
The studies on experimental fitness landscapes typically involve constructing a library of all possible combinations of binary mutations at a small number of sites. The first study included in the present analysis measured the minimum inhibitory concentrations (MIC) of an antibiotic for a complete spectrum of mutants with modified TEM -lactamases; the transition from the antibiotic-sensitive to the antibiotic-resistant form requires five mutation, so the landscape encompassed 120 mutational trajectories between the most distant points on the landscape (or 32 sequences) [25]. The logarithm of MIC was used as the proxy for fitness. In the second study, catalytic activity of 419 sesquiterpene synthase mutants that differed by at most 9 substitutions was measured [28]. We used the catalytic specificity (propensity for producing a particular reaction product rather than a broad spectrum of products) of the mutant enzymes as the proxy for fitness. Before performing the analysis, the fitnesses in the experimental landscapes are mapped onto the interval to enable meaningful quantitative comparisons of the roughness measures.
|
10.1371/journal.pmed.1002534 | HIV treatment eligibility expansion and timely antiretroviral treatment initiation following enrollment in HIV care: A metaregression analysis of programmatic data from 22 countries | The effect of antiretroviral treatment (ART) eligibility expansions on patient outcomes, including rates of timely ART initiation among those enrolling in care, has not been assessed on a large scale. In addition, it is not known whether ART eligibility expansions may lead to “crowding out” of sicker patients.
We examined changes in timely ART initiation (within 6 months) at the original site of HIV care enrollment after ART eligibility expansions among 284,740 adult ART-naïve patients at 171 International Epidemiology Databases to Evaluate AIDS (IeDEA) network sites in 22 countries where national policies expanding ART eligibility were introduced between 2007 and 2015. Half of the sites included in this analysis were from Southern Africa, one-third were from East Africa, and the remainder were from the Asia-Pacific, Central Africa, North America, and South and Central America regions. The median age of patients enrolling in care at contributing sites was 33.5 years, and the median percentage of female patients at these clinics was 62.5%. We assessed the 6-month cumulative incidence of timely ART initiation (CI-ART) before and after major expansions of ART eligibility (i.e., expansion to treat persons with CD4 ≤ 350 cells/μL [145 sites in 22 countries] and CD4 ≤ 500 cells/μL [152 sites in 15 countries]). Random effects metaregression models were used to estimate absolute changes in CI-ART at each site before and after guideline expansion. The crude pooled estimate of change in CI-ART was 4.3 percentage points (95% confidence interval [CI] 2.6 to 6.1) after ART eligibility expansion to CD4 ≤ 350, from a baseline median CI-ART of 53%; and 15.9 percentage points (pp) (95% CI 14.3 to 17.4) after ART eligibility expansion to CD4 ≤ 500, from a baseline median CI-ART of 57%. The largest increases in CI-ART were observed among those newly eligible for treatment (18.2 pp after expansion to CD4 ≤ 350 and 47.4 pp after expansion to CD4 ≤ 500), with no change or small increases among those eligible under prior guidelines (CD4 ≤ 350: −0.6 pp, 95% CI −2.0 to 0.7 pp; CD4 ≤ 500: 4.9 pp, 95% CI 3.3 to 6.5 pp). For ART eligibility expansion to CD4 ≤ 500, changes in CI-ART were largest among younger patients (16–24 years: 21.5 pp, 95% CI 18.9 to 24.2 pp). Key limitations include the lack of a counterfactual and difficulty accounting for secular outcome trends, due to universal exposure to guideline changes in each country.
These findings underscore the potential of ART eligibility expansion to improve the timeliness of ART initiation globally, particularly for young adults.
| In 2009 and 2013, the World Health Organization (WHO) recommended that HIV patients with CD4 counts ≤350 and ≤500 cells/μL, respectively, initiate antiretroviral treatment (ART).
The expansion of ART eligibility criteria has the potential to increase ART initiation rates, especially among healthier patients; however, it could also lead to “crowding out” of persons with more advanced disease and lower rates of ART initiation among these patients.
While many countries have adopted WHO guidelines, the impact of ART eligibility expansions on timely ART initiation has not been studied on a large scale.
We examined the changes in timely ART initiation after national ART eligibility criteria were expanded to CD4 ≤ 350 and/or CD4 ≤ 500 in 22 countries, using data on 284,740 adult ART-naïve patients at 171 sites in the International Epidemiology Databases to Evaluate AIDS (IeDEA) network.
Site-level cumulative incidence of ART initiation (CI-ART) within 6 months of enrollment increased by 4.3 percentage points after national ART eligibility expansion to CD4 ≤ 350 and by 15.9 percentage points after expansion to CD4 ≤ 500.
At the individual level, increases were greatest among patients 16–24 years old at enrollment and those newly eligible for ART. No change or small improvements in CI-ART were also observed among patients already eligible for ART before eligibility expansion.
At the site level, sites with the lowest initial levels of CI-ART experienced the greatest increases following guideline expansions.
Overall, ART eligibility expansions were followed by appreciable improvements in timely ART initiation.
Many clinics can support ART initiation among newly eligible patients with less advanced disease without negatively affecting ART initiation rates among those with more advanced disease.
These findings illustrate the potential of ART eligibility expansion to improve the timeliness of ART initiation and patient outcomes along the care cascade globally, particularly for younger adults, in support of the Joint United Nations Programme on HIV/AIDS (UNAIDS) 90-90-90 targets, thereby reducing morbidity, mortality, and onward HIV transmission.
| Individual and population-level benefits of antiretroviral treatment (ART) are maximized when treatment is initiated soon after HIV infection occurs [1–5]. ART slows the progression of HIV to AIDS and reduces infections and mortality among those living with HIV [6,7]. In addition, ART lowers viral load in HIV-infected individuals, thereby reducing the risk of onward HIV transmission [2,7–11]. Recent studies have shown that immediate versus delayed initiation of ART reduces risks of AIDS and severe opportunistic illnesses [12]. Some studies suggest that earlier initiation of ART may also improve rates of retention in care and medication adherence, and promote more rapid achievement of viral load suppression [13–15]. As a result of this evidence, the World Health Organization (WHO) now recommends immediate initiation of ART for all people diagnosed with HIV, regardless of CD4 count [16].
A systematic review of retention in HIV care between testing and treatment among patients who were ineligible to initiate ART at the time of diagnosis estimated that less than one-third of patients remained continuously engaged in HIV care until the point of ART eligibility [17]. Other reviews of losses along the HIV care continuum in sub-Saharan Africa have estimated that 54%–69% of those ineligible for ART at the time of HIV care enrollment are lost to care prior to ART initiation, with the poorest retention rates among adolescents, young adults, and those enrolling in care at earlier stages of infection [18,19]. In contrast, among patients eligible for ART at the time of enrollment into care, the proportions lost to care prior to ART initiation were 25%–36% [18,19]. In a recent analysis from South Africa, which compared outcomes among patients just below and just above the ART eligibility cutoff, patient eligibility for immediate treatment was associated with a 25 percentage point (pp) increase in ART initiation and an 18 pp increase in 12-month retention [20].
Using longitudinal data on patients enrolling in HIV care from the International Epidemiology Databases to Evaluate AIDS (IeDEA) network, we sought to assess changes in timely ART initiation at the original site of care enrollment following major expansions in HIV treatment guidelines across multiple countries and regions and to identify factors associated with these changes.
We conducted a systematic search for current and historical ART eligibility guidelines of countries participating in IeDEA, including sources such as the Joint United Nations Programme on HIV/AIDS (UNAIDS) Database of National HIV Guidelines [24], the International Association of Providers of AIDS Care (IAPAC) Global HIV Watch [25], the HIV Treatment Guideline Database, health ministry websites, search engines, and published literature, including ART guideline reviews [26–28]. For countries with no publicly available ART guidelines, we obtained information from in-country HIV clinicians, researchers, and ministry of health officials. We identified the time points when each country expanded ART eligibility to include asymptomatic people living with HIV (PLWH) with CD4 counts ≤350 cells/μL and ≤500 cells/μL. We also identified concurrent expansions of guidelines related to expanded treatment for pregnant women and patients coinfected with tuberculosis, regardless of other eligibility criteria (i.e., CD4 counts or clinical staging). If the exact month of ART eligibility expansion was unknown, a midyear value was used.
This cohort study utilized deidentified data approved for use by local ethical committees in each of the IeDEA regions included in the analysis.
Sites eligible for inclusion had to have HIV care data available on patients for the period after care enrollment but preceding ART initiation (pre-ART data), as well as data on both ART initiators and noninitiators. Patients had to be at least 16 years old at enrollment (18 in North America) and have at least 12 months of possible follow-up between enrollment and database closure. Patients were excluded if they transferred to an IeDEA site from another clinic, lacked a clinic visit and lab record, or were known not to be ART naïve at enrollment (with ART defined as any regimen of at least 3 antiretroviral drugs, excluding treatment taken for prevention of maternal-to-child transmission).
To ensure that there was no differential selection into the sample by ART eligibility, distributions of CD4 counts at enrollment were assessed with respect to possible discontinuities at the point of ART eligibility (CD4 = 350 and CD4 = 500 for the 2 analyses), which would indicate potential exclusion of non-ART-eligible patients from contributing sites’ cohorts (S1 Text).
The outcome of interest was site-level absolute pp change in cumulative incidence of ART initiation (CI-ART) within 6 months of enrollment at each site—namely, the difference in CI-ART between those enrolling in HIV care 6–18 months prior to national ART eligibility expansion (“pre” or baseline period) and those enrolling 6–18 months thereafter (“post” period). Pp changes were computed separately for ART eligibility expansions to CD4 counts ≤350 cells/μL and ≤500 cells/μL. Patients enrolling in the 6 months immediately preceding and following guideline introduction were excluded to allow for early or lagged implementation of national guidelines at the site level and to ensure that the 6-month CI-ART outcome for the “pre” period occurred before the guideline change. Additionally, to ensure sufficient data for reliable outcome estimation, sites that had fewer than 30 study eligible patients who enrolled in either of the “pre” and “post” periods bracketing ART eligibility expansion were excluded.
To derive the outcome measure, we estimated overall 6-month CI-ART separately for the “pre” and “post” ART eligibility expansion periods via competing risks regression with the Aalen-Johansen estimator, treating death and pre-ART loss to clinic as competing events. To minimize the risk of misclassifying temporarily disengaged patients as lost, pre-ART loss to clinic was defined as not returning to clinic for at least 12 months, with no evidence of subsequent return.
The CI-ART was estimated via competing risk methods, with the Aalen-Johansen estimator, and treating death and loss to clinic as competing risks, as outlined in a July 2016 concept proposal to the IeDEA Executive Committee (S2 Text), which also prespecified inclusion and exclusion criteria, outcome definitions, and key stratification variables. As discussed in the appendix (S2 Text), other analytic approaches, including the use of site-level metaregression and the pre-post within-site comparisons to derive the outcome, were selected after data became available.
To assess variability of change in 6-month CI-ART across sites, forest plots were used to visualize site-level outcomes, and between-site variances (tau-squared) and p-values from chi-squared tests of between-site heterogeneity were calculated.
We calculated pooled summary estimates of the pp change in 6-month CI-ART across all sites for each major expansion in treatment eligibility criteria, along with corresponding 95% confidence intervals (CIs). We stratified all estimates across several site and patient characteristics. Site characteristics included IeDEA region (see “Data Sources”), site setting (urban versus rural), and site-level summary measures of the patient population enrolling at the site, including overall proportion of female clients (≤45%, 45%–65%, >65%); baseline 6-month CI-ART (grouped in quartiles) among all patients and among those previously eligible for treatment as a site-level measure of the preexisting service level and capacity to expand; and proportion of patients eligible for treatment in the “post” period, under expanded guidelines, as a measure of unmet need for treatment. Patient characteristics included sex, age at enrollment (dichotomized into 16–24 years and ≥25 years, to align with the 15–24 age category that is commonly used by WHO [29]), and CD4 count at enrollment. CD4 count at enrollment was categorized to reflect patients’ ART eligibility status vis-à-vis each guideline change: newly eligible (CD4 count 201–350) and previously eligible (CD4 ≤ 200) for the ART eligibility expansion to CD4 ≤ 350 cells/μL and newly eligible (CD4 351–500) and previously eligible (CD4 ≤ 350), in analyses of ART eligibility expansion to CD4 ≤ 500 cells/μL.
Linear random effects metaregression models were used to examine factors associated with changes in 6-month CI-ART while controlling for concurrent expansions of other ART eligibility criteria and to assess the predicted magnitude of change in 6-month CI-ART following ART eligibility expansions. Bivariate and multivariable associations between the pp change in 6-month CI-ART and site and patient population characteristics were examined through linear random effects metaregression, adjusted for concurrent expansions of ART eligibility criteria for pregnant women and tuberculosis patients. To meet the assumption of linearity in the association between continuous independent variables and the outcome, cohort size was log transformed.
Covariates used in metaregression models included variables related to site characteristics (i.e., baseline 6-month CI-ART, baseline median enrollment CD4, median age at enrollment, and proportion of female patients at a site), as well as binary variables related to concurrent guideline changes extending HIV treatment to all pregnant women and tuberculosis patients. Site-level covariates were derived separately for each of the 2 ART eligibility expansions (to CD4 ≤ 350 cells/μL and to CD4 ≤ 500 cells/μL), based on each site’s patient population in the periods bracketing each ART eligibility expansion.
Analyses were completed in Stata 14.
Overall, 260 IeDEA sites had provided longitudinal data on at least 60 patients—the minimum for potential inclusion in the analysis. Thirty sites were further excluded because of lack of data on pre-ART patients. Among the remaining 230 sites, 225 were in countries where eligibility expansions of interest occurred during the period under study. Among these sites, 171 had sufficiently large cohorts for inclusion in this analysis (i.e., at least 30 patients meeting the eligibility criteria described above in each 12-month period bracketing guideline expansion).
The analysis of ART eligibility expansion to treat asymptomatic PLWH with CD4 count ≤350 cells/μL included 145 sites (with 169,717 patients) in 22 countries where guidelines changed between 2007 and 2012. Six of these countries revised national treatment eligibility guidelines before WHO’s November 2009 recommendation [30] to initiate treatment at CD4 ≤ 350, by a mean of 15 months (range 5–27 months). Among the 16 countries that adopted expanded guidelines after WHO’s recommendation, the mean time elapsed was 20 months (range 6–31 months).
The analysis of ART eligibility expansion for asymptomatic PLWH with CD4 count ≤500 cells/μL included 152 sites (with 128,552 patients) in 15 countries where guidelines changed between 2009 and 2015. Three of these countries introduced new guidelines prior to WHO’s June 2013 recommendation [31] on treatment eligibility (mean of 28 months; range 11–42 months). Among the remaining 12 countries, national guideline change followed WHO recommendations by a mean of 9 months (range 1–24 months).
For both analyses, half of the sites were from Southern Africa (50% and 51%, respectively; Table 1).
During the periods preceding ART eligibility expansions to CD4 ≤ 350 and CD4 ≤ 500, the median proportion of patients initiating ART within 6 months of enrollment was 53.3% and 57.5% of patients, respectively (Table 2).
Patient demographics were similar across sites, with a median age of 34 years and female patients comprising more than 60% of patients on average. Baseline median CD4 counts at enrollment into HIV care were higher among sites included in the analysis of CD4≤500 ART eligibility expansion than among the CD4≤350 ART eligibility expansion sites (273 cells/μL versus 223 cells/μL). However, a larger proportion of patients were missing data for CD4 count at enrollment during both the “pre” and “post” periods in the analysis of the second ART eligibility expansion (median 33% versus 14% and 39% versus 20%, respectively; Table 1).
Marked changes in CI-ART were observed after each expansion in treatment eligibility: 4.3 pp (95% CI 2.6–6.1 pp) after the expansion to CD4 ≤ 350 cells/μL and 15.9 pp (95% CI 14.3–17.4 pp) after the expansion to CD4 ≤ 500 cells/μL (Table 2). There was considerable heterogeneity in site-level changes in CI-ART for both guideline expansions, and heterogeneity chi-squared tests were statistically significant (both p <0.001). However, a much larger proportion of sites experienced positive changes in 6-month CI-ART following the latter ART eligibility expansion to CD4 ≤ 500 cells/μL (Fig 1A and 1B). Overall, 33% of sites (48/145) had statistically significant increases in 6-month CI-ART after the expansion to CD4 ≤ 350 cells/μL, 57% of sites (83/145) had no statistically discernible change, and 10% (14/145) had statistically significant decreases. After expansion to CD4 ≤ 500 cells/μL, 74% of sites (112/152) had statistically significant increases, 26% of sites (39/152) had no statistically discernible change, and <1% (1/152) had a statistically significant decrease.
In stratified analyses of the guideline expansion to treat patients with CD4 ≤ 350 cells/μL, changes in 6-month CI-ART were most pronounced at sites in Asia-Pacific, East Africa, and Central Africa (12.0, 7.6 and 7.3 pp, respectively), at sites with higher proportions of newly eligible patients in the “post” period (4.9 pp and 7.9 pp for clinics in the top 2 quartiles of the proportion of patients newly eligible for treatment, respectively), and sites in the lowest quartile of “pre” period 6-month CI-ART (11.8 pp). There was also a trend in the association between lower “pre” period 6-month CI-ART among previously eligible patients specifically and greater site-level change in CI-ART after guideline expansion (Table 2).
Following the CD4 ≤ 500 expansion, the largest changes in CI-ART were observed in Southern and East Africa (17.6 pp and 15.8 pp, respectively). As observed in the CD4 ≤ 350 analysis, there was a trend of greater change in 6-month CI-ART with lower “pre” period 6-month CI-ART (overall and among previously eligible patients specifically); sites in the lowest quartile of “pre” period 6-month CI-ART had the largest changes in CI-ART following ART eligibility expansion (26.2 pp). However, in contrast to trends observed following expansion to CD4 ≤ 350, there was no trend in CI-ART across quartiles of proportion of newly eligible patients in the “post” period (Table 2).
In analyses stratified by patient characteristics, no significant differences in the pp change in CI-ART were observed by sex or age after expansion to CD4 ≤ 350 cells/μL. However, after expansion to CD4 ≤ 500 cells/μL, appreciably larger increases were observed among women when compared to men (16.8 versus 13.7 pp) and among patients aged 16–24 years at enrollment, compared to those 25 years or older (21.5 versus 15.5 pp). Striking differences were also observed in pooled estimates of change in CI-ART by CD4 count at enrollment. Specifically, while no or minimal improvement was observed among patients who were already ART eligible before guideline expansion, a large increase was noted among those newly eligible for treatment under the expanded ART criteria (18.2 pp among newly eligible patients after expansion to CD4 ≤ 350 cells/μL and 47.4 pp among newly eligible patients after expansion to CD4 ≤ 500 cells/μL; Table 2, Fig 2).
Six-month CI-ART during the period before ART eligibility expansion was the strongest correlate of change in CI-ART following eligibility expansion. In bivariate analyses, the predicted change in 6-month CI-ART after expansion to CD4 ≤ 350 was 9.1 pp for a site with “pre”-period CI-ART of 40% and 1.4 pp for a site with “pre”-period CI-ART of 60%. After expansion to CD4 ≤ 500, the predicted changes in CI-ART were 24.1 pp and 14.6 pp, respectively. In adjusted metaregression models, “pre”-period 6-month CI-ART was the strongest correlate of pp change in 6-month CI-ART after ART eligibility expansion (Table 3). Adjusted changes in CI-ART following both ART eligibility expansions were greatest among those sites with lower pre-expansion levels of CI-ART; each 10-unit decrease in the “pre”-period 6-month CI-ART was associated with an average 3.9 and 5.3 pp change in CI-ART after CD4 ≤ 350 and CD4 ≤ 500 cells/μL eligibility expansions, respectively. Increasing cohort size was also positively associated with change in CI-ART (Table 3).
Metaregression models adjusting for site, patient population, and guideline characteristics explained 24.1% and 38.1% of between-site variability, with tau-squared reduced from 126.1 to 95.7 and from 110.4 to 68.3 in the CD4 ≤ 350 and CD4 ≤ 500 analyses, respectively.
Our study found that expansions in ART eligibility criteria for people with HIV infection were followed by substantial increases in rates of timely ART initiation at the original clinic of enrollment, with much larger increases among those newly eligible under the expanded guidelines compared with those eligible under prior guidelines. Importantly, with expansion to CD4 ≤ 500, increases were most pronounced for persons aged 16–24 years, a critical population that has historically been particularly difficult to engage in care. For both ART eligibility expansions examined (i.e., to CD4 ≤ 350 and CD4 ≤ 500), changes in 6-month CI-ART were greatest at sites that had lower rates of ART initiation prior to ART eligibility expansion and, in metaregression analyses, “pre”-period rates of ART initiation were significantly and inversely associated with changes in CI-ART following guideline change. Larger increases in CI-ART overall were observed following the adoption of CD4 count ≤500 cells/μL treatment eligibility guidelines, with statistically significant increases in CI-ART observed at 74% of sites and a dramatic shift in the pooled estimate of CI-ART across all sites included in the analysis.
Concern has been expressed that expanding ART eligibility may result in sick patients being “crowded out” by asymptomatic patients [32]. Importantly, and consistent with research from South Africa [33,34], increases in ART initiation among newly eligible patients on average were not offset by decreases in ART initiation rates among previously eligible patients, meaning that access to treatment by sicker patients with lower CD4 counts did not decrease when patients with higher CD4 counts became eligible and started treatment. In fact, sites with the lowest baseline ART initiation rates among previously eligible patients saw the greatest overall increases in ART initiation after guideline expansion, which also suggests that broadened treatment eligibility did not compromise access for sicker patients.
Another major finding of our study is that increases in timely ART initiation were largest among young adults—a group that has been consistently shown to have lower rates of retention in care both prior to and after ART initiation, as well as lower rates of ART initiation at early stages of infection [35–38]. Improving timely ART initiation among young adults is particularly important, given that this age group accounts for almost half of new HIV infections among adults [39].
To our knowledge, no analyses to date have assessed differences in ART initiation following major changes in HIV treatment eligibility criteria across multiple countries and regions. While the results of this analysis are promising, further research is needed to assess how increases in ART initiation at higher CD4 counts and among asymptomatic patients affect retention in the HIV care continuum over the longer term, including ART adherence, and sustained HIV virologic suppression. While some studies have suggested increased loss to follow-up on ART among patients with higher CD4 counts [40,41], evidence increasingly supports improved retention among patients initiating with higher CD4 counts [15], including those immediately eligible for ART versus those just under the eligibility threshold [20], and same-day (or “rapid”) ART initiators [42], with any increases in loss to follow-up more than offset by eliminating opportunity for pre-ART attrition [43].
The results of this study are important, given a recent systematic review of national HIV treatment cascades, which indicated that the most common “break point” in the HIV care cascade is between diagnosis and initiation of ART [10], which is critical for achieving optimal HIV care outcomes at the individual and population level [44]. In conjunction with efforts to improve HIV testing and linkage to care [45,46], in part via point-of-care CD4 testing to assess patients’ ART eligibility [47], accelerating the adoption of expanded treatment eligibility criteria, including WHO’s current “Treat All” recommendation, could help reduce rates of pre-ART loss to clinic and increase rates of timely ART initiation. Mathematical models suggest that improving rates of HIV testing, linkage, and immediate ART initiation, combined with other prevention interventions, could nearly stop HIV transmission [48].
At the country level, however, treatment guidelines often lag behind WHO guidance, creating missed opportunities to realize the preventive potential of ART [26,27]. A recent analysis of HIV treatment policy changes in sub-Saharan Africa found that the average time for the adoption of the 2013 guidelines was 10 months (range 0 to 36 months) [26]. Compounding delays in the national adoption of global guidance are lags in the practical implementation of new treatment policies at the service delivery level, where outdated practices among frontline health workers, treatment readiness requirements, and resource constraints lead to further delays in ART initiation among eligible HIV patients [49]. Confirming the importance of service delivery factors in promoting more timely ART initiation, a large intervention study in Uganda recently demonstrated a substantial increase in the proportion of eligible patients initiating ART within 14 days of eligibility determination from 38% to 80% (a 42 pp increase) through relatively simple and low-cost changes at the health facility level [50]. Our study results highlight the importance of promoting ART eligibility expansions, including universal test-and-treat guideline scenarios, with complementary strategies at the facility level to maximize timely diagnosis and ART initiation rates among ART-eligible persons.
Major strengths of this study include the analysis of data from a large number of patients and sites across several world regions. Patient- and site-level characteristics were used to provide setting-specific estimates and to adjust for some differences in patient mix across sites. Where known, concurrent expansions of ART eligibility criteria regarding pregnancy and tuberculosis status were also controlled for, to account for varied scopes and foci of guideline changes. Findings from this large and heterogeneous sample of real-world service delivery settings may be generalizable to many contexts, especially where substantial room remains to expand ART eligibility guidelines. Additionally, the use of a 12-month buffer period around the national guideline change date helped both to reduce exposure misclassification and to control for differences in patient mix and seasonal patterns of ART initiation in within-site pre-post comparisons.
However, even with the use of a buffer period, the lack of data on the exact timing of site-level implementation of expanded ART eligibility guidelines likely contributed to imprecision of the site-level CI-ART estimates, as implementation of expanded guidelines may have preceded national adoption at some sites and lagged at other sites. Furthermore, the inclusion of 171 sites from 22 countries necessarily means considerable heterogeneity of ART referral and record-keeping practices, which may have also affected the precision of the estimates. With half of all sites and over 60% of patients in this analysis being from Southern Africa, overall estimates are strongly influenced by this region. However, changes in CI-ART were similar across regions, especially in the expansion to CD4≤500 analysis, albeit with some regions having lower volumes of data (Asia-Pacific and South and Central America) and therefore less precise CI-ART estimates.
While many sites experienced increases in timely ART initiation, particularly after treatment eligibility was expanded to CD4 ≤ 500 cells/μL, no change in CI-ART and occasional decreases in CI-ART were also observed, possibly as a result of unmeasured site-level factors (e.g., medication or reagent shortages, staffing changes, and other resource constraints). We did not have data on many facility- and patient population-related factors that could help illuminate the reasons behind observed decreases in ART initiation at selected sites after country-level guideline expansion. In addition, random fluctuation or regression to the mean may help explain some decreases and increases (for example, changes observed among sites with lowest “pre”-period ART initiation rates).
Nuanced expansions in national ART eligibility criteria related to clinical stage and pregnancy and tuberculosis status, as well as special populations, such as sex workers and serodiscordant couples, could not be reflected in this analysis. Furthermore, assumptions of a midyear policy change for countries where the exact month of ART eligibility expansion was unknown, in combination with possible early or delayed site-level implementation of national guidelines, may have obscured the effects of guideline change at the site level.
The increases in CI-ART that we observed may also overestimate the true impact of ART eligibility expansions, given our inability to directly account for secular trends in CI-ART and lack of a counterfactual within countries during the periods bracketing ART eligibility expansions. However, the change in timely ART initiation among those previously eligible after the guideline expansion is likely a good proxy for what would have happened to those newly eligible under the guideline expansion had the guideline not changed. The larger increase in timely ART initiation among those newly eligible (18.2 pp for CD4 ≤ 350 and 47.4 pp for CD4 ≤ 500) versus those previously eligible (no change for CD4 ≤ 350 and a small increase of 4.9 pp for CD4 ≤ 500) is more consistent with a guideline effect than a secular trend.
While our study design precludes causal attribution, recent research in South Africa, including studies that used regression discontinuity analysis to account for secular trends prior to and following eligibility expansions [33,45], as well as to examine ART initiation rates just above and just under eligibility threshold [20], lend support for a causal association between ART eligibility expansions and increases in ART initiation.
Our study of a global sample of HIV clinics indicates that ART eligibility expansion was followed by substantial increases in timely ART initiation at the original site of enrollment, including among young adults aged 16–24 years, as well as at clinics with lower rates of timely ART initiation prior to treatment eligibility expansions. That many clinics can support the initiation of ART among newly eligible patients with less advanced disease without an adverse impact on ART initiation rates among those with more advanced disease may reflect a transition away from the emergency treatment era in some locations. While we did not have sufficient data available to assess the changes in CI-ART following the implementation of the currently recommended Treat All approach, these findings underscore the utility of ART eligibility expansion as an essential strategy in support of the UNAIDS 90-90-90 targets globally. As countries increasingly adopt WHO’s Treat All recommendation, future research should explore its impact on timely ART initiation, as well as other longer-term HIV care outcomes, including retention, ART adherence, and sustained virologic suppression.
The content of this publication is solely the responsibility of the authors and does not necessarily represent the official views of any of the governments or institutions mentioned above.
|
10.1371/journal.ppat.1005890 | XRN1 Is a Species-Specific Virus Restriction Factor in Yeasts | In eukaryotes, the degradation of cellular mRNAs is accomplished by Xrn1 and the cytoplasmic exosome. Because viral RNAs often lack canonical caps or poly-A tails, they can also be vulnerable to degradation by these host exonucleases. Yeast lack sophisticated mechanisms of innate and adaptive immunity, but do use RNA degradation as an antiviral defense mechanism. We find a highly refined, species-specific relationship between Xrn1p and the “L-A” totiviruses of different Saccharomyces yeast species. We show that the gene XRN1 has evolved rapidly under positive natural selection in Saccharomyces yeast, resulting in high levels of Xrn1p protein sequence divergence from one yeast species to the next. We also show that these sequence differences translate to differential interactions with the L-A virus, where Xrn1p from S. cerevisiae is most efficient at controlling the L-A virus that chronically infects S. cerevisiae, and Xrn1p from S. kudriavzevii is most efficient at controlling the L-A-like virus that we have discovered within S. kudriavzevii. All Xrn1p orthologs are equivalent in their interaction with another virus-like parasite, the Ty1 retrotransposon. Thus, Xrn1p appears to co-evolve with totiviruses to maintain its potent antiviral activity and limit viral propagation in Saccharomyces yeasts. We demonstrate that Xrn1p physically interacts with the Gag protein encoded by the L-A virus, suggesting a host-virus interaction that is more complicated than just Xrn1p-mediated nucleolytic digestion of viral RNAs.
| Like other eukaryotes, Saccharomyces cerevisiae is chronically infected with viruses. It is fascinating to consider how S. cerevisiae deals with viral infection, because yeast have limited mechanisms of immunity. Our paper focuses on Xrn1p, an enzyme that is important for the destruction of irregular cellular RNAs in all eukaryotic cells. Xrn1p also degrades viral RNAs, owing to the fact that viral RNAs share biochemical characteristics with aberrant cellular mRNAs. Xrn1p was previously known to efficiently control the replication of a S. cerevisiae virus called “L-A.” We find that two different L-A viruses of Saccharomyces yeasts are best controlled by the Xrn1p from their own host species compared to the Xrn1p from other species. Importantly, these Xrn1p from different species are functionally equivalent in all other ways. This would suggest that while the important cellular functions of Xrn1p have been conserved over millions of years, the interaction with L-A-like viruses has been dynamic and constantly redefined by evolution. The identification of species-specific host proteins, like Xrn1p, is recently being appreciated as a key criterion for understanding why viruses infect the species that they do.
| Degradation of mRNAs is a process essential to cell viability. Degradation pathways eliminate aberrant mRNAs, and also act to control gene expression levels. This process typically begins with host enzymes that perform either deadenylation or decapping on mRNAs targeted for degradation [1]. Following decapping, mRNAs are typically degraded by the 5’ to 3’ cytoplasmic exonuclease, Xrn1 [2,3]. Alternatively, after deadenylation, mRNAs can be subject to 3’ to 5’ degradation by the cytoplasmic exosome [4–6]. Viral transcripts and viral RNA genomes usually do not bear the canonical 5’ methylated cap structures or the 3’ polyadenylated (poly(A)) tails typical of cellular mRNAs, making them vulnerable to destruction by these host mRNA degradation pathways. In fact, it has been observed that Xrn1 and components of the exosome efficiently restrict virus replication in eukaryotes as diverse as mammals and yeasts [7–11]. As a result, mammalian viruses have evolved diverse countermeasures to prevent degradation by these proteins [7,8,12–18]. Still unknown is whether the host proteins like Xrn1 and components of the exosome can co-evolve with viruses to circumvent viral countermeasures. While such tit-for-tat evolution is common in mammalian innate immunity pathways, mRNA degradation is essential to the host and would be expected to be subject to strong evolutionary constraint.
Saccharomyces yeasts are known to harbor very few viruses [19]. Further, all yeast viruses are unable to escape their host cell, and instead are transmitted through mating or during mitotic cell division. Almost all described species of Saccharomyces yeasts play host to double-stranded RNA (dsRNA) viruses of the family Totiviridae [20,21]. In fact, most commonly used S. cerevisiae laboratory strains are infected with a totivirus named L-A (Fig 1A) [22]. When initially synthesized, the RNAs produced by the L-A virus RNA-dependent RNA polymerase lack both a cap structure [23,24] and a poly(A) tail [25], and are vulnerable to degradation by yeast Xrn1 (denoted Xrn1p) [8] and the cytoplasmic exosome [14,26]. 3’-to-5’ degradation of viral RNAs by the cytoplasmic exosome is linked to the action of the SKI complex (Ski2, Ski3, Ski7, and Ski8), which acts to funnel aberrant RNAs into the nucleolytic core of the exosome [5,6]. The disruption of exosome and SKI complex genes has been shown to cause higher expression of viral RNAs, higher virus genome copy number, and an overproduction of virus-encoded toxins (i.e. the “superkiller” phenotype) [11,14,27]. In addition, the 5’-to-3’ exonuclease Xrn1p degrades viral transcripts and genomes of several RNA viruses in yeasts [8,24,28].
Viruses and their hosts exist in a constant state of genetic conflict, where what is advantageous for one party is often disadvantageous for the other. Both genomes experience selection for mutations that benefit their own fitness but, particularly in yeast where viruses are strictly intracellular, the virus will be bounded in this process by the fact that if it begins to replicate too well, it may kill its host. Co-evolutionary battles between hosts and viruses play out in the physical interaction interfaces between interacting host and viral proteins (reviewed in [30–32]). One party is selected to reduce these interactions, and the other party is selected to strengthen them. For instance, there are many examples showing that mammalian restriction factors are selected to better recognize their viral targets, while viruses are continuously selected to escape that interaction, or to encode an antagonist protein that neutralizes the restriction factor. Because there is often no stable equilibrium in these systems, this process of tit-for-tat evolution between host and virus can cycle over and over, causing unusual signatures of evolution in both the host and virus proteins engaged in this interaction. While host protein complexes (host proteins interacting with other host proteins) can sometimes become co-evolved, this process of within-species refinement of protein-protein interactions is not the same as the dynamic and recurrent selection for new amino acids at interaction interfaces between host and pathogen proteins. The two scenarios can be disentangled using a metric that looks for codons that have accumulated a significantly higher rate of nonsynonymous mutations (dN) than even synonymous mutations (dS). The signature of dN/dS > 1 commonly results from the repeated cycles of selection that occur in genetic conflict scenarios [33], but has not been shown to be driven by subtler processes like the refinement of within-host physical interactions. Highly diverged host proteins reinforce species barriers, making it difficult for viruses to move from their current host species into new host species (for example, [34,35]). Since yeast have limited antiviral strategies, we reasoned that evolutionary pressure on the RNA quality control pathways to thwart the replication of RNA viruses might be especially intense. This led us to investigate the unique evolutionary scenario involving a restriction system employing proteins critical to RNA turnover and cellular homeostasis.
In this study, we analyzed whether or not any components of the yeast RNA degradation pathways mentioned above are evolving under positive natural selection, potentially indicative of tit-for-tat coevolution with viruses. We identified this evolutionary signature in at least two genes involved in RNA metabolism, RRP40 and XRN1, and then undertook an in depth functional analysis of XRN1. To test the hypothesis that Xrn1p has been honed by co-evolution to target and restrict totiviruses, we made a series of S. cerevisiae strains where XRN1 is replaced with wild-type orthologs from other Saccharomyces species. All XRN1 orthologs fully complemented an XRN1 knockout strain of S. cerevisiae, as assessed by several assays. On the other hand, we found that XRN1 orthologs were different in their ability to control the replication of the L-A virus. Xrn1p from S. cerevisiae was most efficient at controlling the L-A virus that chronically infects S. cerevisiae, and Xrn1p from S. kudriavzevii was most efficient at controlling the L-A-like virus (SkV-L-A1) that we discovered within S. kudriavzevii. All XRN1 orthologs were equivalent in their interaction with another virus-like parasite, the Ty1 retrotransposon. Our identification of signatures of positive selection and species-specific virus restriction suggests that XRN1 can be tuned by natural selection to better restrict totivirus in response to the evolution of these viruses over time. We show that the structure of Xrn1p affords the flexibility to change in response to selective pressure from totiviruses, while also maintaining cellular functions.
We first looked for evidence of positive selection (dN/dS > 1) within the genes encoding the major components of the SKI complex, the exosome, and Xrn1p (Fig 1B). Importantly, signatures of positive selection do not identify the genes that are most important for controlling viral replication. Rather, these statistical tests are designed to identify host proteins that are involved in direct physical interactions with viruses, and which also have the evolutionary flexibility to change in response to viral selective pressure, becoming species-specific in the process. For this reason, we would neither expect to identify signatures of positive selection in all genes known to be involved in controlling totiviruses, nor in all genes encoding components of a complex like the exosome. For each gene, we collected sequences from six divergent species of Saccharomyces (S. cerevisiae, S. paradoxus, S. mikatae, S. kudriavzevii, S. arboricolus, and S. bayanus) [36–38] and created a multiple sequence alignment. We then analyzed each alignment for evidence of codons with dN/dS > 1 using four commonly employed tests for positive selection [39,40]. We see some evidence for positive selection of specific codon sites in several of these genes, however, only XRN1 and the exosome subunit gene RRP40 passed all four tests (Fig 1C and S1 Table). Other genes are determined to be under positive selection by some tests, and may be of interest to explore further. Of XRN1 and RRP40, the impact of XRN1 on viral replication has been more directly substantiated [7–9,13–16,41–48], so we focused our attention on this gene. However, it should be noted that RRP40 encodes a component of the cytoplasmic exosome, which, in conjunction with the SKI complex, is clearly linked to the restriction of L-A [11,14,27].
We next tested if S. cerevisiae XRN1 has been tailored by co-evolution with the L-A virus. Double-stranded RNA (dsRNA) purified from a S. cerevisiae xrn1Δ strain migrates as a distinct band of 4.6 kilobase pairs (Fig 2A), which is consistent with the size of the L-A virus genome, and its identity was further confirmed by RT-PCR (S1 Fig). We confirmed a strong reduction in dsRNA when the xrn1Δ strain was complemented with plasmid-mounted XRN1 from S. cerevisiae under the transcriptional control of its native promoter (Fig 2A), consistent with the published role of Xrn1p as an L-A restriction factor [8,14,24,27]. On the other hand, catalytically-dead versions of Xrn1p (E176G and Δ1206–1528) did not suppress L-A dsRNA levels (Fig 2B), as has been previously described [49]. We next performed heterospecific (other species) complementation by introducing the XRN1 from S. mikatae, S. kudriavzevii, or S. bayanus into the S. cerevisiae xrn1Δ strain. These species were chosen as they are representative of the diversity found within the sensu stricto complex of Saccharomyces yeasts. Strikingly, no other Xrn1p was able to reduce L-A dsRNA to the same extent as Xrn1p from S. cerevisiae (Fig 2A). Xrn1p from S. mikatae, the closest relative to S. cerevisiae in this species set, was capable of slightly reducing L-A dsRNA abundance. Xrn1p from S. bayanus and S. kudriavzevii appear to have levels of dsRNA similar to xrn1Δ, indicating little or no effect on L-A copy number. In summary, we find that XRN1 orthologs vary in their ability to restrict the S. cerevisiae L-A virus. This is somewhat surprising for a critical and conserved gene involved in RNA quality control, but consistent with the signatures of positive selection which suggest that certain parts of this protein are highly divergent between species.
We next used a functional and quantitative assay to confirm the species-specific effects of XRN1 on virus replication. This assay exploits the dsRNA “killer virus” (also known as M virus). The killer virus is a satellite dsRNA of L-A that is totally dependent on L-A proteins for replication. It uses L-A-encoded proteins to encapsidate and replicate its genome, and to synthesize and cap its RNA transcripts [12]. The killer virus encodes only a single protein, a secreted toxin referred to as the killer toxin [19,50,51]. The result is that “killer yeast” colonies, i.e. those infected with both L-A and the killer virus, kill neighboring cells via the diffusion of toxin into the surrounding medium (Fig 2C). Importantly, resistance to the killer toxin is provided by the pre-processed, immature form of the toxin, supplying killer yeast cells with an antidote to their own poison [50]. It has been shown previously that Xrn1p can inhibit the expression of the killer phenotype by degrading uncapped killer virus RNAs [14,52]. Therefore, we use the presence and size of kill zones produced by killer yeasts as a quantitative measurement of killer virus RNA production in the presence of each Xrn1p ortholog.
A strain of S. cerevisiae lacking XRN1, but harboring both the L-A and killer virus (xrn1Δ L-A+ Killer+), was complemented with each XRN1 ortholog. Clonal isolates from each complemented strain were grown to mid-log phase, and 6 x 105 cells were spotted onto an agar plate seeded with a lawn of toxin-sensitive yeast. After several days’ incubation at room temperature, kill zones around these culture spots were measured and the total area calculated. The transformation of xrn1Δ L-A+ Killer+ with S. cerevisiae XRN1 produced an average kill zone that covered 0.68 cm2 (n = 14). However, transformation with XRN1 from S. mikatae, S. bayanus, or S. kudriavzevii produced significantly larger kill zones covering 0.92 cm2 (n = 11), 0.96 cm2 (n = 17) and 0.97 cm2 (n = 17), respectively. The kill zone produced by xrn1Δ L-A+ Killer+ yeast expressing S. cerevisiae XRN1 was significantly smaller than those produced by yeast expressing any of the other XRN1 orthologs (Tukey—Kramer test, p<0.05) (Fig 2D). The smaller kill zones in the strain expressing S. cerevisiae XRN1 are consistent with lower levels of killer and L-A derived RNAs. In summary, this assay also supports a species-specific restriction phenotype for XRN1.
It has been observed that over-expression of XRN1 can cure S. cerevisiae of the L-A virus, presumably by degrading viral RNA so effectively that the virus is driven to extinction [8,28]. Therefore, we developed a third assay to test the ability of XRN1 orthologs to control L-A, in this case by assessing their ability to cure S. cerevisiae of the virus. Plasmids expressing HA-tagged and untagged Xrn1p were transformed into a killer strain of S. cerevisiae with its genomic copy of XRN1 intact. This was followed by the analysis of more than 100 purified clones for virus curing, that is, the absence of the killer phenotype as indicated by the loss of a kill zone when plated on a lawn of sensitive yeast. Importantly, the introduction of an empty plasmid fails to produce any cured clones (n = 103) (Fig 3A and 3B). Provision of an additional copy of S. cerevisiae XRN1 cured 49% of clones (n = 159) (Fig 3A and 3B). Cured clones remained cured (i.e. non-killers) when purified and tested again for their ability to kill sensitive yeasts (n = 20). Over-expression of XRN1 from S. mikatae, S. kudriavzevii, and S. bayanus was unable to efficiently cure the killer phenotype, resulting in only 12% (n = 129), 8% (n = 120), and 9% (n = 123) cured clones, respectively (Fig 3A, blue bars). The loss of L-A from cured strains was also verified by RT-PCR. We detected no L-A or killer RNAs within the four cured clones analyzed (Fig 3C). These data show that XRN1 from all Saccharomyces species have the ability to cure the killer phenotype, however, XRN1 from S. mikatae, S. kudriavzevii, and S. bayanus are considerably less efficient than S. cerevisiae XRN1. Taken together, we show that viral restriction by XRN1 is species-specific. These data are consistant with a model where viral restriction can be refined through sequence evolution in XRN1.
We next tested the presumption that the XRN1 orthologs are functionally equivalent for cellular processes when expressed within S. cerevisiae. We first confirmed that XRN1 orthologs successfully complemented the severe growth defect of S. cerevisiae xrn1Δ, by measuring the doubling time of S. cerevisiae xrn1Δ with or without a complementing XRN1-containing plasmid (Fig 4A). The knockout of XRN1 also renders cells sensitive to the microtubule-destabilizing fungicide benomyl [49], and we observed that all XRN1 homologs convey equal resistance to benomyl on solid medium (Fig 4B). It has been previously reported that over-expression of XRN1 is toxic to S. cerevisiae, a phenotype that has been suggested to be due to a dominant negative interaction of Xrn1p with other essential cellular components, such as the decapping complex [49]. Growth upon medium containing 2% galactose was equivalently reduced for strains carrying GAL1 inducible XRN1 genes from each species, whereas the strain over-expressing GFP grew normally (Fig 4C, right). Finally, the Ty1 retrotransposon is another intracellular virus that replicates within Saccharomyces species and often co-exists with L-A within the same cell. Interestingly, Xrn1p is not a restriction factor for Ty retrotransposons, but rather promotes their replication [43,47,48,53–57]. We found no significant difference between the mean values for retrotransposition in the presence of Xrn1p from S. cerevisiae, S. mikatae, S. kudriavzevii, or S. bayanus (one-way ANOVA, F3, 8 = 0.36, p = 0.78), indicating that the evolutionary differences between divergent XRN1 genes do not affect the ability of Ty1 to replicate within S. cerevisiae (Fig 4D). Collectively, these data indicate the cellular functions of Xrn1p have remained unaffected during yeast speciation, while the interaction with L-A viruses has changed.
We next mapped the region responsible for the species-specific restriction by XRN1. To better understand the structural organization of Xrn1p from S. cerevisiae, we used Phyre [58] to generate a template-based homology model of the exonuclease using the solved structure of Kluyveromyces lactis Xrn1p (Fig 5A). A linker region within the N-terminal domain, the far C-terminal domain, and domain D2 were not included in the model as there is a lack of information regarding the structural organization of these regions. Importantly, modeled domains contained three of the residue positions that we identified as evolving under positive selection (S1 Table), and all of these (blue) fall in and near the D1 domain (orange) (Fig 5A). As expected because of the selection that has operated on them, these residue positions under positive selection are more variable in sequence between species than are surrounding residues (two are shown in Fig 5B). All residues under positive selection are surface exposed and are far from the highly conserved Xrn1p catalytic domain (96% identity, across the Saccharomyces genus) and catalytic pocket (red). The other sites of positive selection fall within the last 500 amino acids of Xrn1p, which is less conserved compared to the rest of the protein (83% identity, across the Saccharomyces genus) (Fig 5C).
To define the importance of the two regions that we identified as containing signatures of positive selection, we replaced portions of S. kudriavzevii XRN1 with the equivalent portions of S. cerevisiae XRN1, and assayed for a region of S. cerevisiae XRN1 that would convey the ability to cure the killer phenotype. We found that an XRN1 chimera encoding the last 775 amino acids from S. cerevisiae (Sc-775) was sufficient to cure 56% of clones analyzed, and this was very similar to S. cerevisiae XRN1 (57%) (Fig 5D). Conversely, when the last 777 amino acids from S. kudriavzevii (Sk-777) were used to replace the same region within S. cerevisiae XRN1, only 9% of clones were cured (S3 Fig). This focused our construction of further chimeras to the second half of the protein, which also contains all of the codons under positive selection and has less amino acid conservation between S. cerevisiae and S. kudriavzevii (82% protein identity, compared to the N-terminius of Xrn1p with 95% identity).
Initial analysis of the highly diverged C-terminal tail revealed that the last 461 amino acids of S. cerevisiae Xrn1p were unable to convey efficient L-A restriction to S. kudriavzevii Xrn1p (S3 Fig). For this reason, we focused further chimeric analysis on the region encompassing the D1, D2, and D3 domains [59]. We swapped into S. kudriavzevii Xrn1p the D2+D3, D1, or D1-D3 domains of S. cerevisiae Xrn1p, and saw increasing rescue of the ability to cure the L-A virus (Fig 5D). All chimeric XRN1 genes were functionally equivalent with respect to their cellular functions, as all were able to establish normal growth and benomyl resistance in S. cerevisiae xrn1Δ (S4 Fig). Species-specific restriction maps predominantly to D1, with contribution from the neighboring D2 and D3 domains. Together, our data suggest that the exonuclease activity of Xrn1p is important for virus restriction and is preserved across species, but that evolution has tailored a novel virus interaction domain (D1-D3) that targets the enzymatic activity of Xrn1p against L-A in a manner that changes over time.
It’s hard to imagine that Xrn1p proteins from different species are differentially recognizing viral RNA, since they are all equivalent in their host functionalities within S. cerevisiae. We considered the possibility that there might be host-virus interactions beyond Xrn1p and the viral RNA. It has been shown that Xrn1p targets uncapped viral RNA transcripts rather than affecting dsRNA propagation [52]. As totivirus transcription only occurs in the context of a fully-formed capsid [60] and capsids are assembled entirely from the L-A Gag protein [61], it would seem plausible that Xrn1p may interact directly with Gag to target virus-derived uncapped RNAs. We introduced epitope tags onto Xrn1p (HA-tag) and the major capsid protein of L-A, Gag (V5-tag), and expressed both tagged and untagged versions of each protein from plasmids introduced into S. cerevisiae xrn1Δ (Fig 6A). Bead-bound antibodies specific for either HA or V5 were used to immunoprecipitate Xrn1p or Gag, respectively. We found that Gag (V5-tagged) was able to immunoprecipitate Xrn1p (HA-tagged) from S. cerevisiae and S. kudriavzevii (Fig 6A, top panel). Reciprocally, Xrn1p-HA from both S. cerevisiae and S. kudriavzevii were able to immunoprecipitate Gag-V5 (Fig 6A, bottom panel). The interaction between Xrn1p and Gag appears not to be mediated by single-stranded RNAs, as their digestion by RNase A in the whole cell extract did not affect the co-immunoprecipitation of Gag by Xrn1p (S5 Fig). We next performed these experiments with a monoclonal antibody specific to L-A Gag, so that endogenous L-A Gag protein could be immunoprecipitated. This reaction co-immunoprecipitated both S. cerevisiae and S. kudriavzevii Xrn1p (Fig 6B). Qualitatively, the relative efficiencies of Gag interaction with both S. cerevisiae and S. kudriavzevii Xrn1p appear similar in all assays, which seems at odds with our model that suggests that evolutionary differences within Xrn1p are a direct determinant of totivirus interaction. There are several possible interpretations. First, Gag might be antagonizing Xrn1p rather than being the species-specific target of Xrn1. Second, there may be a third component in this interaction which makes manifest the species-specificity of Xrn1p. Finally, a trivial explanation could be that coimmunoprecipitations are not very quantitative, and maybe there is in fact a difference in interaction with Gag between the Xrn1p of different species. Nonetheless, these data demonstrate a previously undescribed interaction that goes beyond Xrn1p interaction with viral RNA and warrants careful in vitro study.
We next wished to test our findings against other related yeast viruses. Indeed, the S. cerevisiae totivirus L-A-lus has been shown to have limited susceptibility to XRN1 from a different strain of S. cerevisiae [28]. We also wanted to test viruses of other species, but the only fully characterized totiviruses within the Saccharomyces genus are from S. cerevisiae. To identify totiviruses of other species, we screened Saccharomyces species from the sensu stricto complex for the presence of high molecular weight viral RNAs, and discovered a ~4.6 kbp dsRNA molecule within S. kudriavzevii FM1183 isolated from Europe (Fig 7A) [38]. We cloned the 4.6 kbp dsRNA molecule using techniques described by Potgieter et al. [62] and sequenced the genome of the virus using Sanger sequencing. We named the virus SkV-L-A1 (S. kudriavzevii virus L-A isolate number 1; Genbank accession number: KX601068). The SkV-L-A1 genome was found to be 4580 bp in length, with two open reading frames encoding the structural protein Gag and the fusion protein Gag-Pol (via a -1 frameshift) (Fig 7B). Conserved features of totiviruses were identified and include a conserved catalytic histidine residue required for cap-snatching (H154), a -1 frameshift region, packaging signal, and replication signal (Figs 7B and S6). Phylogenetic analysis of the Gag and Pol nucleotide and protein sequences firmly places SkV-L-A1 within the clade of Saccharomyces totiviruses represented by L-A and L-A-lus [28,63], as opposed to the more distantly related Saccharomyces totivirus L-BC (Figs 7C and S6) [64].
To test the effect of XRN1 upon SkV-L-A1, plasmids expressing XRN1 orthologs were introduced via LiAc transformation into S. kudriavzevii infected with SkV-L-A1. These plasmids were able to express XRN1 from each species, although we find that the expression is variable, with S. mikatae Xrn1p expressing at a level higher than the others (Fig 7D). The expression of these proteins did not affect the overall growth rate or colony morphology of S. kudriavzevii (S7 Fig). Because of the lack of an observable killer phenotype in this strain (likely because a killer toxin-encoding satellite dsRNA is not present), heterospecific XRN1 were expressed within S. kudriavzevii and analyzed for their ability to spontaneously cure SkV-L-A1, as we did previously with L-A in S. cerevisiae (Fig 3). We did not observe any virus curing by any orthologs of Xrn1p, but believe that this could be because the high-copy plasmids that we used in this experiment in S. cerevisiae are unable to drive Xrn1p expression in S. kudriavzevii high enough to actually cure the virus. However, we have observed previously that Xrn1p can reduce the abundance of totivirus RNAs (Fig 2A), so we further analyzed the XRN1-transformed clones of S. kudriavzevii for changes in SkV-L-A1 RNA levels using reverse transcriptase quantitative PCR (RT-qPCR). Total RNA was extracted from clones of S. kudriavzevii and converted to cDNA using random hexamer priming. cDNA samples were amplified using primers designed to specifically target SkV-L-A1 GAG and the cellular gene TAF10. The empty vector control was used as the calibrator sample, and TAF10 expression was used as the normalizer to calculate the relative amount of SkV-L-A1 RNAs present within each XRN1 expressing S. kudriavzevii cell line using the comparative CT method [65]. We found that expression of XRN1 from S. kudriavzevii (n = 10) reduced the relative levels of SkV-L-A1 RNAs by 40% (Fig 7E), even though this Xrn1p was expressed at the lowest levels (Fig 7D). This is in contrast to XRN1 from S. mikatae (n = 9) and S. bayanus (n = 8) that only showed a 13% increase or 15% decrease in SkV-L-A1 RNAs, respectively. S. cerevisiae XRN1 was able to reduce SkV-L-A1 RNAs by 27% and is noteworthy due to the close evolutionary relationship between SkV-L-A1 and other L-A-like viruses from S. cerevisiae (Figs 7C and S6). These data suggest that Xrn1p is a species-specific restriction factor in different Saccharomyces yeasts, and that coevolution of totiviruses and yeasts has specifically tailored the potency of Xrn1p to control the replication of resident viruses within the same species.
In the Saccharomyces genus, Xrn1p, the SKI complex, and exosome are all important for controlling the abundance of totivirus RNAs. We find that XRN1 and the exosome component RRP40 are somewhat unique in their strong signatures of positive natural selection. We speculated that positive selection might be driven by selection imposed by totiviruses. As speciation occurs and viruses mutate in unique ways in each lineage, new allelic versions of these antiviral genes that enable better control of totivirus replication would experience positive natural selection. Indeed, we found this to be the case, with S. cerevisiae Xrn1p restricting the S. cerevisiae L-A virus better than any other ortholog of XRN1, and S. kudriavzevii Xrn1p restricting S. kudriavzevii SkV-L-A1 virus the best. The exact nature of the host-virus protein-protein interaction that is driving this evolutionary arms race is not clear. To thwart XRN1, the totiviruses are known to synthesize uncapped RNAs with an exposed 5’ diphosphate, which is a suboptimal substrate for Xrn1p-mediated decay [24]. Further, it has been shown that the totivirus Gag protein has a cap-snatching activity that cleaves off caps from host mRNAs and uses them to cap viral transcripts, protecting them from Xrn1p degradation [12,14]. We have found that Xrn1p interacts with L-A Gag, and that this interaction is not mediated by the presence of single-stranded RNAs. What remains unknown is whether Xrn1p is targeting Gag as part of the restriction mechanism, or whether Gag is targeting Xrn1p as a counter defense. As we did not observe an obvious species-specific differences in the interaction between Xrn1p and L-A Gag by coimmunoprecipitation, we cannot clearly define the observed role of sequence variation in Xrn1p. This may be because of the low sensitivity of our assay system, or because direct binding of Xrn1p by L-A Gag is ubiquitous and that the rapid evolution of XRN1 results from another intriguing facet of virus-host interaction and antagonism. However, we now know that the interaction between L-A and Xrn1p goes beyond the simple recognition of L-A RNA by Xrn1p. We can speculate that Xrn1p may compete with Gag for access to uncapped viral RNAs as they are extruded into the cytoplasm, or that interaction with unassembled Gag allows the recruitment of Xrn1p to sites of virion assembly resulting in viral RNA degradation. Alternately, it is possible that the target of Xrn1p is simply L-A RNA, and that the interaction with Gag reflects a viral countermeasure where Gag is redirecting or otherwise altering the availability of Xrn1p to degrade L-A RNA. Indeed, there are several examples of mammalian viruses that redirect or degrade Xrn1p to aid in their replication [17,18,66].
The literature suggests that Xrn1p is a widely-utilized restriction factor against viruses, as it has been reported to have activity against mammalian viruses [9,16], yeast viruses [8,24], and plant viruses [46]. The potent 5’-3’ exonuclease activity of Xrn1p has resulted in viruses developing a rich diversity of strategies to protect their RNAs. For instance, Hepatitis C virus recruits MiR-122 and Ago2 to its 5’ UTR to protect its RNA genome from Xrn1p degradation [7,16]. The yeast single-stranded RNA narnavirus uses a different strategy to protect its 5’ terminus, folding its RNA to form a stem-loop structure that prevents Xrn1p degradation [8]. In some cases, viruses even depend on Xrn1p to digest viral RNA in a way that benefits viral replication, for example, preventing the activation of innate immune sensors [41]. Flavivirus (West Nile and Dengue virus) genomes also encode RNA pseudoknot and stem-loop structures that arrest the processive exonuclease activity of Xrn1p, producing short subgenomic flavivirus RNAs (sfRNAs) that are important for viral pathogenicity [13,67]. Members of the Flaviviridae, Herpesviridae, Coronaviridae, and yeast Totiviridae have all been shown to encode proteins that initiate endonucleolytic cleavage of host mRNAs, revealing exposed 5’ monophosphates that are substrates for Xrn1p degradation. This is thought to interfere with host translation and to produce uncapped RNA “decoys” that potentially redirect Xrn1p-mediated degradation away from viral RNA [14,15]. Xrn1p degradation, Xrn1p relocalization, virus-encoded capping enzymes, cap-snatching mechanisms, RNA-protein conjugation, recruitment of host micro-RNAs, cleavage of host mRNAs as “decoys”, and viral RNA pseudoknots are all utilized to prevent Xrn1p-mediated viral RNA destruction [7,8,12–18]. All of this evidence suggests that viruses can employ various methods to escape or harness the destructive effects of Xrn1. Our data now suggests that Xrn1p in yeast is not a passive player in the battle against viruses, but rather that hosts can be selected to encode new forms of Xrn1p that can overcome virally encoded defense strategies.
To rationalize the model of an antagonistic relationship between L-A and Saccharomyces species, it is important to consider the fitness burden of strictly intracellular viruses. Prevailing wisdom assumes that infection of fungi by viruses is largely asymptomatic and benign, especially when considering that their intracellular lifecycle ensures an evolutionary dead-end if they kill or make their host unfit. Indeed, within laboratory yeast strains, the association between L-A and S. cerevisiae appears to be at equilibrium, with no major biological differences between strains infected or not infected by L-A [68]. Therefore, the relationship between L-A and the Saccharomyces yeasts could be viewed as mutualistic or even commensalistic [68–70]. Mutualism is particularly striking in the context of the L-A / killer virus duo that provides the host cell with the “killer” phenotype, a characteristic that is broadly distributed throughout fungi [71]. If an infected yeast cell can kill other yeasts around it using the killer toxin, it no longer has to compete for resources within that environmental niche, an evolutionarily advantageous situation [51,69,70]. Indeed, there are other examples of host-virus mutualism in fungi [72,73]. However, there are many observations that lead one to believe that the relationship between intracellular viruses and their hosts is not benign and static. Firstly, there is a measurable fitness cost to killer toxin production by S. cerevisiae within unfavorable environmental conditions that inactivate the toxin, allow for regular cellular dispersal and/or are nutrient rich [69,70]. Secondly, virus infection of pathogenic fungi can also cause hypovirulence (a reduction in fungal pathogenesis), an outcome that is being exploited to treat agricultural disease [74–77]. Thirdly, many wild and domesticated strains of S. cerevisiae are free of totiviruses (and therefore also of killer), suggesting that there is selection against the ongoing maintenance of these viruses [20,28,71]. Fourthly, the continued maintenance of RNAi systems in fungi also correlates with the loss of the killer phenotype and is known to antagonize fungal viruses [71,78]. However, a virus of the fungi Cryphonectria parasitica has been shown to antagonize and escape restriction by RNAi without crippling its host [78]. This antagonistic relationship appears similar to the equilibrium of Saccharomyces yeasts and totiviruses, and suggests that in the absence of effective RNAi, additional antiviral defenses may be biologically relevant (i.e. Xrn1p). In line with this view of a dynamic relationship between hosts and intracellular viruses, we show that totiviruses from different Saccharomyces species are best controlled by the Xrn1p of their cognate species, and that disruption of this equilibrium can result in excessive virus replication (Fig 2), virus loss (Fig 3), or a reduction in viral RNA (Figs 2 and 7). Signatures of positive selection that we have detected in Saccharomyces XRN1 are also consistent with a host-virus equilibrium that is in constant flux due to the dynamics of a back-and-forth evolutionary conflict (Figs 2 and 6).
There are several examples of mammalian housekeeping proteins engaged in evolutionary arms races with viruses. (By “housekeeping” we refer to proteins making critical contributions to host cellular processes, as opposed to proteins dedicated to immunity.) In most of these other examples though, the housekeeping protein is hijacked by viruses to assist their replication in the cell (rather than serving to block viral replication). For instance, many viruses hijack cell surface receptors to enter cells. We and others have shown that entry receptors are quite evolutionarily plastic, and that mutations can reduce virus entry without compromising host-beneficial functions of the receptor [34,79–83]. For example, the antagonistic interaction of Ebola virus (and/or related filoviruses) with the bat cell surface receptor, Niemann-Pick disease, type C1 (NPC1), has driven the rapid evolution of the receptor without affecting the transport of cholesterol, critical to the health of the host [34]. Numerous such examples highlight how essential housekeeping machineries, not just the immune system, are critical for protecting the cell from replicating viruses. This study highlights an interesting evolutionary conundrum that does not apply to classical immunity genes: as Xrn1p appears to be an antiviral protein, it must be able to evolve new antiviral specificities without compromising cellular health and homeostasis.
XRN1 from S. cerevisiae, including 1000 bp of the 5’ and 3’ UTRs, was amplified by PCR from genomic DNA prepared from S. cerevisiae S288C. This PCR product was cloned into the plasmid pAG425-GAL-ccdB by the “yeast plasmid construction by homologous recombination” method (recombineering) [84] to produce pPAR219. Briefly, pAG425-GAL-ccdB was amplified by PCR to produce a 5000 bp product lacking the GAL-1 gene and the ccdB cassette. The PCR primers used to amplify pAG425-GAL-ccdB contained additional DNA sequence with homology to the UTRs of XRN1 from S. cerevisiae. Both PCR products were used to transform BY4741, with correctly assembled plasmids selected for by growth on complete medium (CM)–leucine. The XRN1 open reading frame (HA-tagged and untagged) from S. mikatae, S. bayanus, or S. kudriavzevii was introduced into pPAR219 between the 5’ and 3’ UTRs from S. cerevisiae XRN1 using recombineering to produce pPAR225, pPAR226, and pPAR227, respectively. As a negative control, NUP133 was cloned into the pPAR219 plasmid backbone to produce pPAR221, which was used to allow growth of xrn1Δ on medium lacking leucine without XRN1 complementation. The LEU2 gene was replaced by TRP1 using recombineering techniques to produce the plasmids pPAR326, pPAR327, pPAR328, and pPAR329. Using PCR and recombineering, we also constructed chimeric XRN1 genes by exchanging regions of S. kudriavzevii XRN1 (pPAR227) with the corresponding regions of S. cerevisiae XRN1 (pPAR219). XRN1 inducible plasmids were constructed by cloning PCR-derived XRN1 genes into pCR8 by TOPO-TA cloning (Thermo Fisher). Utilizing Gateway technology (Thermo Fisher), XRN1 genes were sub-cloned into the destination vector pAG426-GAL-ccdB for over-expression studies [85]. The same pCR8/Gateway workflow was also used to clone and tag GAG from a cDNA copy of the L-A totivirus (pI2L2) to produce pPAR330 and pPAR331. The DNA sequences from all constructed plasmids can be found in S2 File. A list of all relevant plasmids can be found in S2 Table.
The S. cerevisiae killer strain (BJH001) was created by the formation of a heterokaryon from the mating of the haploid strains BY4733 (KAR1) and 1368 (kar1) [86]. The resultant daughter heteroplasmon cells were selected by growth on CM—uracil and the ability to produce zones of growth inhibition indicative of the presence of L-A and the killer virus. The inability to grow on CM lacking histidine, leucine, tryptophan or methionine was also used to confirm the genotype of BJH001. BJH006 was created by replacing XRN1 with the KANMX4 gene using homologous recombination within BJH001 [87]. A list of relevant yeast strains and species used in this study can be found in S3 Table.
1 x 109 yeast cells (~10 mL) were harvested from a 24–48 h overnight culture grown to saturation. Strains of S. kudriavzevii and S. mikatae were grown at room temperature, all other strains were grown at 30°C. The flocculent nature of some strains of wild yeasts made it challenging to accurately determine the exact number of cells present in some cultures. In these cases, the size of the cell pellet was used as an approximate measure of cell number relative to S. cerevisiae. Harvested cells were washed with ddH2O, pelleted, and washed with 1 ml of 50 mM EDTA (pH 7.5). Cells were again harvested and the pellets suspended by vortexing in 1 ml of 50 mM TRIS-H2SO4 (pH9.3), 1% β-mercaptoethanol (added fresh), and incubated at room temperature for 15 min. The cell suspension was centrifuged and the supernatant removed and the cell pellet suspended in 1 ml of BiooPure-MP (a single-phase RNA extraction reagent containing guanidinium thiocyanate and phenol) (Bioo Scientific) and vortexed vigorously. 200 μl of chloroform was added and vortexed vigorously before incubation for 5 min at room temperature. The aqueous phase and solvent phase were separated by centrifugation at 16,000 x g for 15 min at 4°C. The aqueous phase was transferred to a new tube and 1/3 volume of 95–100% ethanol added and mixed well by vortexing. The entire sample was loaded onto a silica filter spin column (Qiagen plasmid miniprep kit) and centrifuged for 30 s at 16,000 x g. The flow-through was discarded and the column washed twice with 750 μl of 100 mM Nacl/75% ethanol by centrifugation at 16,000 x g for 30 sec. The column was dried by centrifugation at 16,000 x g for an additional 30 sec. The dsRNA was eluted from the column by the addition of 100 μl of 0.15 mM EDTA (pH 7.0) and incubation at 65°C for 5 min before centrifugation at 16,000 x g for 30 sec.
dsRNA that was extracted from 1 x 109 yeast cells using our rapid extraction of viral dsRNA protocol was used as template for superscript two-step RT-PCR (Thermo Fisher). cDNA was created using a primer specific for the negative strand L-A genomic RNA– 5’ CTCGTCAGCGTCTTGAACAGTAAGC. Primers 5’-GACGTCCCGTACCTAGATGTTAGGC and 5’-CTCGTCAGCGTCTTGAACAGTAAGC were used to specifically target and amplify cDNA derived from negative strand L-A virus RNAs using PCR with Taq (New England Biolabs). The plasmid pI2L2 was used as a positive control for the RT-PCR reaction as it contains a cDNA copy of the L-A virus genome [88]. Alternatively, we collected total RNA from ~1 x 107 actively growing yeast cells using the RNeasy total RNA extraction kit (Qiagen) and synthesized cDNA using primers to target both the positive and negative strand of either L-A (5’-AAGATATTCGGAGTTGGTGATGACG and 5’-TCTCCGAAATTTTTCCAGACTTTATAAGC) or killer virus (5’-GCGATGCAGGTGTAGTAATCTTTGG and 5’-AGTAGAAATGTCACGACGAGCAACG). The same primers were used to detect L-A and killer virus specific cDNAs using PCR with Taq polymerase (New England Biolabs).
We assayed Ty1 retrotransposition in S. cerevisiae xrn1Δ, using the previously described Ty1 retrotransposition reporter system [89], and confirmed that XRN1 deletion causes a dramatic reduction in Ty1 retrotransposition (~50-fold) [43]. To test the effect of XRN1 evolution on Ty1 replication, we introduced XRN1 from S. cerevisiae, S. mikatae, S. kudriavzevii, or S. bayanus into xrn1Δ and assayed Ty1 retrotransposition.
Yeast lysates were prepared using the Y-PER reagent (Thermo Fisher) from 100 μl volume of log-phase yeast cells as per manufacturer’s instructions or by bead beating as described previously [90]. HA-tagged Xrn1p was detected via Western blot using a 1:5000 dilution of a horseradish peroxidase conjugated anti-HA monoclonal antibody (3F10—# 12013819001) (Roche). Adh1p was detected using a 1:10000 dilution of rabbit polyclonal anti-alcohol dehydrogenase antibody Ab34680 (Abcam). V5-tagged proteins were detected using a 1:5000 dilution of a mouse monoclonal antibody (R960-25) (Life Technologies). Native L-A Gag was detected using a 1:1000 dilution of a mouse monoclonal antibody (gift from Nahum Sonenberg). Secondary antibodies were detected using ECL Prime Western Blotting Detection Reagent on a GE system ImageQuant LAS 4000 (GE Healthcare Life Sciences).
Nucleotide sequences from six species of Saccharomyces yeasts were obtained from various online resources, where available [36,38,91]. Maximum likelihood analysis of dN/dS was performed with codeml in the PAML 4.1 software package [39]. Multiple protein sequence alignments were created using tools available from the EMBL (EMBOSS Transeq and Clustal Omega) (www.embl.de). Protein alignments were manually curated to remove ambiguities before processing with PAL2NAL to produce accurate DNA alignments [92]. DNA alignments were fit to the NSsites models M7 (neutral model of evolution, codon values of dN/dS fit to a beta distribution, with dN/dS > 1 not allowed) and M8 (positive selection model of evolution, a similar model to M7 but with an additional site class of dN/dS > 1 included in the model). To ensure robustness of the analysis, two models of codon frequencies (F61 and F3x4) and multiple seed values for dN/dS (ω) were used (S1 Table). Likelihood ratio tests were performed to evaluate which model of evolution the data fit significantly better. Posterior probabilities of codons under positive selection within the site class of dN/dS > 1 (M8 model of positive selection) were then deduced using the Bayes Empirical Bayes (BEB) algorithm. REL and FEL analysis was carried out using the online version of the Hyphy package (www.datamonkey.org) S1 Table [40]. Analysis of XRN1 was performed using the TrN93 nucleotide substitution model and the following phylogenetic relationship (Newick format):
((((((S. paradoxus-Europe, S. paradoxus-Far East), (S. paradoxus-North America, S. paradoxus-Hawaii)), S. cerevisiae), S. mikatae), S. kudriavzevii), S. arboricolus, S. bayanus);
GARD analysis found no significant evidence of homologous recombination within any dataset. MEGA6 was used to infer the evolutionary history of totiviruses using the Maximum Likelihood method. Appropriate substitution models were selected using manually curated DNA and protein alignments. The tree topologies with the highest log likelihood were calculated, with all positions within the alignment files containing gaps and missing data ignored. The reliability of the generated tree topologies was assessed using the bootstrap test of phylogeny using 100 iterations. Bootstrap values >50% are shown above their corresponding branches.
YPD plates containing 15 μg ml-1 of benomyl were prepared as described previously [49]. Yeast strains expressing XRN1 or containing an empty vector were grown overnight at 30°C in CM—leucine. Cell numbers were normalized and subject to a 10-fold serial dilution before spotting onto YPD agar plates with or without benomyl, and grown at 37°C for 72 h.
S. cerevisiae carrying multi-copy plasmids encoding XRN1 or GFP under the control of the GAL1 promoter were grown overnight at 30°C in CM—uracil with raffinose as a carbon source. Cell numbers were normalized and subject to a 10-fold serial dilution before spotting onto CM—uracil agar plates containing either 2% raffinose or galactose. Plates were grown at 30°C for 72 h.
Plasmids encoding various XRN1 genes were used to transform BJH006. Purified single colonies of killer yeasts were inoculated in 2 ml CM—leucine cultures and grown to mid-log phase. YPD “killer assay” agar plate supplemented with methylene blue (final concentration 0.003% w/v) and pH balanced to 4.2 with sodium citrate, were freshly inoculated and spread with S. cerevisiae K12 and allowed to dry. Thereafter, 1.5 μl of water containing 6 x 105 cells was spotted onto the seeded YPD plates and incubated at room temperature for 72 h. The diameter of the zones of growth inhibition were measured and used to calculate the total area of growth inhibition.
The curing of the killer phenotype was measured by transforming S. cerevisiae BJH006 with approximately 100 ng of plasmid encoding various XRN1 genes using the LiAc method. The addition of 1000 ng or as little as 10 ng of plasmid had no affect on the percentage of colonies cured using this assay. After 48 h of growth, colonies were streaked out and grown for a further 48 h. Clonal isolates of killer yeasts were patched onto a YPD “killer assay” plate (see kill zone measurement protocol) that were previously inoculated with S. cerevisiae K12, and incubated at room temperature for 72 h. The presence or absence of a zone of inhibition was used to calculate the percentage of killer yeast clones cured of the killer phenotype.
PHYRE was used to create a template-based homology model of S. cerevisiae Xrn1p using the solved structure of K. lactis Xrn1p as a template [58,59]. The structure was determined with an overall confidence of 100% (36% of aligned residues have a perfect alignment confidence as determined by the PHYRE inspector tool), a total coverage of 81%, and an amino acid identity of 67% compared to K. lactis Xrn1p. PDB coordinates for the modeled structure can be found in S1 File. Structural diagrams were constructed using MacPyMOL v7.2.3.
Strains were grown in CM lacking the appropriate amino acids in order to retain the relevant plasmids. For co-immunoprecipitations involving L-A Gag-V5 and Xrn1p-HA, 50 mL cultures (CM—tryptophan—leucine, 2% raffinose) were used to inoculate 500 mL cultures (CM—tryptophan—leucine, 2% galactose) at OD600 ~0.1. Cells were harvested at OD600 0.7, after ~14 h of growth at 30°C with shaking. Cultures used for the immunoprecipitation of native Gag were grown in the same manner, but in CM—leucine medium containing 2% dextrose. Immunoprecipitation of yeast and viral proteins were performed as previously described [90] with the following modifications: 2–4 mg of protein was used per co-immunoprecipitation. Approximately 50 μg of protein was loaded for the whole-cell extract “input”, as determined by Bradford Assay (~2% of total input), and was compared to 10–20% of each co-immunoprecipitation. Sepharose beads were substituted for Dynabeads MyOne Streptavidin T1 or Dynabeads Protein G (Thermo Fisher Scientific). For immunoprecipitation of Xrn1p-HA, we used an anti-HA-Biotin, rat monoclonal antibody (3F10—#12158167001) (Roche), and for Gag-V5 a mouse monoclonal antibody (R960-25) (Life Technologies). RNase A was added to whole cell extracts at a concentration of (80 μg mL-1) and incubated with Dynabeads during immunoprecipitation for 2 hours at 4°C. RNAse is in excess in our co-IP experiments, because significant RNA degradation occurred at concentrations of RNase 8-fold lower than we used (S5 Fig). RNA from samples with and without the addition of RNase A was recovered from yeast whole cell extracts after co-immunoprecipitation using Trizol according to manufacturer’s guidelines (Thermo Fisher). The extent of RNA degradation was measured using a 2200 TapeStation Instrument and a RNA screentape, as per manufacturer’s instructions (Agilent). An RNA integrity number (RIN) was calculated for each sample based upon criteria that reflect the quality of the RNA sample, as described previously [93].
dsRNAs were isolated from S. kudriavzevii as described above and processed according to the protocol of Potgieter et al. [62], with the following modifications: Reverse transcription reactions were carried out using Superscript IV (Thermo Fisher), PCR amplification was performed by Phusion polymerase (Thermo Fisher), and cDNAs were cloned into pCR8 by TOPO-TA cloning (Thermo Fisher) before Sanger sequencing.
S. kudriavzevii was transformed with plasmids expressing XRN1 from various Saccharomyces species, and an empty vector control using the LiAc method. The transformation was carried out at room temperature and heat shocked at 30°C. S. kudriavzevii transformants were recovered on CM—tryptophan and grown at room temperature. Clones were derived from two independent transformation reactions and grown to mid-log phase at room temperature. Total RNA was extracted from these cultures by first treating the cultures with Zymolase 100T (final concentration 100 μg mL-1) for 2 hours at room temperature in buffer Y1 (1 M Sorbitol, 100 mM EDTA (pH 8.0), 14 mM β-mercaptoethanol). Yeast spheroplasts were treated with Trizol to extract total cellular RNA, followed by a digestion of residual DNA by Turbo DNase for 30 min at 37°C (Thermo Fisher). The RNeasy RNA cleanup protocol was used to remove DNase from the RNA samples (Qiagen), which were then stored at -80°C. RNA was converted to cDNA using Superscript III and random hexamer priming, as per manufacturers recommendations. cDNA samples were diluted 10-fold with distilled RNase-free water and used as templates for qPCR. Primers designed to recognize the RNAs corresponding to GAG of SkV-L-A1 (5’-TGCTTCTGATTCTTTTCCTGAATGG-3’ and 5’-GCCACTTACTCATCATCATCAAAACG-3’) and the cellular transcripts from TAF10 (5’-ATGCAAACAATAGTCAAGCCAGAGC-3’ and 5’-TCACTGTCAGAACAACTTTGCTTGC-3') were used to amplify cDNA using SYBR Green PCR Master Mix (Thermo Fisher) on a CFX96 Touch (Biorad). TAF10 was used as a cellular reference gene to calculate the amount of viral cDNA within a given sample using the comparative CT method [65].
|
10.1371/journal.pcbi.1005852 | A Bayesian method for detecting pairwise associations in compositional data | Compositional data consist of vectors of proportions normalized to a constant sum from a basis of unobserved counts. The sum constraint makes inference on correlations between unconstrained features challenging due to the information loss from normalization. However, such correlations are of long-standing interest in fields including ecology. We propose a novel Bayesian framework (BAnOCC: Bayesian Analysis of Compositional Covariance) to estimate a sparse precision matrix through a LASSO prior. The resulting posterior, generated by MCMC sampling, allows uncertainty quantification of any function of the precision matrix, including the correlation matrix. We also use a first-order Taylor expansion to approximate the transformation from the unobserved counts to the composition in order to investigate what characteristics of the unobserved counts can make the correlations more or less difficult to infer. On simulated datasets, we show that BAnOCC infers the true network as well as previous methods while offering the advantage of posterior inference. Larger and more realistic simulated datasets further showed that BAnOCC performs well as measured by type I and type II error rates. Finally, we apply BAnOCC to a microbial ecology dataset from the Human Microbiome Project, which in addition to reproducing established ecological results revealed unique, competition-based roles for Proteobacteria in multiple distinct habitats.
| Data from many fields are available primarily in the form of proportions, also referred to as compositions, which impose mathematical constraints on identifying interactions among components in the underlying systems. In particular, correlations cannot be calculated directly from proportions or from count data that give rise to them. Methods that work around this difficulty generally do so by imposing strong assumptions about the distribution of underlying data or associated correlations, and these in turn often prevent quantifying uncertainty in the resulting estimates of correlation. We developed a statistical model (BAnOCC: Bayesian Analysis of Compositional Covariance) that both estimates correlations between counts or proportions and provides a posterior distribution for each correlation that quantifies how uncertain the estimate is. BAnOCC does well at controlling the number of false positives in simulated data and can be practically applied to a wide range of proportional data types.
| A long-standing goal of applied statistics in many fields has been identifying features associated significantly by a measure such as correlation [1,2]. When the features to be associated form a composition, inference of the correlation matrix is subject to the well-known problem of spurious correlation [3–6]. Compositional data in particular are vectors of proportions that sum to a fixed constant (typically one); they are usually thought of as the result of sum-normalizing an unobserved (or unrecorded) and unconstrained basis, following the terminology of [6]. The resulting sum-constraint of the compositional data means that any pairwise correlation measured using such data can be non-zero even if all the pairwise correlations on the unobserved count scale are zero, a phenomenon called spurious correlation [3]. The fact that all the features sum to one also makes the correlation matrix on the unobserved counts (that is, the basis correlation matrix) non-identifiable without untestable, though perhaps not unreasonable, assumptions [7–10]. Any method thus offers at best a partial reconstruction of the unobserved count correlation matrix, and the interest in characterizing such correlations in fields from geology to ecology has led to a variety of approaches.
In the context of microbial ecology, several methods have been proposed to identify significant ecological relationships from compositions; virtually all rely on some form of sparsity assumption and infer quantities relating to the log-transformed unobserved counts (hereafter referred to as the log-basis). The only technique that does not rely on a sparsity assumption is ReBoot [7], which estimates a “compositionally-corrected” correlation matrix using a permutation-based method. Friedman and Alm [8] proposed SparCC, which estimates the log-basis correlation matrix under the assumption that the correlations are on average small in magnitude. Fang et al. [9] noted that the resulting estimate is not guaranteed to be positive definite or that the elements will lie inside [–1, 1] and proposed CCLasso to estimate the log-basis correlation matrix using a LASSO penalty on the off-diagonal elements of the variance-covariance matrix. Ban et al. [10] similarly proposed REBACCA to estimate the log-basis correlation matrix; they use the same LASSO penalty function but a different likelihood function. Kurtz et al. [11] proposed SPIEC-EASI to estimate the log-basis precision matrix when the number of features is large by using sparse graph estimation techniques.
These approaches have difficulty quantifying uncertainty in the estimates, cannot incorporate uncertainty from the choice of tuning parameter, and are not flexible in the quantities they estimate. Friedman and Alm [8] proposed an inferential procedure based on the bootstrap, but offered no theoretical justification. Fang et al. [9] and Kurtz et al. [11] focused solely on estimation, while Ban et al. [10] used a subsampling method from Shah and Samworth [12] to stabilize the selection error rate. The LASSO-based methods [9–11] typically choose a shrinkage parameter and subsequently infer the log-basis covariance or precision matrix. Friedman and Alm [8], Fang et al. [9], and Ban et al. [10] all use the log-basis covariance matrix for network construction, while Kurtz et al. [11] use the log-basis precision matrix. This means that investigators typically must choose whether a precision or correlation matrix is best, and often use the resulting estimate with little guidance as to its uncertainty.
We address these issues by providing a flexible, fully Bayesian approach to identify correlations in compositional data. It is able to quantify uncertainty through the associated posterior and estimates both the log-basis correlation and precision matrix by modeling the composition directly. The graphical LASSO prior of [13] is used to estimate a sparse log-basis precision matrix (and hence a sparse log-basis correlation matrix) through a LASSO penalty, mitigating the non-identifiable nature of the unobserved count correlation matrix. We have implemented the resulting method as BAnOCC (Bayesian Analysis of Compositional Covariance). In this study, we also use a first-order Taylor expansion to approximate the compositional covariance as a function of the mean and variance of the unobserved counts. While not necessary to the development of our method, this expansion helps us explore the situations in which a naïve approach (ignoring the sum-constraint) might work. This approximation shows not only that the spurious correlation between two features can take any value in [−1,1] even if none of the features are correlated on the unobserved count scale, but also that both the variances and means of the unobserved counts control the magnitude and direction of the spurious correlation. Thus, we provide a novel characterization of the surprisingly broad circumstances under which compositionality can impede straightforward identification of the correlation matrix, and we provide the BAnOCC model to overcome this in datasets where it is possible.
The model assumes that a single subject’s composition, Ci = (Ci,1,…,Ci,p)T, is generated by the normalization of that subject’s unobserved and unconstrained counts, Xi = (Xi,1,…,Xi,p)T. That is, Ci=Xi∑j=1pXi,j. We also assume that the unobserved counts for all subjects are independent and identically distributed (iid); this implies that the compositions are iid as well because the transformation is per-subject.
We also introduce notation for the covariance and correlation among the features. The covariance matrix of the unobserved counts is denoted by ΣX = [σX,jk], to be inferred from C1,…,Cn. Similarly, the covariance matrix of the composition is denoted by ΣC = [σC,jk]. To construct the network of feature interactions, the relevant null hypotheses (one for each feature pair j and k) are that features j and k have a covariance of zero (σX,jk = 0); this is equivalent to testing if they are uncorrelated (ρX,jk = 0). We then define the unobserved count and compositional correlation matrices as RX = [ρX,jk] and RC = [ρC,jk], respectively.
BAnOCC assumes that the unobserved counts follow a log-normal distribution and that their correlation matrix is sparse; it is parametrized with the log-basis precision matrix and the log-basis mean (Fig 1). Posteriors for the parameters of the model (and thus functions of them which are of interest) are inferred using MCMC sampling. This fully Bayesian treatment of the problem gives several advantages: a full posterior distribution to quantify the uncertainty in the estimates, the ability to place a prior on the sparsity parameter, and estimates of any function of the log-basis precision matrix, including the log-basis covariance and correlation matrices.
BAnOCC models the unobserved and unconstrained counts using a log-normal distribution with parameters based on the moments of the log-basis: Xi∼iidLN(m,S), such that m=E(log{X}) and S = Var(log{X}). This continuous approximation of the underlying unobserved count data is expected to perform well when the underlying counts have a large dynamic range. In ecology, for example, the log-normal distribution is used to model the (discrete) abundance across species [14,15]. In microbial ecology specifically, the logistic normal is sometimes assumed to be the generating distribution of the composition [10,11]; further, the (discrete) read counts are often simulated using a log-normal distribution [16,17]. The log-normal distribution also allows the totals to be easily integrated out of the likelihood.
The likelihood is parametrized by the log-basis precision matrix O = S−1 and the log-basis mean m, and other parameters of interest like the log-basis covariance matrix S are sampled as transformations of these. By parametrizing using O, we are able to leverage a graphical LASSO prior to enforce sparsity on O and by extension S. Conveniently, the assumption of the log-normal distribution obviates the need to sample the covariance of the unobserved counts to determine the existence and direction of an association between two features on the unobserved count scale. This results because when some element of S, sjk, is zero, then the corresponding element of ΣX,σX,jk∝esjk−1 will also be zero; further, the non-zero elements of S and ΣX will have the same sign (though not the same magnitude).
Under the log-normal assumption, the complete likelihood of the observed composition ci and the latent total ti=∑j=1pxi,j is given by
L(m,O|ci,ti)=exp[−12{log(citi)−m}TO{log(citi)−m}](ti)(2π)p/2|O|−1/2∏j=1pci,j,
(1)
where ci=(ci,1,…,ci,p−1,1−∑j=1p−1ci,j). A detailed derivation can be found in S1 Text. Fitting this likelihood directly is computationally expensive, as the presence of the latent totals necessitates exploring a space whose dimension depends on both n and p. However, (1) factors into two portions: a part dependent on the compositions ci, and the kernel of a log-normal distribution for the totals ti=∑j=1pxi,j with parameters mi*=1TO(m−log{ci})s2* and s2*=11TO1 (where 1 is a vector of 1’s). Integrating over the totals in (1) (S1 Text) gives the more computationally tractable marginal likelihood
L(m,O|ci)=| O|1/2exp{ −12(m− logci)TO(m−logci)−(m*i)2S2*}(2π)1/2(s2*)1/2(2π)p/2Πj=1pcij.
In order to mitigate the non-identifiability of the precision matrix O, BAnOCC uses a shrinkage prior to conservatively estimate the sparsest O consistent with the observed relative abundance data. This is the graphical LASSO prior of [13]:
p(O|λ)=C-1∏j=1pExpojj|λ2{∏k=i+1pLaplaceojk|λ}1O∈M+,
where 1O∈M+ is an indicator function that O is positive definite, Exp(x|λ) has the exponential density of the form p(x) = λe−λx1x>0, and Laplace (x|λ) has the Laplace density of the form p(x)=λ2e−λ|x|. In comparison to variable selection priors such as spike-and-slab [18], the graphical LASSO prior is more scalable to high dimensions at the cost of being unable to generate estimates that are exactly zero [19]. We deal with this by using the resulting posterior samples to conclude whether a correlation is likely to be zero or not. The choice of λ is key to the degree of shrinkage imposed by this prior. We placed a gamma prior on λ in lieu of specifying it a priori; this is possible because [13] showed that the normalizing constant C does not depend on λ. The prior for m is the conditionally-conjugate normal prior N(n,L) with mean n and covariance matrix L. Hyperparameter choice for the two priors (on m and λ) is discussed in more detail below.
BAnOCC samples the posterior using Stan’s C++ implementation and R interface [20]. Multiple quantities can be estimated from BAnOCC, including the log-basis precision, covariance, and correlation matrices. In our simulations and application, we estimated the log-basis correlation RlogX because it is interpretable and nicely scaled; we used the posterior median as the point estimate and the 95% credible intervals for wjk to determine whether the correlation between features j and k was non-zero.
The interpretation of the prior parameters on m is relatively straightforward, while that of the shrinkage parameter λ is less clear. Because log-basis means m have a normal distribution, em represents the median unobserved counts, which conveniently have a log-normal distribution with parameters n and L. Therefore, we could parametrize the prior on m by the expected median unobserved counts nLN = exp{n + 0.5diag(L)} and uncertainty of the median unobserved counts LLN=nLNnLNT(eL−1). The prior on the shrinkage parameter λ has a shape parameter a that determines how much prior probability mass is placed on λ values close to zero, and a rate parameter b that determines how the probability mass is spread across the entire domain. In particular, a ≤ 1 forces an asymptote at zero, while a > 1 does not.
When little or no prior data is available, weakly informative priors can be used. Any prior on λ should have high probability mass close to zero and so should have a ≤ 1. Larger values of a will “soften” the asymptotic behavior at zero (S1 Fig). The value of the rate parameter b should be chosen to so that most prior probability mass is on sensible values for λ. The degree of shrinkage implied by λ does not appreciably change for λ > 1 (S2 Fig), and so a b of around 5 will give a reasonable uninformative prior distribution for λ. For the log-basis means, m∼N(0,lI) can be used, with l a large value such as 100. An overlarge value for l can make computation less efficient and put prior mass on grossly implausible values of em, so an l of 500 or less is reasonable.
Prior subject-matter information can be incorporated into the priors for both λ and m, but most easily into the prior on m. If the data have few features, a smaller shape hyperparameter a should be employed to upweight values of λ that yield high shrinkage. The implied prior on the median unobserved counts em could be sampled to provide an empirical distribution of the total counts ∑j=1pemj; this could be assessed for gross deviations from what might be considered reasonable, or agreement with known ranges if such data are available.
The implementation of BAnOCC is publicly available with source code, documentation, and tutorial data as an R/Bioconductor package at http://huttenhower.sph.harvard.edu/banocc.
We first aimed to identify what characteristics of compositional data impede or facilitate the accurate estimation of the unobserved count correlation matrices in general. Such characteristics should delineate when BAnOCC or any other technique for estimating the unobserved count correlation would perform well. A first-order Taylor expansion approximates the compositional covariance as a function of the mean and covariance of the unobserved counts. Because the compositional correlation is a function of the compositional covariance, the resulting approximation also explains how the correlation behaves. Letting X represent the unobserved counts and C the composition, with the mean of X denoted by μX = (μX,j)T and the approximate average proportions by ω=(μX,1∑j=1pμX,j,…,μX,p∑j=1pμX,j)T, the Taylor expansion yields
ΣC≈(1∑j=1pμX,j)2(I−ω1T)ΣX(I−ω1T)T.
(2)
Here I is the p × p identity matrix, and 1 is a p-dimensional vector of 1’s. Eq (2) allows us to approximate the behavior of the compositional covariance from the parameters of the unobserved counts that generate it. For a detailed derivation, see S1 Text.
Surprisingly, when no features are correlated on the unobserved count scale, the spurious correlation can take any value in [−1,1] depending on the properties of the unobserved counts (Fig 2). This is suggested by considering Eq (2) when σX,jk = 0 for all j ≠ k, then
σC,jk≈(1∑l=1pμX,l)2[ωjωk∑lσX,ll−ωjσX,kk−ωkσX,jj]forj≠k.
(3)
The weights ωj and the variances σX,ll can be configured arbitrarily to force σC,jk either to the extreme positive or extreme negative end of the spectrum. In particular, we see three types of strong spurious correlations (Fig 2B–2D): “negative dominant”, “positive dominant”, and “negative mixed”. These three types of correlations are thus representative of a range of expected real-world behaviors, and we included them in subsequent simulation studies of BAnOCC and previous models.
“Negative dominant” spurious correlation (Fig 2B) occurs when features j and k in the unobserved counts have (1) high mean and (2) high variability compared to the remaining (l ≠ j,k) features. Intuitively, the remaining features must contribute minimally to the total mean or total variance in the unobserved counts. When normalized, the sum-constraint thus forces a negative correlation between features j and k because they behave as if they were the only two features in the composition.
In the “positive dominant” spurious correlation type (Fig 2C), features j and k in the unobserved counts have (1) small variability and (2) high mean relative to the remaining (l ≠ j,k) features. The positive correlation in the composition results because the variability in the sum of the remaining feature abundances causes the compositions for features j and k to be shrunk or stretched in the same direction when the data are normalized.
Finally, “negative mixed” spurious correlations are the result of “positive dominant” type bases where feature k and the remaining features have switched roles (Fig 2D). After normalization, the variability in feature k forces feature j to move in the opposite direction to accommodate the remaining features.
Eq (3) also offers an alternative explanation for the negative covariance between features in a Dirichlet distribution. A Dirichlet distribution with parameters α1…,αp results when each feature is independent on the unobserved count scale and has a Gamma(αj,β). The mean and variance of a Gamma distribution are αβ and αβ2, respectively, implying that in the unobserved counts, a feature with high mean will also have high variance, and vice versa. This captures “negative dominant” correlations well, but fails to capture “positive dominant” or “negative mixed” correlations, which result when at least one feature has high mean but low variance in the unobserved counts.
Eqs (2) and (3) further suggest that the overall effect of normalization on the correlation estimate as the number of features p increases depends on the characteristics of μX and ΣX. In ecological applications, it is often assumed that if p is large and the compositional means are similar across the p features, then the correlation estimates based on the composition and unobserved counts are not likely to be very different [8,10]. Part of the appeal of this reasoning is that it does not rely on information about the unobserved and unconstrained counts. Expanding Eq (2), we can see that ΣC ∝ ΣX − ω1TΣX − ΣX1ωT + ω1TΣX1ωT. If the means are very similar to each other, this affects only the weights ω given to the offset ω1TΣX − ΣX1ωT + ω1TΣX1ωT. Small weights render the offset negligible only in the case where the unobserved variance on the unobserved counts ΣX is not too large: the behavior of the offset as the number of features increases depends on the similarity of the means (through ω) and on the variances of the additional features in the unobserved counts (through ΣX).
Thus when analyzing compositional data, one cannot know with certainty in which data the correlations are strongly affected by the normalization, much less the magnitude and direction of the change in correlation structure induced by normalizing. The information loss due to normalization implies that ΣX is non-identifiable without assumptions about its structure. However, knowing how the unobserved and unobserved counts affect the spurious correlation allows simulation of datasets that have specific types of spurious correlation for testing the performance of estimation methods in these cases.
Using the information from this theoretical analysis, we tested BAnOCC on two types of datasets. The first comprised small datasets generated using the model itself but designed to be challenging by incorporating negative dominant correlations. Second, we also simulated larger, more realistic datasets using an independent model specific to microbial community structure, sparseDOSSA [21].
For the former, four small datasets with 1,000 samples and nine features each were generated according to four scenarios. The “simple” scenario had no true correlations and no negative dominant correlation; the “high spurious” scenario had no true correlations but the presence of a negative dominant correlation; the “retained spike” scenario had several true correlations and no negative dominant correlation; and the “reversed spike” scenario had several true correlations and a negative dominant correlation between two features that are positively correlated in the unobserved abundances (see details in S2 Text and data in S1 Data). On these data, we used hyperparameters nj = 0, L = 1000I, a = 0.5 and b = 5 (S3 Fig).
Realistic data were generated using the SparseDOSSA model [21], which generates each feature from a zero-inflated, truncated log-normal distribution with subsequent rounding and estimates the feature-specific parameters by fitting to a given real-world template dataset. We induced correlations between features by using a multivariate distribution with a log-basis correlation that had off-diagonal elements set to one of four different correlation strengths ({−0.7,−0.3,0.3,0.7}). To ensure that strong compositional effects were present, we used a template with low-diversity community structure [22] with 14 pseudomicrobial features. The correlations were set so that the non-zero elements of the log-basis precision matrix and the log-basis covariance matrix would be the same; we used seven correlations (see details in S2 Text and data in S2 Data). We used hyperparameters a = 0.5, b = 5, nj = 3, and L = 30I (S4 Fig).
Using our first set of simulated data for evaluation, we compared the estimation and inference from BAnOCC with that from CCLasso [9], a frequentist LASSO-based method that chooses the shrinkage parameter using K-fold cross validation (Fig 3). BAnOCC had much lower false positive rates than CCLasso, resulting from the model’s ability to use the posterior distribution to account for estimate uncertainty while CCLasso, being LASSO-based, used a non-zero point estimate to determine significance of an effect.
BAnOCC and CCLasso both estimate the log-basis correlation matrix accurately, and both are a substantial improvement on a naïve approach (row 2 of Fig 3). In particular, both BAnOCC and CCLasso have much lower false positive rates than Pearson correlation. Over all the null associations, Pearson correlation had a staggering false positive rate of 82%; CCLasso had almost 14% false positives as a result of many small but non-zero estimates; BAnOCC, because it uses the posterior credible intervals to evaluate uncertainty, had a false positive rate of about 3%. BAnOCC cannot estimate the log-basis correlations wjk to be exactly zero because of the continuous prior, but the null associations whose 95% credible intervals cover zero have very small estimates (all are less than 0.15, 75% are less than 0.05).
The association between features 1 and 5 in the “reversed spike” dataset was difficult for both BAnOCC and CCLasso. Both gave a small, negative estimate (-0.001 for BAnOCC and -0.113 for CCLasso). BAnOCC displays a slight bias toward positive correlations instead of the moderate negative correlation that was present in the underlying unobserved abundances, as shown by several false positive associations in this dataset. This behavior is common among many methods, including SparCC and SPIEC-EASI (S5 Fig). It results from the fact that when a negative-dominant structure is present, positive correlations become much more likely to be real than negative ones, an interesting observation to consider when interpreting real-world results from any of these methods.
BAnOCC and CCLasso agree well with the true magnitude and direction of the non-zero associations that both methods conclude are significant. For these associations, the relative difference with the true value is less than 15% for both methods. When the associations were rejected, the 95% credible interval from BAnOCC covered the true value, indicating its utility for evaluating the uncertainty of the estimate. The false negative rates were 25% for BAnOCC and 0% for CCLasso, a direct result of the higher tolerance for false positives CCLasso exhibits. In practice, this has the expected effect of dramatically lowering BAnOCC’s false positive rate in recovering true correlations from compositional data.
We compared BAnOCC’s performance as measured by type I and type II error rates to a range of previous methods (Fig 4): simplicial variation [23], SparCC [8], CCLasso [9], SPIEC-EASI [11], ReBoot [7], and Spearman correlation (directly on the composition as a negative control). Of the two frequentist LASSO-based methods (CCLasso and REBACCA [10]), CCLasso alone had an R package interface; because they employ highly similar approaches, they should yield similar results. For a positive control, we also applied Spearman correlation to the unconstrained (and usually unobserved) counts (Table 1 and S3 Text).
Overall, BAnOCC controlled the type I error rate for all correlation strengths (Fig 4A) while maintaining comparable power compared with other recent methods (Fig 4B). These results held true in a more even community with larger features, in which BAnOCC was the sole method to fully control the type I error rate (S8 Fig). As the number of samples increased, all methods increased in power (S9 Fig), while the type I error rates remained fairly constant (S10 Fig).
Only BAnOCC and SparCC controlled type I error while maintaining high power for all correlation strengths (see also AUC boxplots in S6 Fig). Both behaved similarly to Spearman correlation applied to the unconstrained abundances, which represents the best possible performance (as it uses the unconstrained data rather than the composition—this is impossible in practice, when only the composition is available). SparCC’s type I error rate was slightly inflated in a larger dataset with more features, while BAnOCC continued to control the type I error rate at the nominal level (S8 Fig). As other authors have noted, SparCC does not guarantee that its log-basis correlation estimate has bounded elements nor that it is positive definite [9]. By contrast, BAnOCC not only estimates a positive definite correlation matrix with bounded elements, but also can infer network edges based on the precision matrix as well.
Several methods proved to control the type I error rate poorly: Spearman correlation exemplifies this as a negative control, but simplicial variation, SPIEC-EASI using GLASSO and to a lesser extent CCLasso were comparable. ReBoot, by design, attenuates the type I error rate of Spearman correlation, but does not control it perfectly. The high type I error rates are also somewhat expected in simplicial variation, but SPIEC-EASI using GLASSO may not be performing as expected, especially since in contrast the Meinshausen-Bühlmann neighborhood selection method did control type I error. This may also possibly be because the neighborhood selection infers each element of the matrix one at a time, while GLASSO infers the matrix all at once; this makes the GLASSO optimization a more difficult problem.
Feature 5 in the template dataset has a large mean and variance, while feature 3 has a small mean and variance. This results in a strong negative spurious correlation in the composition, which gives rise to interesting behavior of essentially all methods when detecting this association. When the true association is negative, many compositionally-appropriate methods such as BAnOCC, SparCC, and SPIEC-EASI (MB) do poorly at detecting the true correlation (Fig 4B) because the negative correlation is difficult to attribute to the unobserved counts rather than spurious correlation. Conversely, more naïve methods such as simplicial variation and Spearman correlation do very well at detecting a weak negative correlation between these two features because this becomes a strong negative correlation in the composition. This simulated example thus provides some insight into the form of sensitivity / specificity tradeoff that applies in the constrained, information-loss setting of identifying true correlations from compositions.
As an example application, we inferred a correlation network among microbial taxa profiled using ecological data from the Human Microbiome Project [22] (Fig 5). Microbial community sequencing generates compositions by assigning sequencing reads to microorganisms; since nucleotide sequencing depth is arbitrary, the resulting counts are not informative regarding the unobserved and unconstrained counts and are often normalized to relative abundances. Co-variation patterns in such data are of interest because they suggest ecological interactions, such as mutualism (positive correlation) or predation (negative correlation) [7].
The microbial taxonomic relative abundance data used here consisted of 523 microbial features measured across 700 total samples using MetaPhlAn2 v2.0_beta1 [24] in July of 2014 (available in S3 Data), further excluding from all networks markers removed in the subsequent version’s database (v2.0_beta2). These samples were in turn drawn from 127 individuals at six distinct body sites. Microbial ecology differs at each body site [22], providing examples for BAnOCC analysis that ranged from diverse, relatively even communities (such as stool) to less diverse, highly skewed ecologies (such as the vaginal posterior fornix). For each of three representative body sites (stool, posterior fornix, and buccal mucosa), we selected the first time point from each subject, collapsed taxonomic information to the genus level, and then removed features with relative abundance less than 0.0001 in at least 50% of samples. With too few features, little to nothing can be concluded about the true correlations; so if fewer than 10 features remained we lowered the prevalence cutoff until 10 features were retained.
The hyperparameters for the gamma prior on λ were a = 0.5 and b = 5 for all body sites, ensuring that we gave substantial weight to sparser precision matrices. For all body sites, we used the prior variability of the log-basis means L = 30I; each body site, however, had a different nj so that the distribution of the sums of medians were similar across different body sites (see S11–S14 Figs). We further compared BAnOCC’s inferred network using the log-basis correlation matrix with that from CCLasso, and BAnOCC’s inferred network using the log-basis precision matrix with SPIEC-EASI. There is broad agreement between the methods as to which edges are significant, with very few edges discrepant between the methods (S15 Fig).
In stool, BAnOCC inferred several positive associations between genera within the family Bacteroidales, in particular Bacteroides, Odoribacter, Parabacteroides and Alistipes (Fig 5A). Until recently, these genera were classified as part of the same genus [25]. This supports the common observation that closely (but not too closely) related taxa tend to have positive ecological associations [26]. Additionally, positive associations in the buccal mucosa (Fig 5B) connect taxa that are known to physically co-aggregate; in particular, Fusobacterium interactions with species from the Porphyromonas and Capnocytophaga genera (among others) are crucial in biofilm formation [27] and have been previously recovered from 16S-based ecological analyses [7]. Lastly, we can see the well-documented negative association between the Lactobacillus genus in the posterior fornix with several genera associated with dysbiosis such as Gardnerella and Prevotella [28] (Fig 5C).
Two interactions newly suggested by this analysis involved the Proteobacteria across multiple body sites, and specifically in stool and the oral cavity (buccal mucosa). The genera Escherichia and Haemophilus represent the two major proteobacterial residents in these habitats, respectively, and both were involved in predominantly negative interactions with more typical, abundant members of these communities (e.g. Faecalibacterium and Eubacterium in the gut, Leptotrichia or Corynebacterium in the mouth). These clades are highly phylogenetically diverged and tend to carry larger, more generalize genomes and pan-genomes [29,30]; this suggests that they will overgrow in these habitats only in unusual situations, exemplified by E. coli’s abundance in the gut primarily during inflammation [31]. Further details may be provided by future analyses using BAnOCC or related methods on species or strain-level ecological profiling.
Here, we describe BAnOCC, a Bayesian method for inferring the log-basis correlation structure from compositional data. Assuming a log-normal distribution on the unobserved and unconstrained counts, the model estimates the log-basis correlations using a sparsity-inducing shrinkage prior on the log-basis precision matrix. It is part of a family of several recently proposed LASSO-based methods [9–11] which provide a more rigorous approach to correcting for compositional effects than earlier methods [7,8]. Unlike the other LASSO-based correlation-inference methods that summarize pairwise associations using a single point estimate, BAnOCC yields uncertainty estimates of the precision, covariance, and correlation parameters. Simulation results show that BAnOCC performs as well as or better than existing methods in controlling type I error while maintaining power for network edge detection from compositional data. Finally, we applied the method to assess microbial relationships in the human microbiome, confirming established interactions and suggesting novel ones for future validation.
Analysis using a Taylor series approximation provided one of the first characterizations of properties that make true correlations “difficult” to recover from compositions, or conversely “easy” to miss as false negatives. In particular, this depends not only on the more intuitive number and evenness of feature means, but also on the distribution of their variance. This allowed us to simulate designedly difficult test cases for BAnOCC and a variety of published methods, in contrast to previous simulation studies that relied primarily on relatively simple synthetic data [7–10]. In most studies, spurious correlation is noted to be commonly present and of varying magnitudes and directions [11]. However, the possible sensitivity of methods to the type of spurious correlation encountered has not been explored and is an important contribution to the characterization of existing and future methods.
We anticipate several computational and statistical refinements that may further improve BAnOCC’s performance. While BAnOCC uses 95% credible intervals for inference, these can be overly conservative [32]. Alternative thresholding methods may improve on this, such as the scaled neighborhood criterion [32] or the partial-correlation based approach of [33] and [13]. A discrete-continuous mixture prior such as the G -Wishart prior [34] or the covariance selection prior [35] on the log-basis correlation matrix would further allow the posterior probability that wjk = 0 to be nonzero, and this quantity could be used as a threshold.
For applications specifically on count data, such as microbial compositions, the data could be modeled more accurately by adding a hierarchical layer. This would generate measurement counts conditional on the unobserved and unconstrained counts, making the observed compositions a function of normalized measurement counts. The degree of zero-inflation observed in ecological data could also be modeled directly using a hurdle or mixture model, or a multinomial distribution for the measurement counts. This would provide a particularly targeted approach for microbial ecology, in which more detailed data (at the species or strain level [24]) could be further incorporated. We thus hope to refine both the accuracy of compositional correlation inference and the applications to microbial community data in future studies.
|
10.1371/journal.pbio.1000415 | BiP Binding to the ER-Stress Sensor Ire1 Tunes the Homeostatic Behavior of the Unfolded Protein Response | The unfolded protein response (UPR) is an intracellular signaling pathway that counteracts variable stresses that impair protein folding in the endoplasmic reticulum (ER). As such, the UPR is thought to be a homeostat that finely tunes ER protein folding capacity and ER abundance according to need. The mechanism by which the ER stress sensor Ire1 is activated by unfolded proteins and the role that the ER chaperone protein BiP plays in Ire1 regulation have remained unclear. Here we show that the UPR matches its output to the magnitude of the stress by regulating the duration of Ire1 signaling. BiP binding to Ire1 serves to desensitize Ire1 to low levels of stress and promotes its deactivation when favorable folding conditions are restored to the ER. We propose that, mechanistically, BiP achieves these functions by sequestering inactive Ire1 molecules, thereby providing a barrier to oligomerization and activation, and a stabilizing interaction that facilitates de-oligomerization and deactivation. Thus BiP binding to or release from Ire1 is not instrumental for switching the UPR on and off as previously posed. By contrast, BiP provides a buffer for inactive Ire1 molecules that ensures an appropriate response to restore protein folding homeostasis to the ER by modulating the sensitivity and dynamics of Ire1 activity.
| Secreted and membrane-spanning proteins constitute one of every three proteins produced by a eukaryotic cell. Many of these proteins initially fold and assemble in the endoplasmic reticulum (ER). A variety of physiological and environmental conditions can increase the demands on the ER, overwhelming the ER protein folding machinery. To restore homeostasis in response to ER stress, cells activate an intracellular signaling pathway called the unfolded protein response (UPR) that adjusts the folding capacity of the ER according to need. Its failure impairs cell viability and has been implicated in numerous disease states. In this study, we quantitatively interrogate the homeostatic capacity of the UPR. We arrive at a mechanistic model for how the ER stress sensor Ire1 cooperates with its binding partner BiP, a highly redundant ER chaperone, to fine-tune UPR activity. Moving between a predictive computational model and experiments, we show that BiP release from Ire1 is not the switch that activates Ire1; rather, BiP modulates Ire1 activation and deactivation dynamics. BiP binding to Ire1 and its dissociation in an ER stress-dependent manner buffers the system against mild stresses. Furthermore, BiP binding accelerates Ire1 deactivation when stress is removed. We conclude that BiP binding to Ire1 serves to fine-tune the dynamic behavior of the UPR by modulating its sensitivity and shutoff kinetics. This function of the interaction between Ire1 and BiP may be a general paradigm for other systems in which oligomer formation and disassembly must be finely regulated.
| The secreted and membrane-spanning proteins that eukaryotic cells use to sense and respond to their environments and to communicate with other cells are functional only when they attain their proper three-dimensional structures. Folding of these proteins takes place in the endoplasmic reticulum (ER), aided by molecular chaperones. Degradation pathways help to discard misfolded proteins. When cells experience environmental stresses, nutrient depletion, or certain differentiation cues, the ER folding and degradation machineries can become overwhelmed and the cell risks accumulating and secreting malfunctional and potentially harmful proteins [1]. Such conditions of ER stress activate the unfolded protein response (UPR) [2], resulting in an expanded ER [3],[4] and increased expression of genes encoding ER chaperones, ER associated degradation machinery, and other components of the secretory pathway [5]. As such, the UPR provides a feedback loop that helps cells maintain high fidelity in protein folding and assembly.
The UPR plays a fundamental role in maintaining cellular homeostasis and is therefore at the center of many normal physiological responses and pathologies. For example, when the severity of ER stress exceeds the capacity of the UPR to restore homeostasis, mammalian cells commit to apoptosis [2]. Furthermore, the UPR is activated in many cancer cells [6],[7],[8] as well as during familial protein-folding and neurodegenerative diseases [9],[10]. Deficiencies in UPR signaling can also lead to diabetes [11]. Thus, the UPR constitutes an important control module whose core signaling machinery, which is conserved from yeast to humans, proves critical for cell physiology.
Misfolded secretory proteins accumulate in the ER lumen. The UPR is initiated in that compartment when the transmembrane sensor molecule Ire1 self-associates and activates its cytoplasmic endoribonuclease domain [12],[13],[14],[15]. Activated Ire1 transmits the signal by removing a non-conventional intron from its mRNA substrates, HAC1 mRNA in yeast and XBP1 mRNA in metazoans, which upon subsequent ligation are translated to produce potent transcriptional activators of UPR target genes [16],[17],[18]. Since the Hac1 protein is short-lived (half-life of ∼2 min) [18],[19], Ire1 activity is the key determinant of the magnitude and duration of the UPR.
Despite early clues for Ire1's role as a central UPR regulator, the mechanism by which it senses unfolded proteins remains disputed. One model proposes that Ire1 activity is mainly regulated by the ER-resident chaperone BiP (Kar2 in yeast). In this model, BiP inhibits Ire1 activity by binding to it in the absence of stress. During stress, BiP is titrated away by unfolded proteins, leaving Ire1 free to oligomerize and activate. This model was suggested because immunoprecipitation experiments showed that Ire1 interacts with BiP in unstressed cells and dissociates from BiP under ER stress conditions [20],[21],[22]. Site directed mutagenesis of BiP yielded mutants that do not bind to Ire1 [23], but since they failed to support growth when expressed as the only copy of BiP, they are difficult to interpret mechanistically in view of the many pleiotropic functions of BiP. By contrast, mutants of Ire1 lacking the juxtamembrane segment of its lumenal domain that is responsible for BiP binding retained regulation: mutant Ire1 was inactive in the absence of ER stress and activated in its presence [15],[22],[24],[25], thus suggesting that BiP release and rebinding are not causal for switching Ire1 on and off.
An alternative model of Ire1 regulation postulates that unfolded proteins bind to the lumenal domain of Ire1, triggering Ire1 self-association and activation of its cytoplasmic effector domains. Support for such activation of Ire1 by direct binding to unfolded proteins stems from structural studies of the Ire1 lumenal domain that revealed a putative peptide binding groove [24]. Mutational probing experiments demonstrated that the residues pointing into the groove are required for signaling [24].
Recently a hybrid, two-step model for UPR regulation has been proposed in which both BiP and unfolded proteins regulate Ire1: initial dissociation of BiP from Ire1 drives its oligomerization, while subsequent binding to unfolded proteins leads to its activation [15]. This model posits that BiP regulates Ire1 oligomerization, yet oligomerization is not sufficient for Ire1 activation. However, in vitro experiments demonstrated that the oligomerization state of the cytoplasmic domains of Ire1 determines the rate of enzymatic activity [12].
Thus, while genetic and biochemical analyses of the UPR have been immensely successful in elucidating many aspects of the UPR's unusual signal transduction mechanism, a coherent model of Ire1 regulation and the involvement of BiP has remained elusive. In this work, we study the UPR as a coordinated homeostatic system by carrying out measurements of the time dynamics of the pathway across a wide range of ER stress levels. Using population-based assays of UPR activity complemented with dynamic dose-resolved flow cytometry and a predictive computational model, we dissect the role of BiP in modulating the sensitivity and duration of the UPR. Specifically, by comparing the wild type UPR to a strain bearing a mutant version of Ire1 that lacks the UPR-specific BiP interaction motif, we show that BiP prevents Ire1 from activating in response to low levels of stress and that it aids in Ire1 deactivation once the stress has been alleviated. Using a single cell Ire1 FRET assay, we provide evidence suggesting that BiP performs these functions by sequestering inactive Ire1 molecules. By buffering Ire1, BiP ensures that only appropriate levels of stress trigger the UPR and that the duration of UPR induction matches the magnitude of the stress. These data position BiP as a modulator of the dynamic properties of the UPR.
Most UPR studies to date have been carried out under saturating conditions, where induction of protein folding damage surpasses the homeostatic capacity of the UPR and hence remains unmitigated. To position the experimental system in a physiological regime where cells proliferate efficiently when the UPR functions adequately, we probed the response to depletion of the metabolite inositol [26]. In the absence of inositol in the growth media, Ire1 is required for cells to induce the expression of genes required for inositol synthesis as part of the UPR transcriptional program [27]. To monitor UPR induction dynamics following this stimulus, we depleted inositol in a yeast culture and assayed for Ire1 activity as reflected by the splicing of HAC1 mRNA observed on Northern blots (Figure 1A, see Methods). After a lag phase—presumably the time required to exhaust residual inositol stores—HAC1 mRNA splicing reached a maximal level by 120 min, and then declined during an adaptation phase to recover near basal levels by 240 min. Population growth slowed during the induction phase but was restored upon recovery (Figure S1A). Thus, the UPR indeed functions as a homeostat in response to inositol depletion: the lack of inositol triggers activation of the biosynthetic pathway via Ire1, which initially overshoots and then settles at a new basal level that meets the cells' needs to grow under the new conditions. In this example, our detection of HAC1 mRNA splicing was not sensitive enough to detect a difference between the starting condition and the new basal level. However, blotting for the UPR target INO1 mRNA, which encodes inositol 1-phosphate synthase required for de novo inositol synthesis, demonstrated that the readjusted level at the 240 min time point was elevated compared to the un-induced system (Figure 1A, right panel), as was the expression of a UPR reporter (Figure S1B).
To determine whether similar adaptation also occurs after Ire1 activation in response to other modes of UPR induction, we treated cells with DTT, a reducing agent that counteracts disulfide bond formation and thereby induces protein misfolding in the ER. Disulfide bonds are formed through a relay in which ER client proteins are initially oxidized by protein disulfide isomerase (PDI). PDI is in turn oxidized by the FAD-dependent oxidase Ero1, which is finally oxidized by molecular oxygen [28]. Both PDI and ERO1 are UPR target genes, but since Ero1 directly passes the electrons to molecular oxygen, its abundance limits oxidative capacity. Thus, we reasoned that for moderate amounts of DTT, UPR-mediated induction of ERO1 would compensate for the increased demand for oxidation, allowing Ire1 to deactivate.
To test this, we treated cells with a range of DTT concentrations. Cells treated with 5 mM DTT no longer proliferated, indicating the presence of a maximal ER stress beyond which cells can no longer compensate effectively even in the presence of a maximally active UPR (Figure 1B, black). By contrast, cells treated with 2.2 mM or 1.5 mM DTT continued to proliferate, albeit at rates decreased from control cells (Figure 1B, purple and green). To investigate whether these growth phenotypes correlated with the activation and deactivation of the UPR, we monitored Ire1 activation by measuring HAC1 mRNA splicing as above (Figure S2). Consistent with the observed growth arrest, Ire1 activation was maximal and sustained in 5 mM DTT (Figure 1C, black): HAC1 mRNA was spliced to its full extent 30 min after DTT addition and splicing was maintained at this high level for the duration of the experiment. By contrast, in cells treated with doses of 2.2 mM or 1.5 mM DTT, Ire1 deactivation occurred in 4 h and 2 h, respectively (Figure 1C, blue and green). Therefore, under non-saturating DTT conditions, cells show the same transient Ire1 activity that characterized the response to inositol depletion. Furthermore, the duration of that transient response increased along with the magnitude of the stress.
To ascertain that the Ire1 activation and deactivation phases are reflective of the regulation of UPR target genes, we measured the expression of a synthetic UPR-regulated GFP transcriptional reporter (TR) over time in cells treated with 1.5 or 2.2 mM DTT (Figure 1D, E, see Methods). In these cells, the TR was induced to dose-dependent plateaus after a lag of approximately 30 min. The lag is consistent with the time required for transcription, translation, and GFP chromophore maturation, while the plateaus reflect the accumulation of the long-lived GFP reporter protein (half-life >8 h). Induction of a natural UPR target promoter, ERO1, closely matched the response from the synthetic TR (Figure S3). Therefore, the expression of UPR target genes at any given time is reflected by the rate of GFP production, rather than its abundance. When plotted as a function of the rate of GFP production (dTR/dt; Figure 1E), the TR exhibited activation and deactivation phases at 1.5 and 2.2 mM DTT that mirrored the dynamics of upstream HAC1 mRNA splicing (compare Figure 1C and 1E).
Taken together, the data shown in Figure 1 indicate that under different inducing stimuli, the UPR undergoes induction and adaptation phases that are reflected in the transient splicing activity of its sensor Ire1. Ire1 activity, in turn, is faithfully transmitted to the system's transcriptional output.
To assess whether the activation and adaptation properties of Ire1 are dependent on BiP binding and dissociation, we expressed a mutant form of Ire1, Ire1bipless, lacking a 51 amino acid segment (Ire1Δ475–526,GKSG) that contains the BiP binding site (see Methods, Tables 1, 2). While similar to the Ire1ΔV mutant described in [22], Ire1bipless retains 10 amino acids defined in the crystal structure of the core lumenal domain [24] that were deleted in Ire1ΔV. As previously reported, wild type Ire1 associated with BiP in a co-immunoprecipitation assay in the absence of ER stress (Figure 2A, B) but the association diminished when cells were treated for 1 h with 5 mM DTT (Figure 2A, B). By contrast, no change in the association of Ire1bipless and BiP was observed between stressed and unstressed cells (Figure 2A, B). The residual binding of BiP to Ire1bipless is likely due to non-specific absorption of the notoriously sticky chaperone (Figure 2A, B). As the amount does not change between UPR-induced and uninduced cells, this residual interaction does not reflect a physiologically important regulatory interaction.
To determine whether the diminished association between Ire1bipless and BiP impacts Ire1 activation, we measured HAC1 mRNA splicing in wild type cells and cells expressing Ire1bipless grown in the presence and absence of 5mM DTT for 1 h (Figure 2C). In both wild type and Ire1bipless cells, no detectable HAC1 mRNA was spliced in the absence of stress, and splicing was identically induced in the two strains after treatment with DTT. These data refute any model that poses modulation of the BiP•Ire1 association as the exclusive regulator of Ire1 activity.
Next, we investigated the subcellular localization of Ire1bipless in the presence and absence of ER stress. In response to ER stress, wild type Ire1 oligomerizes in clusters in the ER membrane that appear as discrete foci in fluorescence microscopy images [14],[15]. Similar to wild type GFP-tagged Ire1, GFP-tagged Ire1bipless displayed cortical and perinuclear ER localization in the absence of stress and formed bright foci in cells treated for 1 h with 5 mM DTT (Figure 2D). Quantification revealed that Ire1bipless formed foci of equal magnitude to the wild type protein upon UPR induction. In unstressed cells, however, Ire1bipless displayed a 2-fold increase in the level of clustering compared to wild type Ire1 (Figure 2E), and the foci exhibited considerable cell-to-cell variability (Figure S4, see Discussion).
The increased clustering of Ire1bipless did not apparently lead to activation, since a Northern blot of total RNA from cells bearing Ire1bipless did not show detectable amounts of spliced HAC1 mRNA in the absence of stress (Figure 2C). We considered it possible that splicing occurred at a level below the detection limit of the Northern blot assay. This reasoning is supported by Northern blots for INO1 mRNA, which is a more sensitive indicator of UPR induction as demonstrated above (Figure 1A, right). Indeed, INO1 mRNA was significantly elevated in cells expressing Ire1bipless as compared to cells expressing wild type Ire1 under non-inducing conditions (Figure 2F). Furthermore, there is a notable increase in the basal signal from a UPR reporter in unstressed Ire1bipless cells (Figure S5). Thus, UPR signaling in Ire1bipless cells is leaky.
The propensity of Ire1bipless to form small clusters in the absence of stress prompted us to ask if cells bearing Ire1bipless would be more sensitive than wild type to low levels of stress. To test this notion, we expressed a GFP splicing reporter (SR), in which the first exon of the HAC1 open reading frame is replaced by GFP (Figure S6A). The HAC1 intron represses translation of the mRNA, so GFP is only produced once active Ire1 removes the intron. Using flow cytometry, the SR allowed us to precisely quantify Ire1 activity over time in wild type and Ire1bipless cells. The SR did not compete with endogenous HAC1 mRNA for Ire1 when wild type cells were treated with 5 mM DTT for 1 h (Figure S6B), and similar to the TR, the GFP encoded by the SR decayed with a half-life of >8 h.
When wild type cells expressing the SR were treated with increasing concentrations of DTT, the SR was induced to dose-dependent plateaus (Figure 3A), and the rate of GFP production displayed the peak and decline behavior characteristic of the splicing of endogenous HAC1 mRNA (dSR/dt; Figure S7A). Consistent with the data shown in Figure 1, cells expressing wild type Ire1 were insensitive to DTT at concentrations below 1.5 mM as apparent from the absence of SR induction. By contrast, hac1Δ cells were hypersensitive to DTT: they induced the SR to near maximal levels at all doses (Figure 3B), and the rate of GFP production remained high until the reporter saturated (Figure S7B). In the absence of HAC1, Ire1 activation fails to initiate a transcriptional response, and the stress is never alleviated.
Interestingly, Ire1bipless cells showed an intermediate SR phenotype. Ire1bipless cells were more sensitive to DTT than wild type cells, becoming activated at 0.66 mM DTT and saturated at 1.5 mM DTT (Figures 3C, S7C). These data are consistent with the notion that increased clustering in Ire1bipless cells in the absence of DTT is coupled with sensitization, which allows activation at low levels of stress.
To validate that our data are consistent with a model of Ire1 regulation that includes interactions with unfolded proteins and BiP and to provide hypotheses for how BiP could specifically contribute to Ire1 regulation, we built a computational model of the UPR with the following assumptions (see Text S1). Ire1 can exist in one of three states: (i) as a free inactive monomer, (ii) as an inactive complex bound to BiP, or (iii) as an active complex bound to an unfolded protein (Figure 4A). Further, free BiP can bind to unfolded proteins and either productively aid in their folding or nonproductively dissociate. Unfolded proteins are either reduced or oxidized depending on the redox potential of the ER and must be oxidized in order to fold. In the model, the redox potential is set by the ratio of DTT to Ero1. When bound to an unfolded protein, the active Ire1 complex initiates the production of the Hac1 transcription factor, which in turn increases the production of BiP and Ero1 to close the UPR feedback loop. To explicitly model the measured experimental output (GFP fluorescence), the active Ire1 complex was set to trigger the production of a simulated SR in addition to producing Hac1. We extracted available model parameters from the literature and fitted remaining parameters to a subset of the experimental data (Figure S8, see Supporting Information for details). Using this “wild type” model as a baseline for comparison, we generated a “hac1Δ” model in which no induced production of BiP or Ero1 exists and an “Ire1bipless” model in which the interaction between Ire1 and BiP is disabled (Figure 4A).
The functional form of the dissociation of the active Ire1/unfolded protein complex was a modeling choice. Significantly, a model in which this dissociation was assumed to be linear did not reproduce the difference between the wild type and Ire1bipless when the SR time courses were simulated (Figure S9). Instead, a nonlinear, cooperative dissociation function of the active Ire1-unfolded protein complex was required to recapitulate the data; i.e., the dissociation rate of the active Ire1-unfolded protein complex must decrease in proportion to the concentration of the active oligomeric complex raised to a power greater than one. Given that Ire1 signals by clustering into foci, this nonlinear dissociation function can be thought of as a consequence of having to disassemble a cooperative enzyme complex (Figure S10, see Discussion). When simulated with such nonlinear dissociation of the active Ire1 complex, the model robustly recapitulated the DTT titration time course results in wild type, hac1Δ, and Ire1bipless cells (Figure 3D–F). When the SR time course was simulated with the wild type Ire1 model, doses of DTT of 1.5 mM and below produced less than 10% activity, 2.2 mM DTT produced an approximately half-maximal response, 3.3 mM DTT produced a response of approximately 75% of the maximum, and 5 mM DTT produced a near saturating response (Figure 3D). By contrast, simulation of the hac1Δ model produced near saturating responses to all doses, recapitulating the hypersensitivity measured in vivo (Figure 3E). Furthermore, simulation of the Ire1bipless model yielded an intermediate phenotype in which 0.66 mM DTT produced 15% activity, and doses of 1.5 mM DTT and above saturated the response (Figure 3F). Importantly, this agreement between the model simulations and experimental data was an emergent property of the functional interactions in the system, which arose independently of the choice of parameter values (Figures S11, S12).
In addition to accounting for the increased sensitivity of Ire1bipless compared to the wild type in the DTT titration time course experiments, our computational model predicted that Ire1bipless should exhibit delayed shutoff dynamics compared to the wild type after DTT is removed (Figure 4B).
This prediction can be rationalized in intuitive terms. When DTT is removed, disulfide bonds can form and proteins can mature. Thus the concentration of the ligand for Ire1 activation starts to decrease, and individual Ire1 molecules dissociate from the active oligomer. When wild type Ire1 dissociates, it can either rejoin the signaling complex (through interaction with an unfolded protein), or it can bind to BiP. Therefore, Ire1 deactivation proceeds rapidly since the inactive free form can be sequestered away by binding to BiP. In contrast, Ire1bipless lacks the ability to interact with BiP. Thus, while DTT removal will still prompt the dissociation of Ire1 from the active oligomer as the concentration of unfolded proteins decreases, the inability of Ire1bipless to bind to BiP increases the probability that an inactive Ire1bipless monomer will be recaptured by an unfolded protein and reactivate. As a result, Ire1bipless deactivation would proceed more slowly than that of wild type Ire1.
To test this prediction experimentally, we performed a DTT washout experiment in which wild type and Ire1bipless cells were treated with 5 mM DTT for 1 h to fully activate Ire1 in both strains. Subsequently, DTT was removed by filtration, cells were washed and resuspended in fresh media, and samples were collected over time to assay for HAC1 mRNA splicing by Northern blot (Figure 4C). Additional samples of wild type cells were collected to assay for the association of Ire1 and BiP by immunoprecipitation (Figure S13). Confirming the model predictions, we found that while Ire1 deactivated after 60 min in the wild type, Ire1bipless retained activity for 120 min. As expected, Ire1 deactivation correlated with re-association with BiP (Figure S13). These results point to a role for BiP binding in promoting Ire1 deactivation once stress has been alleviated.
To pursue the mechanism through which Ire1 deactivation proceeds, we hypothesized that, since Ire1 signals through assemblies of high-order oligomers, BiP binding may sequester breakaway Ire1 monomers, therefore promoting de-oligomerization of active Ire1 complexes. If this were the case, Ire1bipless cells should exhibit slower disappearance of Ire1 oligomers than wild type cells upon removal of stress.
To directly test this hypothesis, we co-expressed GFP- and mCherry-tagged versions of Ire1 or Ire1bipless and employed a microscopy-based fluorescence resonance energy transfer (FRET) assay [29] to quantify Ire1 self-association (Figures 5A, S14, see Methods). In an otherwise wild type scenario, the FRET signal displayed a broad dynamic range, from 0.01 a.u. (s.e.m. = 0.02, n = 36) in untreated cells in which the Ire1 fluorescence displayed a diffuse ER localization to 0.73 a.u. (s.e.m. = 0.06, n = 41) in cells treated with 5 mM DTT for 4 h, in which Ire1 is maximally clustered into foci (Figure S6B). In Ire1bipless cells, the basal FRET signal in the absence of DTT was elevated to 0.17 a.u. (s.e.m. = 0.09, n = 53), but the maximum FRET signal in the presence of DTT (0.71 a.u., s.e.m. = 0.08, n = 32) was comparable to wild type. As expected, wild type cells displayed transient increases in FRET signal that returned to baseline levels over the course of the experiment after treatment with 2.2 or 1.5 mM DTT (Figure 5B, C). In contrast, Ire1bipless cells were sensitized and displayed transient increases in FRET signal only when treated with 0.66 mM or 0.99 mM DTT but showed persistent strong FRET signal when treated with 1.5 mM or 2.2 mM DTT. These data recapitulate the role of BiP in buffering the Ire1 to low levels of stress (Figure 3).
To assess the role of BiP in the de-oligomerization of Ire1, we performed a DTT washout experiment and measured Ire1 FRET over time in wild type and Ire1bipless cells (Figure 5D, E). After treatment of both strains with 5 mM DTT for 1 h, we washed the cells in fresh media lacking DTT and imaged the cells over time. Consistent with the deactivation kinetics of wild type and Ire1bipless cells as measured by Northern blot, wild type Ire1 de-oligomerization proceeded rapidly and the FRET signal returned to baseline after 60 min. By contrast, the Ire1bipless FRET signal remained higher than basal levels at 120 min. Taken together, these data indicate that BiP binding to Ire1 contributes to the efficient de-oligomerization of active Ire1 complexes.
In this work, we investigated the homeostatic properties of the UPR in response to a range of physiological stress levels. Using time-resolved measurements of the induction and adaptation kinetics of the wild type UPR and a mutant UPR in which the sensor molecule Ire1 is not modulated by the chaperone BiP, we established a model for dynamic UPR regulation. In this model, Ire1 is principally activated when unfolded proteins bind to it directly. In a dynamic equilibrium, binding to unfolded proteins pulls Ire1 into oligomeric clusters and away from the chaperone BiP. Oligomerization, which occurs as a direct consequence of unfolded protein binding to Ire1's lumenal domain, is necessary and sufficient for Ire1 activation, and as such is the central control point in the UPR. Rather than regulating the first step of Ire1 activation, BiP provides superimposed modulation of the UPR's dynamic properties. Specifically, BiP assumes a dual role in which it simultaneously acts as a buffer to reduce the system's sensitivity to low stress levels and as a timer to tune the response time to the magnitude of stress by assisting in Ire1 deactivation once homeostasis is restored to the ER. The model establishes the UPR as a dynamic system whose capacity is adjusted to efficiently counteract a large spectrum of stress magnitudes and suggests a long-sought role for BiP binding to Ire1.
When cells experience protein folding stress in the ER, the UPR is activated to increase the ER's folding capacity. For manageable stress magnitudes, the UPR is capable of restoring folding homeostasis. However, if the magnitude of the stress surpasses the capacity of the UPR, yeast cells sustain maximal Ire1 signaling and cease to proliferate (Figure 1B, C). Within the physiological regime of ER stress, the response of Ire1 to moderate DTT inputs (1.5 mM and 2.2 mM DTT, Figure 1C) displayed transient activation dynamics, followed by adaptation to near basal levels. Interestingly, the duration of Ire1 activity—not the maximal amplitude of its activity—correlated with the magnitude of the stress. Since the Hac1 transcription factor is short-lived, the length of the Ire1 activation pulse should determine the duration of UPR target gene activation by Hac1 [18],[19]. This in turn determines the volume of the ER and the concentration of ER chaperones, components of the degradation machinery, and other cytoprotective proteins that are produced to combat the stress. This mode of signal regulation in which the duration of the output matches the magnitude of the input is known in engineering as “pulse-width modulation.” It is widely employed to reduce noise in engineered control systems by transforming an analog signal (amplitude) into a digital all-or-none pulse of varying length [30].
Although in principle real-time information about the folding status of the cell could be conveyed exclusively through the interaction of unfolded proteins with Ire1 to determine the duration of UPR induction, we find that BiP plays an important role in modulating the length of the Ire1 activation pulse (Figures S6A,C, 5B,C). Perhaps this modulating role of BiP reflects the necessity for precise tuning of the Ire1 pulse beyond what can be achieved through Ire1 and unfolded proteins alone. Interestingly, it was recently shown that a mutant of mammalian Ire1α shares salient properties with Ire1bipless: it does not bind to BiP, retains ER stress inducibility, and displays increased basal activity [31]. Therefore, it seems likely that the role of BiP in buffering Ire1 oligomerization is conserved in mammalian cells. Moreover, as the transmembrane kinase PERK, which in metazoan cells functions in a parallel UPR signaling branch to Ire1, shares close sequence homology to Ire1's lumenal domain, lessons learned for Ire1 modulation by BiP are likely to also apply to PERK regulation.
Precise tuning, and subsequently the buffering role of BiP, becomes all the more important since the UPR is linked to crucial cell fate decisions such as commitment to apoptosis [32]. The decision to commit to apoptosis might depend directly on the time of exposure to stress or on a thresholding mechanism through which either the extent of cellular damage or UPR machinery are assessed. Both scenarios would translate into an enhanced commitment to apoptosis in the absence of BiP modulation of Ire1.
As detailed above, precision homeostasis in the UPR requires the pathway-specific interaction of Ire1 and BiP. Disruption of this interaction in vivo leads to increased sensitivity to low levels of stress (“leakiness”), coupled to slower deactivation of Ire1 once stress is removed (Figure 4C). By using FRET to measure Ire1 self-association, we found that BiP performs these functions by aiding Ire1 de-oligomerization (Figure 5C–E). In vitro, Ire1 functions as a cooperative enzyme with a Hill coefficient >8, and the active species are large oligomers [12]. This high cooperativity could translate in vivo to a switch-like response of Ire1 to small changes in the concentration of unfolded proteins. For example, it follows from basic principles of enzyme kinetics that if Ire1 signals in clusters of 16 molecules, a mere 35% increase in unfolded proteins would cause Ire1 to go from 10% to 90% active. In this light, BiP's role as a binding partner that desensitizes Ire1 can be viewed as a gatekeeper that prevents triggering of the Ire1 activation switch following small or transient fluctuations in the local concentration of unfolded proteins. By doing so, BiP works to ensure that Ire1 is only activated when the stress is sufficient to warrant a response, thus improving information quality in the signaling pathway [33].
It is formally possible that in addition to loss of its UPR-specific BiP interaction Ire1bipless retains its ER-stress dependent activation, yet displays altered activation dynamics due to non-native conformational interactions. However, since Ire1bipless oligomerizes and activates in a ligand-specific manner to the same extent as wild type Ire1, we contend that in the simplest scenario, Ire1bipless, like the previous “bipless” mutant Ire1ΔV [22],[25], is a structurally sound molecule that is activated by the same mechanism that activates wild type Ire1.
Though similar to Ire1bipless, Ire1ΔV was not shown to be hypersensitive to DTT or to deactivate after washout with delayed kinetics [22]. However, Ire1ΔV did display hypersensitivity to heat shock and delayed deactivation kinetics in response to ethanol [22]. While the discrepancies between Ire1bipless and Ire1ΔV may be due to differences in experimental resolution, the elevated response of Ire1ΔV to heat shock and ethanol is consistent with the notion that BiP buffers Ire1 to these mild ER stresses.
Our study of the intricate UPR dynamics was guided by a computational model which was able to recapitulate our data and generate useful predictions. In the model, BiP serves as a buffer to the pool of inactive Ire1. By binding to free Ire1, BiP sequesters the inactive form of Ire1 and both prevents activation at low levels of stress and promotes deactivation once the stress has been overcome (Figures 3D–F, 4B).
This mechanism of Ire1 activation in our model contrasts with the two-step Ire1 activation model [15], in which unfolded proteins first trigger BiP dissociation from Ire1 to induce oligomerization, and subsequently bind to the oligomers to activate signaling. As opposed to separating oligomerization and activation into two steps, our model treats unfolded protein binding as the single activating step; Ire1 is in dynamic equilibrium with BiP and unfolded proteins, and its unfolded protein bound state is active. Thus, BiP dissociation, rather than triggering oligomerization, yields monomeric Ire1, which can then either bind to an unfolded protein and activate or re-bind to BiP. We note that the small Ire1bipless foci that formed in the absence of stress resulted in increased expression of INO1 mRNA and increased basal levels of UPR reporter fluorescence (Figures 1A, S5). Thus, we never observed inactive foci, in support of our model that oligomerization and activation occur in the same step.
In addition to this different mechanism of Ire1 activation, our model also proposes a mechanism for Ire1 deactivation. Since BiP and unfolded proteins compete for Ire1, BiP serves as a buffer that allows rapid deactivation of Ire1 as the concentration of unfolded proteins decreases. Finally, in contrast to the static picture of Ire1 activation presented in the two-step model, we present a time-resolved, quantitative model that accurately portrays Ire1 activation in response to any dose of DTT over time in its activation and adaptation phases.
While the computational model reflects our current understanding of Ire1 regulation, it is likely to be an oversimplification. Next generation models could easily improve the verisimilitude by including additional ER processes that are not currently represented in the model (such as glycosylation, ERAD, and BiP's ATP hydrolysis cycle) or better constraining the model parameters by targeted measurements. Yet even with increasing mechanistic detail the requirement for cooperative Ire1 deactivation is likely to persist (Figure S9). This feature, modeled as decreasing Hill function of active Ire1 molecules, is consistent with the notion that Ire1 signals through assemblies of high-order oligomers. As Ire1 oligomers grow in size or number, the percentage of Ire1 molecules that have the ability to be deactivated decreases as many molecules become captured inside macromolecular assemblies. Such cooperativity in Ire1 deactivation can be depicted intuitively as a simple steric consequence of Ire1 oligomerization (Figure S10).
Interestingly, this cooperativity can also be invoked to interpret the increased variability in foci formation in the Ire1bipless mutant cells (Figures S4 and S15). BiP's role can be thought of as a vehicle to help Ire1 traverse the threshold-like inactivation curve. In a wild type cell where focus formation might initiate stochastically, the presence of BiP can accelerate the dissociation of the foci. However, in an Ire1bipless mutant, any stochastically formed focus would be stable for a longer time (Figure 5C–E). If focus dissolution is an all-or-none process, an extreme scenario is one where Ire1 focus formation in wild type and Ire1bipless cells occurs as a pulse train whose low frequency of activation is the same in both populations. However, the duration of each pulse would be longer in Ire1bipless than in wild type cells. This simplified scenario would result in modest differences in foci formation as averaged over the population since the activation probability is itself low. It would nonetheless result in large variations around this average exhibited by individual cells. According to this view, BiP buffering would ensure that activated Ire1 signaling centers assume a more homogeneous size, providing for a consistent input/output relationship and consistent deactivation kinetics. As such, BiP buffering fine-tunes the UPR by filtering noise from the signal transmission process, thereby increasing the information content of the signal and improving the cell's homeostatic control of the ER.
This mode of regulation by which a free pool of a protein is buffered by chaperones may be a widely used mechanism in biology. For example, many kinases interact with cytosolic chaperones, and kinase signaling receptors that oligomerize during activation may hence be buffered similarly. Moreover, dynamic protein assemblies, such as clathrin coats or SNARE complexes, utilize chaperone interactions to aid disassembly [34],[35]. Insights gained from our understanding of the functional consequences of the interaction between BiP and Ire1 may therefore be generally applicable to many other systems, in which protein oligomers have to form and be broken down again in a highly controlled manner.
Reporter constructs and mutant alleles are genomically integrated into wild type or mutant strains. All experiments were conducted in complete, synthetic media (2×SDC: yeast nitrogen base, glucose, complete amino acids).
Ire1bipless is an allele of Ire1 that lacks the 51 amino acid juxtamembrane segment of the lumenal domain. This region is not in the crystal structure of the lumenal domain (Credle et al. [24]). Amino acids 475–526 of Ire1 were removed by 2-step PCR cloning and replaced with a 4 amino acid linker (Gly-Lys-Ser-Gly) on an episomal yeast plasmid (pRS315). The resulting positive, sequenced clone (Ire1bipless) was sub-cloned onto integrative plasmids (pRS305, pRS306), transformed into Ire1Δ cells (YDP002), and shown to complement for growth in the absence of inositol. Imaging constructs of Ire1bipless (GFP- and mCherry-tagged) were created by sub-cloning from the sequenced plasmid into the integrative wild type Ire1-GFP and Ire1-mCherry plasmids used for the FRET experiments. All experiments except the immunoprecipitations were conducted with genomically integrated Ire1bipless constructs.
We cultured cells in 2×SDC media to OD600 = 0.4, collected 50 ml per sample, washed cells in 1 ml 2×SDC and stored pellets at −80°C. Total RNA was extracted by resuspending cells in AE buffer (50 mM NaOAc, pH 5.2, 10 mM EDTA in DEPC-treated water), adding SDS to 1% and acid phenol (pH ∼4) (Fisher) to 50%, and heating at 65°C for 10 min. After spinning out the cell remains, we added chloroform and separated by centrifuging in phase-lock tubes (5 Prime). We precipitated the RNA with ethanol, washed with ethanol, and finally dissolved in 50 µl DEPC water.
RNA samples were quantified by spectrophotometry, added to loading buffer (1×E/formamide/formaldehyde/ethidium bromide/bromphenol blue), and heated at 55°C for 15 min. Samples were cooled on ice for 5 min and loaded. The gel is 1.5% agarose/20% formaldehyde/1×E and is run for 270 min at 100 V. Gels were transferred to nitrocellulose by wicking in 10× SSC for 24 h, and RNA crosslinked with 150 J. Blots were pre-hybridized in Church buffer for 3 h at 65°C, and hybridized overnight with random primer-generated probes from a HAC1 PCR product that incorporated α-32P-CTP using GE ready-to-go beads. Blots were washed in 2× SSC, sealed in plastic, exposed to phosphor-imager screens overnight, imaged with the storm scanner, and quantified with ImageQuant software.
We cultured cells bearing the SR or TR at 30°C in 2× SDC in 96 well deep well plates in an Innova plate shaker at 900 rpm. DTT stocks were made fresh from powder stored at 4°C for each experiment, and always 1 M in 10 ml. From this stock kept on ice, we prepared fresh 5× working stocks to start the experiment by diluting DTT in 1 step into 2× SDC to 37.5 mM (5×7.5 mM) in 10 ml. This 37.5 mM working stock was serially diluted by 1.5-fold increments (6 ml + 3 ml SDC) 10 times to span the range 0.13–7.5 mM. Every 2 h throughout the experiment, we repeated the full dilution series from the 1 M stock, making 1× dilution stocks in 2× SDC. To start the experiment, 200 µl of each 5× stock was added to 800 µl cells in the 96 well plates at time 0. The cells were incubated and shook at 30°C and were sampled every 30 min by 12-channel pipetting 75 µl of each culture into a 96 well microtiter plate. 5 µl of each 75 µl was subjected to flow cytometry analysis using a BD LSR-II equipped with a high throughput sampler, a 488 nm 100 mW laser, FITC emission filter, and FACS DIVA software to compile .fcs files. .fcs files were analyzed in MatLab and/or FloJo. No cuts or gates were applied to cell distributions. Median FITC-A values were calculated for each dose-time point and plotted in ProFit. Errors are calculated from the standard deviation of the median for 3 biological replicates.
We constructed the experimental FRET strain by co-expressing Ire1-GFP and Ire1-mCherry in the same cell from the endogenous IRE1 promoter integrated in the genome of an Ire1Δ strain and constructed bleed-through control strains by expressing either Ire1-GFP or Ire1-mCherry integrated alone in the deletion strain. FRET assays were performed using a Yokogawa CSU-22 spinning disc confocal on a Nikon TE-2000 inverted microscope equipped with 150 mW 488 and 562 nm lasers. Cells bearing the reporters were grown in 2× SDC to mid log phase, diluted to OD600 = 0.1, gently sonicated, and 80 µl added to 96 well glass bottom plates coated with concanavalin-A. Cells were allowed to settle for 20 min before imaging. DTT dilutions were prepared as 5× working stocks as in the titration time course experiments, and 20 µl added to wells at time 0.
Cells were imaged at each time point with 3×3 s exposures: 488 excitation/590 emission (GCh), 562 ex/590 em (ChCh), 488 ex/520 em (GG). Images were processed by first identifying cell boundaries and assigning the 16-bit fluorescence images to individual cells using the open-source cell-id software. Background was calculated by the mean intensity of areas in each fluorescent image not assigned to cells and subtracted from the cellular mean intensities to obtain corrected single cell values for GG, ChCh, and GCh.
The GCh value is a conglomerate of true FRET signal and fluorescent channel bleed-through from the individual fluorophores. The average GCh values from the single-fluorophore control strains were subtracted from the experimental strain GCh values to obtain final corrected values. FRET was calculated for each cell with the formula: F = GCh/(GG*ChCh)∧0.5.
For each time point at each dose, we obtained images of three different fields of cells, collecting a total of 30–60 cells per dose per time point. Mean values were plotted in ProFit and error bars represent the standard error of the mean.
Cells bearing C-terminally HA-tagged Ire1 or Ire1bipless expressed from the IRE1 promoter on 2 micron plasmids were cultured, collected, and stored in the same manner as for the Northern blot analysis. Cell pellets were thawed on ice, resuspended in 1 ml IP buffer (50 mM Tris-HCl, pH 8.0, 150 mM NaCl, 1% Triton X-100, protease inhibitors), and subjected to bead-beating (5×1 min, with 2 min on ice between iterations). Beads and cell debris were centrifuged and the cell free lysate was incubated with anti-HA conjugated agarose beads for 2 h at 4°C. Beads were spun, washed 5× with 1 ml IP buffer, and boiled in SDS-PAGE loading buffer.
Samples were run on BioRad ready-gels (4%–15% acrylamide, Tris/glycine/SDS) for 90 min at 35 mA. The proteins were subsequently transferred to Millipore Immobilon PVDF membranes at 220 mA for 45 min. Blots were blocked in 1% casein in TBS (10 mM Tris, 150 mM NaCl) for 30 min, followed by incubation with primary antibodies overnight. The rabbit polyclonal anti-Kar2 was used at 1∶5000 dilution, and the mouse anti-HA was used at 1∶2000. The next morning, the blots were washed 3× for 10 min with TBS, and then incubated with Li-Cor fluorescently-coupled secondary antibodies, goat anti-mouse 680 and 800, at 1∶10,000 dilution for 30 min. Blots were again washed 3× for 10 min with TBS, scanned with the Li-Cor infrared scanner, and processed with the Odyssey software package.
Wild type and Ire1bipless were cultured to OD600 = 0.4 in 400 ml 2×SDC at 30°. Cultures were brought to 500 ml and treated with 5 mM DTT for 1 h. Cells were sampled, filtered onto nitrocellulose membranes with 1 µm pores, washed with 100 ml 2×SDC, and then resuspended in 500 ml 2×SDC and returned to 30° incubation and sampled as indicated. For the FRET washout experiment, 1 ml cultures were spun, washed, resuspended, and imaged.
|
10.1371/journal.ppat.1003307 | A Refined Model of the Prototypical Salmonella SPI-1 T3SS Basal Body Reveals the Molecular Basis for Its Assembly | The T3SS injectisome is a syringe-shaped macromolecular assembly found in pathogenic Gram-negative bacteria that allows for the direct delivery of virulence effectors into host cells. It is composed of a “basal body”, a lock-nut structure spanning both bacterial membranes, and a “needle” that protrudes away from the bacterial surface. A hollow channel spans throughout the apparatus, permitting the translocation of effector proteins from the bacterial cytosol to the host plasma membrane. The basal body is composed largely of three membrane-embedded proteins that form oligomerized concentric rings. Here, we report the crystal structures of three domains of the prototypical Salmonella SPI-1 basal body, and use a new approach incorporating symmetric flexible backbone docking and EM data to produce a model for their oligomeric assembly. The obtained models, validated by biochemical and in vivo assays, reveal the molecular details of the interactions driving basal body assembly, and notably demonstrate a conserved oligomerization mechanism.
| Gram-negative bacteria such as E. coli, Salmonella, Shigella, Pseudomonas aeruginosa, and Yersinia pestis are responsible for a wide range of diseases, from pneumonia to lethal diarrhea and plague. A common trait shared by these bacteria is their capacity to inject toxins directly inside the cells of infected individuals, thanks to a syringe-shaped “nano-machine” called the Type III Secretion System injectisome. These toxins lead to modifications of the host cell, allowing the bacteria to replicate efficiently and/or to evade the immune system, and are necessary to establish an infection. As a consequence, the injectisome is an important potential target for the development of novel therapeutics against bacterial infection. In this study, we focus on the basal body, an essential region of the injectisome that forms the continuous hollow channel across both membranes of the bacteria. We have used an array of biophysical methods to obtain an atomic model of the basal body. This model provides new insights as to how the basal body assembles at the surface of bacteria, and could be used for the design of novel antibiotics.
| The bacterial injectisome, or type III secretion system (T3SS), is a specialized syringe-shaped protein-export system utilized by many pathogenic Gram-negative bacteria for the injection of virulence proteins (effectors) into host cells. Bacterial proteins destined for both needle assembly and host cell targeting are translocated via the injectisome in a process known as type III secretion [1]. The injectisome can be divided into three major regions: the inner- (IM) and outer-membrane (OM) spanning basal body with associated export apparatus and ATPase components, an extracellular needle, and a terminating pore inserted into the host cell membrane called the translocon (for review see ref [2], [3]). The major structural scaffold of the basal body is comprised largely of three proteins that arrange into a series of highly oligomerized, concentric rings: two intimately associated proteins localized to the IM - PrgK/EscJ/MxiJ and PrgH/EscD/MxiG, and a third protein belonging to the secretin family of OM proteins, InvG/EscC/MxiD (Salmonella enterica serovar Typhimurium SPI-1, Enteropathogenic Escherichia coli (EPEC) LEE and Shigella dysenteriae nomenclature, respectively) (Figure 1A) [4], [5], [6], [7], [8].
S. Typhimurium is an important medical pathogen causing gastroenteritis in infected individuals. Two of the major virulence determining factors are the discrete T3SSs encoded by the Salmonella Pathogenicity Islands (SPI) 1 and 2, which are required for bacterial invasion and replication within host cells [9]. The SPI-1 system, belonging to the mxi-spa evolutionary family which also includes the Shigella dysenteriae T3SS [10], is considered the prototypical T3SS and has been the focus of structural characterization using a variety of techniques including the first cryo-electron microscopy (EM) 3D reconstruction of a T3SS needle complex [7] revealing its supramolecular assembly. More recently, a cryo-EM analysis of the SPI-1 injectisome at ∼10 Å resolution has provided unprecedented detail of the overall architecture of the basal body [11]. We have previously published the x-ray crystallographically determined structures of the periplasmic domain of the S. Typhimurium basal body protein PrgH, in addition to the periplasmic domains of EPEC PrgK homologue EscJ and InvG homologue EscC [6], [12]. These structures defined a common modular domain (ring-building motif; RBM) hypothesized to be involved in ring oligomerization and have been used, along with the available EM data, in all subsequent studies modelling the assembly of these basal body components. The accuracy of these preliminary molecular models [11], [12], [13], however, has been hampered by the use of tentative homology models for some domains, and the lack of structural information entirely for others.
Here, we make new advances in compiling the precise molecular details of the Salmonella SPI-1 injectisome. Specifically, we report the crystal structures of the cytoplasmic domain of the IM ring protein PrgH11–120 (and a new crystal form of its periplasmic domain PrgH170–392 with improved resolution and detail), as well as the structure of the periplasmic domain of the OM ring InvG. (Figure 1B, Table 1). Collectively these structures provide the most cohesive set of T3SS basal body atomic-resolution structures known from a single species.
Using these structures, we have modelled the symmetric ring assemblies by flexible backbone symmetric docking guided by the above EM data (the higher-resolution map EMD-1875 was used unless specified), with the generalized symmetric modelling framework in the program Rosetta [14] (Figures 2 and S1; See Materials and Methods and Appendix S1 for details). This framework makes conformational sampling in symmetric systems tractable by 1) only considering conformations that are consistent with the symmetry of the system and 2) performing a minimal number of energy and derivative evaluations by explicitly simulating only a subset of the interacting monomers and propagating conformational changes to symmetry-related subunits. We implemented a two-step approach to explore symmetric ring conformations consistent with EM data: First, we used an initial global fixed-backbone search starting from a randomized orientation of the monomeric subunits incorporating a score term measuring correlation to the EM data [15], [16] (Figure 2A); this step aims to globally identify ring arrangements that are consistent with the EM data. The candidate fixed-backbone conformations identified in the first step are then explored locally in more detail using symmetric, all-atom refinement with full backbone flexibility [17] (Figure 2B). This step aims at optimizing the symmetric arrangements identified previously by capturing any (small) changes in the backbone conformation. To retain consistency to the EM data, an EM score term with reduced bias is used in the second step. Conformations with the lowest combined score (full atom energy and EM correlation score) from step 2 are then reported as the final models.
For both the cytoplasmic and periplasmic domains of PrgH, we applied a 24-mer oligomerization constraint, according to the established stoichiometry [4], [6], [11]. For the periplasmic domain, we observed that in preliminary modelling runs, residues 361–369 moved during the local perturbation step, and led to a model with a diameter not supported by the EM map, and relatively poor correlation. No visible density was observed for these residues in the crystal form reported previously [12], nor in one of the two molecules in the asymmetric unit of the crystal form reported here, suggesting that these residues are flexible. We therefore performed the docking procedures using residues 173–361 (Figure 2), which led to a model with excellent correlation to the EM map. We note that residues 370–392 were not resolved in any of the crystal forms obtained to date, and therefore these residues were not included in any of the modelling attempts.
The stoichiometry of the T3SS OM secretin has been a matter of debate, with numbers between 12 and 14 proposed [5], [12], [13], [18] and recent studies favouring 12 for the secretins of other systems [19]. Unexpectedly, the recent EM analysis of the SPI-1 basal body suggested a stoichiometry of 15 [11]. We therefore generated ring models with stoichiometries of 12, 14 and 15. To reduce the potential bias of the imposed map averaging, we have carried out the calculations using both the recent high-resolution EM map (EMD-1875) [11] and the previous lower resolution 20× averaged map (EMD-1100) [7]. Based on the docking results, the 12-mer model appears incompatible with the target region of the EM maps [4], having significantly lower interface energies and worse map correlation. Using the lower resolution 20× averaged map, the modelling of the 14- and 15-mer rings results in very similar interfaces with the 14-mer ring having better correlation to the EM map and the 15-mer ring having better all-atom energies. Use of the higher resolution map, however, clearly favours the 15-mer assembly, with the 14-mer configuration having significantly worse correlation to the EM map and Rosetta energy. Assuming a fixed-radius ring, it is expected to have a better-packed interface (and therefore better full-atom energies) for the more compact 15-mer ring than for the 14-mer. The 15-mer shows a similar interface regardless of the map used. Using the high-resolution map, the modelling converged on two opposite modes – a “helix in” conformation and a “helix out” conformation” referring to the orientation of the helix in the N1 domain of InvG (Figure S1B). The “helix in” mode shows excellent agreement with our previous biotinylation [12] and cross-linking [13] experiments, and better fit with the EM map.
To demonstrate the importance of using experimentally determined structures for our modelling protocol, we also generated ring models for the cytoplasmic domain of PrgH and the periplasmic domain of InvG using homology models based on the previously available structures MxiG [20] and EscC [12] respectively. For EscC, this strategy led to a model with a similar interface, but with significantly higher energy. In the case of MxiG this led to a model with a different interface, which did not converge during the all-atom refinement step if the PrgH11–120 structure was used in a similar conformation (data not shown), thus illustrating how the two-step modelling strategy allows for the discrimination between alternative ring arrangements. For molecular docking applications, the accuracy in the final docking solution is inherently limited by the precision in the backbone structure of the monomeric subunits. In the more general case an accuracy of 2 Å backbone RMSD or better in the coordinates of the monomer is needed to accurately predict the structure of the docked state [21]. Therefore, the availability of the high-resolution crystal structures of the monomeric subunits, rather than homology-based models, combined with the ability to internally rank models based on EM data, enables for high-resolution modelling of the symmetric ring complexes (assuming a minimal degree of change in the conformation of the backbone going from the monomeric to the complex state) (Figure 3, Figure S2).
To provide support for our modelled interfaces we have engineered a series of mutant variants and assayed their effect on secretion in Salmonella cultures. Mutants were designed using Rosetta to calculate maximal interface disruptiveness and by visual inspection of the models (Figure 4A). For the cytoplasmic domain of PrgH, we observed that mutation of two leucine residues to alanine or an electronegative glutamate (Leu20 and Leu87) abrogates secretion, while a more conservative mutation to tyrosine has no effect (Figure 4B, bottom). For the periplasmic domain of PrgH, we engineered numerous interface and surface mutations (Figure S3). Notably, we identified a loop (residues 319–324) that mediates several side- and main-chain contacts with the adjacent subunit. Side-chain mutations within this loop (K320L, T324L) have no effect on secretion; however, mutation of the conformationally labile Gly322 to a leucine abrogates secretion (Figure 4B, middle), suggesting that the loop main chain conformation/contacts are more critical to interface integrity. Finally, for InvG we observed a loop (residues 95–99) forming a series of side-chain contacts with the adjacent subunit. Mutation of Gln97 within the loop (which forms a hydrogen bond with the adjacent molecule) to leucine – but not alanine - abrogates secretion (Figure 4B, top). Mutation to alanine of the strictly conserved Asp95 within the loop (Figure S4) (which lacks any direct interactions with the neighbouring protomer) has no effect on secretion. In order to support the hypothesis that the observed secretion-deficient phenotypes were due to disruption of the oligomeric interfaces, circular dichroism analysis was used to ensure the introduction of the mutants did not merely deleteriously affect the fold of the individual domains (Figure S5). Further, we purified assembled injectisomes containing mutations in either of the PrgH or InvG domains. Negative stain EM analysis demonstrates that despite being necessary for secretion the PrgH cytoplasmic domain, as observed in a PrgH130–392 mutant (lacking the N-terminal cytoplasmic domain), is not required for assembly of the basal body (Figure 4C). However, mutant variant G322Y in the periplasmic domain of PrgH, or Q97L in InvG, disrupts needle complex assembly (Figure 4C).
Clearly, a question that arises from our analysis is the lack of phenotype for several predicted interface mutants, mainly for the PrgH periplasmic domain (Figure S3). One might argue that the complexity of the assembled injectisome, where multiple membrane spanning proteins, filaments and accessory proteins, as well as the membrane environment collectively act to stabilize the basal body rings, making the “all or nothing” action of a single point mutant in the context of the assembled T3SS more difficult. Ideally, the mutant phenotype would be assayed in isolated ring oligomers from individual domains, where they would be expected to be more deleterious. However, this is currently unfeasible due to unsuccessful attempts to reconstitute such T3SS sub-assemblies in vitro.
From our analysis, each of the three rings shows an excellent fit to the EM map with correlation coefficients of 0.95, 0.92 and 0.94 for PrgHcytoplasmic, PrgHperiplasmic and InvG models respectively (Figure 3B, Figure S6 and Video S1). The model of the IM PrgH periplasmic ring is broadly similar to previously reported models from our groups and others [11], [13]. The corresponding region of the EM map is rich in detail and the availability of our previously published crystal structure of this domain facilitated the accurate positioning in the map (Figure S6) and reconstruction of the oligomer [11]. Nonetheless, the average backbone RMSD between the presented model and those previously deposited (PDB ID 2Y9J) is 2.9 Å, corresponding to a small rotation of the monomer subunits (Figure S7). Importantly, our PrgH periplasmic domain model possesses fully refined interfaces (Figure S2B).
For the cytoplasmic domain, the structure of the orthologues from Chlamydia (CdsD) (PDB ID: 3GQS), Shigella (MxiG) [20], [22] and Yersina (YscD) [23], [24] recently confirmed predicted homology to the family of forkhead associated (FHA) domains (Figure S8). Tentative ring models were proposed for MxiG [20], [22]; however, comparison to the model presented here is not possible in the absence of available coordinates for the Shigella variant (although from the published figures the models appear to be generally oriented in a similar fashion). FHA domains are frequently involved in interaction with phosphothreonine (pThr)-modified proteins. In Shigella, the PrgH orthologue MxiG was reported to interact with phosphorylated peptides from the secretion apparatus protein Spa33 [20] although a separate study failed to corroborate these results [20], [22], and structures of the Yersinia orthologue YscD revealed an unconserved phosphothreonine binding motif [23], [24]. Protein sequence alignment of the PrgH orthologues indeed illustrates that the conserved residues involved in phosphopeptide binding in FHA domains are poorly conserved (Figure S9). To validate this, we engineered mutations of the PrgH residues corresponding to the proposed phosphothreonine interacting residues in MxiG - Arg35, Gln42 and Asp65 - to alanine. These mutations have no effect on in vivo secretion assays (Figure S10). We therefore conclude that the PrgH cytoplasmic domain is unlikely to interact with pThr-modified proteins. We note however that the FHA phosphopeptide interacting loops in our PrgH structure are accessible on the cytoplasmic face of the ring model, suggesting a potential interface for protein-protein interactions. Indeed, deletion of this domain results in the formation of secretion incompetent immature basal body assemblies lacking needles (Figure 4C) suggesting a role in the correct assembly/coordination of the inner-membrane export apparatus.
For the InvG periplasmic domain, our newly generated ring model is significantly different from existing ones [11], [12], with an average backbone RMSD of over 5.5 Å between our model and the previously deposited one (PDB ID 2Y9K) (Figure S2A and S7)). Importantly, the fully refined interface in our model does not present the atomic clashes present in the previous model. This model was based on homology modelling from the distant EPEC orthologue EscC (22% sequence identity between the periplasmic domains, Figure S4). Superposition of the InvG and EscC crystal structures revealed a similar organization, with the N0 and N1 domains common to the secretin family [19], [25]; however, the relative orientation of these domains is shifted 36 degrees between the two proteins (Figure S11).
Collectively therefore, the basal body model reported here marks a significant advancement from prior models. Of note, the modelling protocol clearly favoured a 15-mer stoichiometry for InvG, in support of recent EM analysis [11]. This result is however in contradiction with the stoichiometry of other secretin proteins [5], [19], [26], [27], [28], and further studies will be required to confirm if this represents a true system-to-system variation.
We have previously observed the presence of a common modular domain, which we termed ring-building motif (RBM), in all the proteins comprising the basal body as well as components of the IM export apparatus [12], [29], [30]. Structural comparison of the intra-subunit interactions of the RBMs in the basal body model supports our previous hypothesis that this domain functions as an oligomerization scaffold, with a conserved interaction mechanism in three RBM mediated interfaces of PrgH and InvG. In all three cases, the conserved interface consists of the N-terminal helix packing against the three-stranded beta sheet of the motif (Figure 5A). Analysis of the electrostatic profile at the RBM interface suggests the driving force for self-association is the complementary charge between the surface formed by the three-stranded beta-sheet and the surface formed by the two helices (Figure 5B). Importantly, similar interactions are present in the oligomeric structure of the EPEC basal body component EscJ [6], suggesting that the RBM plays a common species-independent role in the assembly of the three basal body components. Further, the subsequent observation of an RBM in the T2SS secretin [25], [31] and the intercellular channel complex in sporulating Bacillus subtilis [32], [33] suggests this mechanism is likely applicable to related bacterial systems. It should be noted that the N-terminal RBMs of PrgH (defined in the new crystal form of PrgH170–392 reported here) and EscJ do not form the conserved oligomerization interface and appear involved in intra-domain interactions or membrane association.
All the reported purified domains of the T3SS basal body are monomeric in solution, showing that the membrane-embedded domains and/or the presence of additional proteins are necessary for their oligomerization. Nonetheless, in our earlier structures of the IM protein EscJ from EPEC we observed a 24mer oligomeric packing generated around the 6 fold screw axis of the crystals, the first direct insight into the oligomerization number and interfaces of the ring which has subsequently been supported in re-evaluation of higher resolution EM analysis [11] and has been the basis for all subsequent models of that component and orthologues in the literature. Prompted by this, we investigated if the crystal packing in the structures reported here could be correlated to the interfaces found in our oligomeric models. We observed that the cytoplasmic domain of the second basal body IM ring, PrgH11–120, forms a hexamer in the asymmetric unit of the crystal (Figure S12A), with the two interfaces involved in generation of the hexamer being the most sequence conserved surface region (Figure S12B). Comparison of our subsequent Rosetta-EM generated PrgH11–120 24-mer and hexameric crystal packing reveals that the surfaces used for protein-protein contacts is the same in the 6-mer of the crystal form and in the 24-mer of the oligomeric model (Figure 6A). A ∼20° subunit rotation in the hexamer of the crystal structure accommodates the tighter packing but superposition of the interface secondary structural units shows significant conservation with both interfaces having similar buried surface areas (640 Å2 vs. 546 Å2 for hexamer and 24mer respectively). For the periplasmic PrgH170–392 domain, crystallographic symmetry-related molecules in the crystal produce a dimer that exploits the same general interface as the 24-mer biological assembly (Figure 6B). Similarly, a dimer formed by a unit cell translation from the InvG crystal lattice superposes with a dimer from the modelled 15-mer ring (Figure 6C). Collectively, the crystallographic packing interfaces we observe support and mirror the low energy interfaces generated in our modelled basal body. This would be consistent with a model whereby the soluble domains have low-affinity interaction surfaces for oligomerization, which can be captured by the high concentration in the crystallization experiment. In vivo, oligomerization is likely dependent on membrane localization and potentially nucleation by additional members of the injectisome.
Finally, analysis of the surface electrostatics of the InvG and PrgH periplasmic ring models reveals an acidic surface present on the face of the InvG ring proximal to PrgH, and a complementary basic surface on the corresponding face of PrgH (Figure 5C). This charge complementarity may provide an initial force of attraction in the assembly of the inner and outer membrane components of the basal body. Furthermore, we observed a positively charged interior collar of the InvG ring corresponding to the N0 domain (Figure 5C), where residues are proposed to be in contact with the inner rod and socket [11], [34]. This surface is formed by a set of basic residues lying on one side of the InvG monomer (Figure S13). In particular, we have previously shown that mutation of Lys67 abrogates secretion by altering substrate switching [13]. From these observations, we can propose a model whereby a large number of weak, charge-based interactions between individual subunits of the basal body lead to a stable complex upon assembly. This is most likely the reason why all the purified soluble domains do not oligomerize in vitro, in the absence of the stabilizing trans-membrane domains and lipidic membrane environment. Similar multivalent interactions probably govern the interaction between the basal body and other components of the injectisome, namely the rod and needle. This model is in agreement with the observation that extreme pH conditions lead to the disassembly of the injectisome [7], [11].
While the agreement with the experimental data is compelling, it should be recognized that we have provided models of the ring assemblies, not experimentally determined high resolution structures. In particular, we assume in the initial docking calculations that there are no large scale conformational changes between the monomeric and oligomeric states of individual domains, and that all subunits are identical in the assembled basal body. Furthermore, since the monomers do not spontaneously oligomerize, it is possible that other components, such as the membrane scaffolding, influence the conformation of the oligomeric assembly.
In summary, we report the crystal structures for the cytoplasmic and periplasmic domains of PrgH and InvG, two of the main basal body components of the prototypical Salmonella SPI-1 injectisome. We have refined these structures into the EM density using Rosetta symmetric flexible backbone docking calculations, generating ring models for these domains that exhibit a high degree of correlation to the Salmonella SPI-1 basal body EM map. The modelling procedure produced converged, low-energy interfaces, which were validated by in vivo functional assays. The obtained models provide insights into the discrete interactions occurring during assembly of the T3SS basal body. Analysis of the packing of the previously identified modular domain common to the three basal body components confirms that it represents a common ring-building motif with an electrostatically-driven oligomerization mechanism likely conserved amongst the many clinically important bacteria that rely on a T3SS for their pathogenic effects.
The gene coding for PrgH11–120 was amplified by PCR from Salmonella enterica serovar Typhimurium genomic DNA and cloned into the pET-28(a) plasmid (Novagen) fused to a 6xHis tag at the N-terminus followed by a thrombin cleavage site, using restriction-free PCR [35]. The obtained pET-PrgH11–120 was transformed into BL21(DE3) competent cells, and kanamycin-resistant colonies were used to inoculate LB media containing 50 µg/ml of kanamycin, at 37°C to an OD600 of ∼0.5. Expression was induced by IPTG at 0.1 mM, and the protein was expressed at 20°C for 20 hours. Cells were harvested, resuspended in buffer (50 mM TRIS pH 7.4, 300 mM NaCl, 20 mM imidazole), lysed by sonication and debris pelleted at 45,000 g for 50 min. The protein was purified from the supernatant by passing it through Ni-activated chelating sepharose, and the His-tag was subsequently removed by adding thrombin (Roche) at 1∶1000 dilution for 16 hours at 4°C. The protein was further purified by size-exclusion chromatography using a Superdex75 gel filtration column (GE Healthcare). Expression and purification of PrgH170–392 was performed as described previously [12].
The gene coding for InvG22–178 was amplified by PCR from S. Typhimurium genomic DNA and cloned into the pET-28(a) plasmid (Novagen) fused to a 6xHis tag at the N-terminus followed by a thrombin cleavage site, using restriction-free PCR[35]. The obtained pET-InvG22–178 was transformed into BL21(DE3) competent cells, and kanamycin-resistant colonies were used to inoculate LB media containing 50 µg/ml of kanamycin, at 37°C to an OD600 of ∼0.5. Expression was induced by IPTG at 1 mM, and the protein was expressed at 37°C for 5 hours. Cells were harvested, resuspended in buffer (50 mM HEPES pH 6.8, 150 mM NaCl), lysed by sonication and debris pelleted at 45,000 g for 50 min. The protein was purified from the supernatant by passing it through Ni-activated chelating sepharose, and the His-tag was subsequently removed by adding thrombin (Roche) at 1∶1000 dilution for 16 hours at 4°C. The protein was further purified by size-exclusion chromatography using a Superdex75 gel filtration column (GE Healthcare).
For SeMet-labelled protein, bacteria were grown in minimal media containing 100 mgs each of added L- Lysine, Phenylalanine, and Tyrosine; 50 mgs of L- Isoleucine, leucine, and valine; and 60 mgs selenomethionine per litre. Purification was performed as for unlabelled protein.
All initial crystallization trials were performed by sitting-drop vapour diffusion using a Phoenix drop setter (Rigaku). Crystals of PrgH11–120 (5–10 mg/ml) were grown at 20°C by sitting-drop vapour diffusion using 15–20% PEG 6000, 0.02 M CaCl2, 0.1 M HEPES pH 6.5 as reservoir solution. Crystals of PrgH170–392 were obtained at 20°C in a range of conditions using protein samples at 2 mg/ml concentration or less, and all requiring the presence of PEG precipitants. The crystals were all of a new crystal form (compared to those previously-reported by our group [12]), and optimal diffraction was obtained for crystals grown in 100 mM bicine pH 8.5, 20% PEG 6000.
InvG22–178 crystallized in several conditions at 20°C, with pH above 7.5 and containing PEG precipitating agents, but formed clusters of needles not amenable for data collection. Single crystals could be obtained by decreasing the protein concentration to approximately 4 mg/ml, and the best diffraction was obtained for crystals grown in 100 mM HEPES pH 8.0, 30% jeffamine M-600. Selenomethionine-labelled protein crystallized similarly.
Crystals were cryo-protected by soaking in the crystallization condition supplemented with 30% glycerol and flash-cooled in liquid nitrogen. Data were collected at beamline 8.1 of the Advanced Light Source (ALS) and beamline CMCF-1 of the Canadian Light Source (CLS) at 100 K. For PrgH11–120, a mercury chloride derivative was obtained by soaking in cryo-protectant +2 mM mercury chloride for 10 minutes. A three wavelength MAD experiment was subsequently carried out as well as collection of a high-resolution native dataset in the absence of mercury chloride. For InvG22–178 a three wavelength MAD experiment on selenomethionine-derivative crystals was carried out, and a native dataset was collected.
Data were processed and scaled with HKL2000 [36] and MOSFLM [37]/SCALA [38]. For PrgH170–392, a molecular replacement solution was found with the program PHASER [39] using the PDB file 3GR0 as a search model, and an initial model was built using ARP/WARP [40]. For PrgH11–120 and InvG22–178, structure determination and initial model building were carried out with the PHENIX suite [41], which identified 12 Hg atoms for PrgH11–120 (FOM 0.51) and 6 Selenium sites for InvG22–178 (FOM 0.54). All models were further refined with REFMAC5 [42] and PHENIX, using TLS parameters [43]. Data processing and model refinement statistics are summarized in Table 1. Structure quality was assessed with PHENIX and all models have good stereochemistry with PrgH11–120, PrgH170–392 and InvG22–178 having respectively 96.46%, 98.1% and 97.9% of residues in the favoured region of the Ramachandran plot with no outliers. We note the presence of four PEG molecules and four phosphate ions per asymmetric unit in the PrgH170–392 crystal structure, for a total of 76 ion/ligand atoms. These molecules were present in the crystallization condition or cryoprotectant buffer, and are therefore not likely to be biologically relevant. The coordinates for PrgH11–120, PrgH170–392 and InvG22–178 have been deposited to the PDB, under the accession numbers 4G2S, 4G08 and 4G1I respectively.
Rosetta modelling made use of standard symmetric docking protocols [14], [44] augmented with a term assessing agreement to experimental cryo-EM density [15]. For all symmetric assemblies considered here, symmetric docking calculations were performed in two steps: in a first step, fixed-backbone docking calculations were performed using the monomer conformations from the crystal structures in a randomized orientation. This procedure typically resulted in the identification of a small number of local minima, suggesting potential binding modes for each assembly. In the second, refinement step, we performed a fine-grained local search starting from each binding mode identified in step (1). Here, the rigid body degrees of freedom underwent small random perturbations in a number of Monte-Carlo trajectories. Finally, for each trajectory we performed gradient-based optimization of all backbone, side-chain and rigid body degrees of freedom. The command lines used for each procedure are deposited in Appendix S1. The coordinates for the PrgHcytoplasmic, PrgHperiplasmic and InvG ring models have been deposited to the PDB, under the accession numbers 3J1W, 3J1X and 3J1V respectively.
For complementation assays, mutants were engineered into a plasmid containing the genes coding for PrgH or InvG, described previously [13], [45], using the QuickChange mutagenesis kit (QIAGEN). The obtained plasmids were transformed into electro-competent PrgH- or InvG-deletion strains of S. Typhimurium. Secretion assays were performed as described previously [13], [45]. Briefly, 5 ml LB cultures of S. Typhimurium strains were grown at 37°C overnight, and cells were then pelleted at 6,000 g for 10 min. Proteins in the supernatant were precipitated by adding 10% TCA and pelleted at 6,000 g for 30 min. Pellets were washed in 0.5 ml acetone, re-suspended in 20 µl gel loading buffer, boiled, and ran in a 10% acrylamide SDS-PAGE gel that was stained with Coomassie Blue.
Electro-competent strains of S. Typhimurium lacking the gene for FliC and either InvG or PrgH [45] were transformed with a plasmid containing the genes coding for PrgH or InvG, described above, or the corresponding mutants. These transformants were used to make electro-competent cells, which were transformed with a plasmid containing the gene coding for the T3SS transcription activator HilA. The obtained transformants were then used for purification of the injectisome, as described previously [45].
For EM analysis, purified injectisome samples were diluted in 10 mM Tris pH 8.0, 500 mM NaCl, 5 mM EDTA and 10 mM LDAO. They were prepared on carbon grids and stained with 0.75% uranyl formate using standard procedures. Images were collected with a H7600 Transmission Electron Microscope (Hitachi Hi-Technologies Canada, Inc.) equipped with a side mount AMT Advantage (1 mega-pixel) CCD camera (Hamamatsu ORCA), and operated at an acceleration voltage of 120 kV.
Circular dichroism (CD) spectra were recorded with a nitrogen-flushed Jasco J-810 spectro-polarimeter, at 20°C. Proteins were dialyzed against buffer containing 5 mM Tris pH 8 and 50 mM NaCl prior to analysis. 0.1–0.5 mg/ml protein was used with a path length of 0.1 cm. Data were recorded from 260 to 190 nm using a 2 s time constant, 10 nm min−1 scan speed and a spectral bandwidth of 1 nm. Spectra were corrected for buffer.
The multiple sequence alignments (Figures S4 and S9) were made with ClustalW [46] and the figures generated with ESPript [47]. RMS distances were calculated with PyMol (Schrodinger, LCC). Electrostatic surfaces were calculated with the APBS module [48] in PyMol. Domain angle differences were measured with HingeFind [49]. Conserved residues were mapped on the structures with ConSurf [50]. Map fitting was performed with Chimera [51]. All structure figures were generated with Pymol or Chimera.
|
10.1371/journal.pbio.2003489 | Quantifying the effects of temperature on mosquito and parasite traits that determine the transmission potential of human malaria | Malaria transmission is known to be strongly impacted by temperature. The current understanding of how temperature affects mosquito and parasite life history traits derives from a limited number of empirical studies. These studies, some dating back to the early part of last century, are often poorly controlled, have limited replication, explore a narrow range of temperatures, and use a mixture of parasite and mosquito species. Here, we use a single pairing of the Asian mosquito vector, An. stephensi and the human malaria parasite, P. falciparum to conduct a comprehensive evaluation of the thermal performance curves of a range of mosquito and parasite traits relevant to transmission. We show that biting rate, adult mortality rate, parasite development rate, and vector competence are temperature sensitive. Importantly, we find qualitative and quantitative differences to the assumed temperature-dependent relationships. To explore the overall implications of temperature for transmission, we first use a standard model of relative vectorial capacity. This approach suggests a temperature optimum for transmission of 29°C, with minimum and maximum temperatures of 12°C and 38°C, respectively. However, the robustness of the vectorial capacity approach is challenged by the fact that the empirical data violate several of the model’s simplifying assumptions. Accordingly, we present an alternative model of relative force of infection that better captures the observed biology of the vector–parasite interaction. This model suggests a temperature optimum for transmission of 26°C, with a minimum and maximum of 17°C and 35°C, respectively. The differences between the models lead to potentially divergent predictions for the potential impacts of current and future climate change on malaria transmission. The study provides a framework for more detailed, system-specific studies that are essential to develop an improved understanding on the effects of temperature on malaria transmission.
| Many of the mosquito and parasite life history traits that combine to influence the transmission intensity of malaria (e.g., adult mosquito longevity, biting rate, the developmental period of the parasite within the mosquito, and the proportion of mosquitoes that become infectious) are strongly temperature sensitive. Yet, in spite of decades of research, the precise relationships between individual traits and temperature remain poorly characterized. As a consequence, the majority of studies exploring the influence of local environmental conditions, or prospective impacts of climate change, draw on a combination of studies that utilize different experimental methods and a range of mosquito and parasite species. Here, we use the Indian malaria mosquito, Anopheles stephensi, and the human malaria parasite, Plasmodium falciparum, to thoroughly characterize the influence of temperature on key transmission-related traits. The results reveal a number of novel insights and challenge some longstanding assumptions regarding the nature of mosquito and parasite thermal responses. This study provides an experimental blueprint for further system-specific studies necessary to more fully understand the implications of changing temperatures on malaria transmission.
| Numerous studies show the transmission of malaria to be strongly influenced by environmental temperature [1–10]. This has led to a large body of both theoretical and empirical research which explores the possible effects of temperature on the dynamics and distribution of malaria both in the present day [6, 11–16] and under scenarios of future climate change [4, 7, 8, 17–23]. In spite of the accepted importance of temperature, the thermal sensitivity of individual mosquito and parasite traits that combine directly or indirectly to determine transmission intensity (i.e., adult longevity, biting rate, fecundity, generation time, vector competence, and parasite extrinsic incubation period [EIP]) remains surprisingly poorly characterized. For example, a recent study that explored the influence of temperature on transmission rate of P. falciparum in Africa utilized data from a Latin American malaria vector for biting rate, a North American vector infected with P. vivax for vector competence, a mixture of 6 malaria vector species from Asia, Africa, and North America for parasite development rate, and even a nonmalaria vector (Aedes albopictus) for fecundity [4]. Many other studies rely on similar data [6, 8, 13, 17, 24–26]. The necessity to combine insights from such disparate systems highlights the paucity of data.
Similarly, the iconic degree-day model developed in the 1960s to define the EIP (also called the period of sporogony) of P. falciparum inside the mosquito vector [27, 28] has been applied in a multitude of studies over the years. Yet, it is rarely acknowledged that this relationship derives from a limited number of experiments conducted in the 1930s and 1940s by using Russian populations of native Mediterranean mosquitoes (An. maculipennis). Furthermore, many other historical experiments that explore temperature-sensitive parasite development rate are very poorly replicated (data points sometimes based on single mosquitoes), contain little or no information about the sources of infectious blood or blood infection levels, and explore a limited temperature range [29–33]. Whether these data are sufficient to describe parasite development rate in every malaria transmission system seems extremely unlikely, yet that is the prevailing assumption.
Here, we present a detailed investigation of how temperature affects key mosquito and parasite traits for a single species pairing of An. stephensi and P. falciparum across a range of temperatures relevant to malaria transmission. Specifically, we measured adult mosquito mortality rate, the duration of the EIP (the time for parasites to reach their infectious stage), vector competence (the maximum prevalence of infectious mosquitoes), and biting rate across several temperatures from 21°C to 34°C. We then use these data to generate temperature-dependent models of relative vectorial capacity (rVC) and relative force of infection. Our data and the contrasting models highlight the need to improve empirical understanding of the effects of temperature on malaria transmission in addition to providing an experimental framework for conducting future species-specific research across a range of vector-parasite pairings.
Our principal experiment involved feeding replicate cups of approximately 150, 3 to 5 day-old adult female An. stephensi mosquitoes on human blood infected with P. falciparum. We then transferred these mosquitoes to temperature-controlled incubators set to 21, 24, 27, 30, 32, and 34°C (our pilot experiment also contained groups at 16°C and 18°C, but no sporogony had occurred through day 26 postinfection, so those groups were excluded from further experiments; see S1 Table). Mosquitoes were monitored to assess daily mortality, and subsamples were removed for dissections at set intervals (see Materials and methods) to track the time it took for parasites to invade the salivary glands, and hence, the distribution of time over which mosquitoes became infectious. The experiment was repeated over 2 independent experimental blocks.
We found that mosquito mortality rate was significantly different across temperature (log-rank test, χ2 = 533, df = 5, p < 0.001) and across temperature and block (log-rank test, χ2 = 569, df = 11, p < 0.001). Upon pairwise log-rank comparison analysis, we found several instances of significant effects of block x temperature interactions (S2–S6 Tables). These interactions are likely due to decreased initial mortality in the second block, which allowed for increased sampling time, especially across warmer temperatures (30°C to 34°C). Nonetheless, mortality tended to increase with temperature across both experimental blocks (Fig 1).
Due to logistic constraints in generating large numbers of infected mosquitoes, we tracked both mortality and sporogony within the same mosquito samples and, thus, do not have survival data that encompass the entire mortality curve without censored cases. We used parametric survival analysis to characterize survival of each temperature group (S7 and S8 Tables). We fit several survival distributions (Gompertz, Weibull, exponential, log-logistic, and log-normal) to our data and compared the fits using the Aikake information criterion (AIC) (S7 and S8 Tables). These models have been used to describe survival curves of a diversity of insects in laboratory and field settings [35–42]. Our cumulative survival data were best described by a Gompertz function, in which the mortality hazard increases exponentially with age:
f(x)=αeβx
where x is a given age, α is the initial mortality rate, and β is the constant exponential mortality increase with age [26, 35, 43, 44]. Because of significant interactions driven by block differences, we built a Gompertz function that describes each temperature x block combination separately (Fig 1, Table 1). Overall, block 2 exhibited higher median survival times, along with lower initial mortality rate (α) values than block 1, regardless of the age-dependent exponential increase in mortality (β).
Dissection of mosquitoes revealed an increase in the prevalence of sporozoite-infected mosquitoes over time in all temperatures (Fig 2). To describe this pattern and enable comparisons of EIP between temperatures, we fit a basic logistic curve to the data for each temperature in both experimental blocks:
gx=g1+e−k(x−t)
where at any given day x, the proportion of infectious mosquitoes is dependent on g (the asymptote), which is the maximum sporozoite prevalence and provides a measure of vector competence, k (a rate constant), which describes the instantaneous change in proportion of infectious mosquitoes through time, and t (the inflection point), the time at which 50% of maximum proportion infectious is reached [1,45,46].
At 21°C, 24°C, and 27°C, sporogony was well described by the logistic function applied to the full data series. However, at 30°C, 32°C, and 34°C, the raw data showed an initial increase in prevalence of infectious mosquitoes followed by a decline, which was not accurately described by the logistic function alone (Fig 2). For subsequent modeling analysis (see section on Transmission potential below), we truncate the logistic curves at the day of the peak proportion of infectious mosquitoes and fit exponential curves to characterize the decline in prevalence (calculated by using nonlinear least squares regression in R; S9 Table). This truncation does not affect the calculation of the maximum proportion infectious, the rate constant, nor the inflection point.
EIP is very poorly defined in the literature [4,47–54], so we provide here 3 estimates for each temperature to represent a range of possible interpretations of EIP (i.e., time in days to 10%, 50%, and 90% of maximum infectiousness). In Fig 3, we show how EIP10, EIP50, EIP90, and vector competence (maximum proportion infectious, or g, the asymptote of the logistic curve) values change with temperature in each experimental block (values for logistic model parameters are given in S10 Table). EIPs were shortest at 34°C and increased at cooler temperatures, irrespective of the specific measure of EIP (i.e., EIP10 increased from an average of 6.1 to 11.2 days, EIP50 increased from an average of 7.0 to 15.1 days, and EIP90 increased from an average of 8.0 to 19.0 days; see S11 Table for details). Additionally, the relative and absolute difference between EIP10 and EIP90 increased progressively under cooler conditions (Figs 2, 3A and 3B).
We next conducted an experiment to determine the effect of temperature on the gonotrophic cycle length (i.e., the time from blood meal to oviposition), taking the mean of the first 2 gonotrophic cycles for each temperature. We found that gonotrophic cycle length declined with increasing temperature, although with differences between cycle lengths diminishing as temperature increased (Fig 4). The percentage of mosquitoes laying eggs was lower in the second gonotrophic cycle compared to the first, but there was no obvious effect of temperature on the likelihood of egg laying in either cycle (S12–S14 Tables).
Our empirical data enable us to generate thermal performance curves for biting rate, vector competence, mosquito mortality rate, and parasite development rate.
For estimates of daily biting rate, we followed convention by taking the reciprocal of the mean gonotrophic cycle length [51,52,54–57]. Some mosquitoes, such as An. gambiae and An. funestus, have been shown to take multiple blood meals per gonotrophic cycle [58]. However, there are no data from the field to suggest this behavior for An. stephensi.
For vector competence, we used values of the asymptote (g) of our logistic curves, while for parasite development rate we used the reciprocal of the median EIP (EIP50).
Generating a thermal performance curve for daily mosquito mortality rate is challenging, as we show mortality rate to increase with mosquito age. Accordingly, we follow the methodology described in [4] to fit negative exponential functions to the beginning and end points of the Gompertz distributions and use the exponent to approximate a fixed daily mortality rate for each block and temperature combination (see electronic supplementary material for further methodology and accompanying datasets and figures).
We present the thermal performance curves for these traits in Fig 5, together with equivalent thermal performance curves from the study of Mordecai et al. [4]. Our thermal performance curves exhibit quantitative and qualitative differences to the established thermal performance curves derived from mixed-species data (for additional information comparing specific nonlinear models between this paper and those published previously, see S16 Table).
Having defined the effects of temperature on biting rate, mortality rate, EIP, and vector competence, it is possible to characterize the overall effects of temperature on transmission potential using a metric such as rVC. The rVC is the daily rate at which mosquitoes can transmit parasites to humans (relative to the vector-to-human population ratio, which here we do not define) [52–54, 56, 59], and is described by the following equation:
rVC= a2be(−μn)μ
where a is the daily biting rate, b is vector competence, μ is the daily mosquito mortality rate and n is the length of the EIP50 (see S2 Fig, S15 Table for calculations of rVC using EIP10 and EIP90 as alternatives).
In Fig 5E we show the temperature-dependent model of rVC based on our thermal performance curves. This model suggests a temperature optimum for transmission of 29°C, with an upper maximum threshold of 38°C and a minimum of 12°C. However, standard vectorial capacity models [1, 53–57, 59] and related models such as the basic reproductive rate, R0 [4, 51, 60], assume constant mortality rate per day, a discrete value for EIP at a given temperature, and no change in the proportion of infectious mosquitoes over time (so, no recovery from infection or altered survival rates due to infection). Our empirical data violate these assumptions, and so we also explore the effects of temperature on relative transmission potential using an alternative measure, adapted from the work of Killeen et al. [61] (see also [45,46] for similar methods). This model explicitly uses the full distributions for survival, sporogony, and competence by multiplying the number of mosquitoes alive on any given day (values from our survival curves) by the probability of being infectious (values from our curves for change in proportion of infectious mosquitoes over time). The product of these 2 proportions (area under the intersection of the 2 curves in Fig 6) represents the daily number of infectious mosquitoes, which we term “infectious mosquito days.” The number of infectious mosquito days is then multiplied by our empirical estimates of biting rate for each temperature to give the probable number of infectious bites transmitted by a cohort of mosquitoes over a given time period, assuming all blood meals are taken on humans; this is analogous to the relative force of infection.
Because our survival data were truncated, we extended our survival estimates out to day 50, which is the point at which mosquito survival dropped below 1% in the longest-lived temperature treatment. For each temperature, we used the mean of the daily values calculated from the empirical survival and data from both blocks, and used the model fits only for the tails of distributions where we have no raw data.
In Table 2, we provide values for the predicted number of infectious bites transmitted by a cohort of 100 mosquitoes over a period of 50 days for each temperature (see S3 Fig for schematic outlining the model approach).
We then fit a nonlinear curve (in this case the best fit model was a simple quadratic function) to our data points for mean transmission potential at each temperature (Fig 7A). The resultant thermal performance curve of relative force of infection suggests a temperature optimum of 26°C, with lower and upper thresholds of 17°C and 35°C, respectively. In Fig 7B, we compare the models of rVC and relative force of infection.
This study represents one of the most detailed empirical investigations of the effects of temperature on P. falciparum and a key mosquito vector conducted to date. The results yield several important insights that challenge classical assumptions. We also provide an experimental blueprint for future species-specific explorations of temperature-mediated changes in malaria transmission, and such data will be essential for future predictive and theoretical modeling studies.
Most conventional malaria transmission models assume constant rate of adult mortality per day [1, 4, 50, 53–55, 59]. In contrast, we show that mosquito mortality rate is most accurately described using an age-dependent distribution, in this case, a Gompertz function. Other studies also suggest age-dependent survival (e.g., [40, 62–65]). Here, we show that the age-dependence holds across temperatures, with initial senescence rate (α) increasing at higher temperatures. We acknowledge that ours is a lab-based study and that the empirical survival data were necessarily truncated because of destructive sampling of mosquitoes for dissection (i.e., Kaplan-Meier estimates are interval-censored). Future studies would benefit from monitoring survival in a parallel group of mosquitoes that receive the same infectious blood meal but with no samples removed for dissection.
A recent analysis of survival data for Aedes mosquitoes indicated that age-dependent mortality is more likely to be found in the laboratory as external mortality factors such as predation, encountering insecticides, and nutritional stress tend to be more prevalent in the field, confounding the effects of senescence [40]. However, a previous field study on geographically-distinct populations of A. aegypti in Southeast Asia and Latin America suggests that age-dependent mortality is observable in the field, and that older mosquitoes die at quicker rates than younger mosquitoes in the same cohort [35]. Whether our data are an artifact of prolonged survival under laboratory conditions is unclear, as very few data exist on the nature of the mortality distributions of adult Anopheles spp. under field conditions.
Existing models of P. falciparum sporogony, either the classic degree-day models [27, 28] or more contemporary thermal performance curves [4], provide a single estimate of EIP for a mosquito population at a given temperature. We show that sporogony is not perfectly synchronized between individual mosquitoes, but instead follows a distribution across the mosquito population over time, with temperature affecting the extent of the distribution (i.e., the number of days over which sporogony occurs), as well as the median value (see also [1, 46]). Most empirical studies are vague in reporting whether they are defining the EIP as the first mosquitoes to become infectious (approximating our EIP10), the median value (EIP50), or the time to maximum prevalence (approximating our EIP90). Our results clearly show potential for substantial variation between these measures, particularly at cooler temperatures. In turn, these discrete, single-value measures of EIP can yield markedly different estimates of transmission potential (such as rVC) for the same mosquito population (S2 Fig, S15 Table).
A number of recent studies describe parasite development rate (the reciprocal of the EIP) as a unimodal function, suggesting a decline in growth rate as temperatures increase above the optimum [4, 25, 66]. The unimodal functions result from inclusion of data (often single data points) at high temperatures in which parasites fail to complete development. Yet, there is a substantial mechanistic difference as to whether high temperatures limit transmission because parasite survival/vector competence declines or because parasite growth slows and EIP becomes progressively long (and is assumed to be infinite at the point where no parasites survive). We find no evidence of an increase in EIP at high temperatures. More data are needed to resolve this fundamental issue.
Furthermore, at high temperatures we see a decline in the prevalence of potentially infectious mosquitoes over time, suggesting either that mosquitoes are recovering from infection (i.e., sporozoites are dying or otherwise being cleared from the salivary glands and surrounding hemolymph), or that infectious mosquitoes exhibit differential mortality and die at a quicker rate compared with noninfectious mosquitoes. Our current experimental design does not enable us to determine the mechanism explicitly, but when we compare overall mosquito survival with the rate of decline in the proportion of infectious mosquitoes over time, we see no significant difference in instantaneous hazard rates (S16 Table, S4 and S5 Figs, see supplementary methods for analysis). This outcome is more consistent with enhanced death of infectious mosquitoes rather than parasite clearance. It has been suggested that P. falciparum has no lethal effect on naturally occurring mosquito hosts [67], yet most studies examine malaria infections in the range of 24°C to 28°C. Our data suggest that P. falciparum might impact mosquito survival at higher temperatures (see also [7]). We are not aware of any malaria transmission models that consider possible costs of parasite infection under increased environmental stress (temperature or otherwise).
Finally, we show that the predicted effects of temperature on overall transmission potential differ between a standard vectorial capacity model and an alternative model of force of infection, with further differences within the vectorial capacity model that are dependent on which estimate of EIP is used (See Fig 2, S15 Table). One reason for this difference is that the empirical life history parameters contradict several of the model assumptions implicit in the vectorial capacity approach. These differences between models could have important biological significance. For example, the differences in the upper and lower thermal limits for transmission would generate different patterns of range expansion and contraction in response to climate change. Within the transmission range, a shift in temperature from 24°C to 28°C (e.g., via seasonal change or longer-term climate warming) would be predicted to increase rVC by 34% but decrease relative force of infection by 1%. At the thermal extremes, even small shifts in temperature have quantitatively different outcomes; warming from 32°C to 34°C, for example, reduces rVC by 30% but reduces the force of infection model by 62%.
Our model of rVC derives from the standard formulation developed by Garret-Jones [53], which is itself a simplified version of the original dynamical model as developed by Macdonald [47]. Alternative formulations of vectorial capacity, such as presented in Brady et al. [68], could generate different thermal response curves for transmission as they combine individual life history traits in different ways. Extension of either the rVC model or the model of relative force of infection to a more holistic metric, such as the basic reproductive rate, requires additional information on actual mosquito density (also likely temperature dependent through effects of temperature on larval development rate and survival [66]), as well as susceptibility and rate of recovery of the human host. Regardless of model framework, properly characterizing the thermal performance curves for individual traits remains important, especially for key traits such as the proportion of mosquitoes surviving the EIP, or the frequency of blood feeding, as these are integral to transmission.
In general, we demonstrate that a detailed, system-specific examination of temperature sensitivity yields quantitatively and qualitatively different estimates of temperature-dependent life history traits, compared to the currently accepted relationships that integrate data from diverse studies and a mixture of mosquito and parasite species. For some traits, the differences appear small (e.g., our data on biting rate match previous data quite closely, at least over the temperature range of the current study). For other traits, the differences are substantial; our observed mortality rates are much higher, our parasite development rates are greater than those predicted by standard models at cool temperatures, and unlike contemporary unimodal thermal performance curves [4], we see no evidence for an increase in EIP at high temperatures.
We acknowledge that malaria transmission is not determined by temperature alone [55, 69–71]. Furthermore, we used long-standing lab colonies of a single mosquito–parasite combination; it is likely that parasite development, vector competence, biting rate, and longevity vary between species and between local populations [7, 57, 62, 72, 73], including the potential for local thermal adaptation [72, 74]. Yet, there is little reason to think that our system is more idiosyncratic than any other local malaria vector-parasite pairing in nature. We also focus on describing the effects of constant temperatures, as this is consistent with nearly all other studies to date. However, our own research has shown that daily variation in temperature can influence mosquito and parasite life history traits above and beyond the effects of mean temperature alone [2, 12, 66]. Future studies would benefit from the inclusion of daily temperature variation, particularly at high and low temperature extremes, as variation is likely to play an important role in defining the upper and lower limits of transmission. Inclusion of variation in biotic factors, such as differences in larval habitat quality, would also be valuable as these can further shape transmission potential [46, 75–77]. Such ecological complexities only strengthen the need for more detailed, system-specific studies of the type presented here in order to fully understand the influence of temperature on transmission and generate more informed predictions of the potential impact of climate change.
Temperatures were selected to capture key transmission range for P. falciparum [1, 4, 27, 28, 30]. All mosquitoes were housed in secure cardboard cups inside secondary mesh containment cages. Cages were kept in environmentally controlled reach-in incubators at 21°C, 24°C, 27°C, 30°C, 32°C, and 34°C, each ± 0.5°C and 80% ± 5% relative humidity. To ensure incubators were functioning correctly, we placed battery-powered portable USB dataloggers in each incubator. Daily data were extracted at approximately 9:00 am to ensure temperature and humidity were stable throughout the duration of the experiment.
In vitro cultures of P. falciparum strain NF54 (wild type, Center for Infectious Disease Research, Seattle, Washington) were maintained in RPMI 1640 medium (25 mM HEPES, 2 mM L-glutamine), supplemented with 50 μM hypoxanthine and 10% human A+ serum (Valley Biomedical, Winchester, Virginia). Culture was maintained in an atmosphere of 5% CO2, 5% O2, and 90% N2. Parasite cells were then subcultured into O+ human erythrocytes (Valley Biomedical). Gametocyte initiation occurred at 5% haematocrit and 0.8% to 1.0% mixed-stage parasitemia. The culture was maintained for 17 days and parasite cells were provided fresh media daily.
On the day of the blood meal, gametocyte cultures (approximately 8% gametocytemia for each experimental block) were briefly centrifuged, and the supernatant was removed and discarded. Pelleted erythrocytes were diluted to 40% haematocrit using fresh A+ human serum and O+ human erythrocytes. The mixture was pipetted into glass bell jars fixed with a Parafilm membrane and connected by plastic tubing with continuously flowing water heated to 37°C. Each bell jar was filled with 2 mL of blood culture, which feeds approximately 200 females. Mosquitoes were given 20 minutes to fully engorge, after which the bell jars were removed, as the parasites in culture are no longer viable after 20 minutes. In each cup, >95% of females were observed to have engorged fully. Immediately following the blood meal, mosquitoes were transferred to their respective temperature treatments and maintained by providing cotton balls soaked with 10% glucose and 0.05% para-aminobenzoic acid in water, which were replaced daily.
An. stephensi Liston adult females were from our laboratory colony (originally derived from a long-standing colony at the Walter Reed Army Institute of Research, Silver Spring, Maryland) maintained at standard insectary conditions (27°C ± 0.5°C, 80% ± 5% relative humidity and a 12:12 photoperiod). Three- to 5-day-old females were aspirated into cardboard cups (475 mL), with approximately 150 per cup. Four cups were allocated to each of the 6 incubators set at the different experimental temperatures (21°C, 24°C, 27°C, 30°C, 32°C, and 34°C), totalling approximately 600 females per temperature. Each cup was provided a human blood meal infected with in vitro cultured P. falciparum strain NF54 (wild type, Center for Infectious Disease Research).
Salivary gland sampling began on day 10 post-blood meal for 21°C and 24°C, day 8 for 27°C, day 6 for 30°C, and day 5 for 32°C and 34°C, based on results from a pilot experiment (S1 Fig). For each sample, 8 to 10 mosquitoes were aspirated from each replicate cup into absolute ethanol and salivary glands were dissected. Glands were ruptured beneath a glass cover slip and examined under a light microscope at 40x for presence of sporozoites.
Dead mosquitoes were counted daily in each cup. For survival analysis, mosquitoes removed for dissections each day and those that remained alive at the end of the experiment (day 25) were considered censored cases. We compared a range of plausible mortality curves including exponential, log-logistic, log-normal, Weibull, and Gompertz distributions for each block x temperature combination individually by using the R package flexsurv and selected the best-fit model using AIC (S7 and S8 Tables).
To estimate effects of temperature on the length of the gonotrophic cycle, 3- to 5-day-old females were fed to repletion on a membrane feeder using pork intestine sausage casing filled with human blood heated to 37°C. Fully engorged females (n = 85 per temperature treatment) were then transferred to individual 50 mL plastic tubes covered with mesh and filled with 5.0 mL of distilled water. Each tube was provided a cotton ball moistened with 10% glucose solution that was replenished daily. Daily, tubes were checked for presence of eggs between 9:00 am and 10:00 am. Females in tubes that had laid eggs were then transferred to a clean tube and sugar was withheld for 6 hours, after which these females were offered a second human blood meal on the membrane feeding system (in groups of 5 tubes per feeder, all feeds occurred at 27°C for optimum response).
This allowed for a quantification of the length of time to laying the first and second clutches; the mean of these values was used as gonotrophic cycle length. For females not laying a second clutch, the number of days to laying the first clutch was considered as the mean in the final calculation of mean cycle length. To calculate biting rate, we used the reciprocal of the mean gonotrophic cycle for each temperature. Differences in biting rate were assessed using a Kruskal-Wallis test followed by Dunn’s post-hoc rank sum comparison using the R package pgirmess.
To assess if temperature affected the likelihood of egg laying in general, mating success was also assessed by dissection of spermathecae from females in that had not laid eggs by day 21 post-blood meal. Spermathecae were ruptured under a glass cover slip and observed under a light microscope at 40x magnification. Presence of sperm, whether alive or dead, was considered a successful mating. Data on each individual clutch and mating success can be accessed in the supplementary materials (S12–S14 Tables).
Raw data for survival and infection, R script for statistical analysis, and numerical values for producing each figure can be accessed in the Dryad data repository: http://dx.doi.org/10.5061/dryad.74389 [34].
|
10.1371/journal.pntd.0002199 | A Qualitative Study Exploring Barriers Related to Use of Footwear in Rural Highland Ethiopia: Implications for Neglected Tropical Disease Control | The role of footwear in protection against a range of Neglected Tropical Diseases (NTDs) is gaining increasing attention. Better understanding of the behaviors that influence use of footwear will lead to improved ability to measure shoe use and will be important for those implementing footwear programs.
Using the PRECEDE-PROCEED model we assessed social, behavioral, environmental, educational and ecological needs influencing whether and when children wear shoes in a rural highland Ethiopian community endemic for podoconiosis. Information was gathered from 242 respondents using focus groups, semi-structured interviews and extended case studies. Shoe-wearing norms were said to be changing, with going barefoot increasingly seen as ‘shameful’. Shoes were thought to confer dignity as well as protection against injury and cold. However, many practical and social barriers prevented the desire to wear shoes from being translated into practice. Limited financial resources meant that people were neither able to purchase more than one pair of shoes to ensure their longevity nor afford shoes of the preferred quality. As a result of this limited access, shoes were typically preserved for special occasions and might not be provided for children until they reached a certain age. While some barriers (for example fit of shoe and fear of labeling through use of a certain type of shoe) may be applicable only to certain diseases, underlying structural level barriers related to poverty (for example price, quality, unsuitability for daily activities and low risk perception) are likely to be relevant to a range of NTDs.
Using well established conceptual models of health behavior adoption, we identified several barriers to shoe wearing that are amenable to intervention and which we anticipate will be of benefit to those considering NTD prevention through shoe distribution.
| Consistently wearing shoes may help in preventing onset or progression of a wide range of Neglected Tropical Diseases (NTDs). This study assessed the factors that influenced shoe wearing behaviors among people living in a rural community in highland Ethiopia. In this community, a substantial proportion of people are at risk for podoconiosis, a debilitating lower leg condition that can be prevented by wearing shoes. We conducted semi-structured individual interviews, focus group discussions and extended case studies among 242 adults and systematically analyzed the information. We found that shoe wearing is intermittent, and that different factors such as cost and ability to use the shoes for certain activities (such as farming) influenced consistent shoe wearing for most people. Some factors (such as shoe size, fear of stigma) were more relevant for podoconiosis patients. Social norms were found to be increasingly supportive of shoe wearing, and children exhibited greater desire to wear shoes than adults. These findings have relevance for preventing development and progression of a variety of NTDs in a range of settings.
| Interest is growing in the use of footwear in the primary prevention of certain Neglected Tropical Diseases (NTDs). While evidence for a protective role of footwear against podoconiosis [1]–[3] and chronic larva migrans [4], [5] is relatively strong, evidence for the role of shoes is inconsistent in relation to hookworm, with some studies finding evidence of protection from footwear [6], [7], but other studies finding no effect [8]–[11]. Evidence is also inconsistent for other helminthiases [10], [12]–[15] and Buruli ulcer [16], [17]. For snakebite and tungiasis, evidence of protection is circumstantial, and based on the predilection of bites [18] and lesions [19] for the feet.
While research on the impact of behaviors such as hand-washing [20], face-washing [21] and use of disease-preventing commodities such as insecticide-treated bed nets (ITNs, [22]) is relatively advanced, there is a paucity of research on behaviors related to footwear and their impact on NTDs.
While conducting work on the use of shoes in a rural Ethiopian community endemic for podoconiosis (a NTD triggered by exposure to irritant soils in the tropical highlands [3], [23]), we uncovered considerable information on behaviors and practices relating to shoe use which is relevant to a range of other NTDs. In brief, in southern Ethiopia, shoes are being distributed through a local non-governmental organization to children with the intention of preventing podoconiosis. This non-communicable form of elephantiasis arises from long-term exposure to red clay soils. Ecological and observational evidence suggests that consistent use of shoes prevents disease by protection from soil exposure. Shoe distribution to children of treated patients has been accompanied by messages linking foot hygiene and shoe use to reduced risk of disease. Program implementers considered it vital to understand why children might or might not wear shoes, in order to improve the messaging that might be used alongside distribution. To this end, we drew on several conceptual models to guide our efforts. First, we relied on the PRECEDE-PROCEED model that suggests beginning the process with diagnostic planning to assess social, behavioral, environmental, educational and ecological issues and needs that may influence whether and when children wear shoes [24]. We also considered social cognitive theory of self regulation [25]. Taken together these theories argue for the importance of targeting individuals' beliefs and attitudes about shoe wearing, how these beliefs influence perceived capabilities to prevent podoconiosis, and whether wearing shoes can be effective in reducing their risk for the condition. The data presented in this article arise from a qualitative study aimed to gain deeper understanding of the barriers to consistent use of shoes in a rural setting.
We anticipate that this information will be valuable both for investigators designing future studies to assess the association between shoe use and incidence of NTDs, and for those developing shoe-related prevention programs.
Ethical approval was granted by the Institutional Review Boards of Addis Ababa University Medical Faculty and the National Human Genome Research Institute, National Institutes of Health, USA. Oral consent was obtained from all study participants by a trained research assistant, following the procedures developed and evaluated by Tekola and colleagues using Rapid Ethical Assessment in this community [26]. In brief, Rapid Ethical Assessment is a form of rapid anthropological assessment performed to explore a community's understanding of research, to document how a community prefers to be approached by investigators and to detail how community members wish consent to be given. Rural communities in Wolaita prefer contact through a Mossy Foot Treatment and Prevention Association (MFTPA) staff member prior to individual discussion and consent. The Mossy Foot Treatment and Prevention Association is a local non-governmental organization involved in the prevention and treatment of podoconiosis patients and shoe distribution for their children. Oral consent is preferred by this community [26], was approved by the IRBs mentioned above and was documented by a witness on each occasion following an explanation of the research protocol using the information sheet and consent form.
The study was conducted in Wolaita zone in southern Ethiopia, where the population is estimated to be 1.7 million [27]. Podoconiosis is known to be prevalent in this zone [28]. Most of the villagers are subsistence farmers. The study was entirely qualitative and employed multiple methods (focus group discussion (FGD), in-depth interviews (IDI) and case studies) to gain an in-depth understanding of community perspectives on behaviors related to shoe use, and the predominant facilitators and barriers to wearing shoes. Structured topic guides were used to direct discussions, focusing on local explanations for the causes of podoconiosis, attitudes towards individuals affected by podoconiosis, attitudes towards wearing shoes, and optimal role models and settings for promoting footwear among high-risk children. Case studies enabled deeper and more contextualized information to be gathered around an individual, with information gathered from the individual, from family members and from friends. An article describing community perceptions surrounding risk factors and prevention, including how adults' explanations of disease heredity influence shoe wearing and interpersonal behaviors, has recently been published [29]. The present paper focuses on the perceptions of participants regarding footwear and explores the major factors impeding shoe use in the community.
Participants were recruited using convenience and snowball sampling methods. A total of 242 adults participated from the following three groups: (1) 69 adults affected with and receiving treatment for podoconiosis, “affected”; (2) 129 unaffected adults, with no current sign of or previous history of podoconiosis, “unaffected”; and (3) 44 community and religious leaders “community and religious leaders”. None of the community leaders and religious leaders was currently affected by disease. The study took place in four of 14 communities served by the Mossy Foot Treatment and Prevention Association (MFTPA) a local non-governmental organization involved in the prevention and treatment of podoconiosis patients and shoe distribution for their children. The four sites were selected to represent the diversity of communities served with respect to size, duration of the relationship with MFTPA, and distance from the main office of the MFTPA.
This study was conducted from June to August 2010. The month of June was partly dry while the rest of the study was conducted in the rainy season. This allowed the researchers to observe community shoe wearing practices during both the dry and the rainy seasons. A trained research assistant (Desta Ayode - DA) spent up to three weeks in each of the four communities with Abebayehu Tora (AT) and one other data collector conducting focus group discussions, semi-structured in-depth interviews and extended case studies with research participants. A total of 38 IDIs, 28 FGDs and 7 case studies were conducted in the study sites. All materials used for the study were developed in English, and then translated into Amharic and Wolatigna. The discussion and interviews were conducted in either Amharic or Wolatigna, and were audio-recorded. The audio-recordings were first transcribed in the language in which they were conducted (either in Amharic or Wolatigna), then translated into English. Translations of both study materials and transcripts were checked for consistency and to evaluate accuracy of important concepts.
A total of four coders with different backgrounds (Hendrik de Heer – the Netherlands, Emi Watanabe - Japan, Desta Ayode – Ethiopian resident and Tsega Gebreyesus – Ethiopian diaspora, the latter two Amharic speakers) were involved in developing the coding scheme and coding the data in order to maximize the breadth and depth of the analysis. After initial reading of the transcripts, the interview themes served as a starting point for the codebook, and subthemes were created as they emerged from the data. These overarching themes included barriers and advantages to wearing shoes, beliefs about podoconiosis and perspectives on best settings for interventions to facilitate shoe wearing as a means of prevention of podoconiosis and other diseases. In weekly meetings, any suggested categories or themes to add were discussed and agreed upon by all coders before being added to the list of themes and sub-themes. All coders coded multiple data sources and overlapped with each of the three other coders. Every inconsistency between coders for a given source (e.g. the transcript of a focus group) was resolved through discussion. The first 10% of all transcripts was coded by all four coders and two-thirds of all transcripts were coded by at least two coders. NVIVO-9 Qualitative data analysis software was used to assess all themes in the transcripts (NVIVO, QSR International, Burlington, MA 01803, USA). For example, the major sub-themes that emerged for barriers to consistent shoe-wearing included: i) financial barriers, ii) unsuitability of available shoes for certain activities, iii) low perceptions of adverse consequences as a result of not wearing shoes, iv) difficulty finding appropriate shoe sizes and v) fear of stigmatization as a result of wearing certain shoes. These themes are discussed in greater detail in the results section.
Results are presented in two main categories – barriers to consistent use of shoes, and community perceptions favoring footwear. Barriers to consistent use of shoes are further divided into four categories including those related to: 1) limited financial resources; 2) the unsuitability of shoes for specific activities; 3) a low perception of risk; 4) a fear of stigma. Quotes are attributed according to the age, gender and disease status of the participant.
Although many respondents aspired to wear shoes, the reality was rather different. Many adults possessed shoes, and most parents stated that they were trying to buy shoes for their children; however, respondents observed that shoes were not worn regularly by most members of their community. As stated above, barriers faced by all community members included financial constraints, poor access to footwear appropriate to a range of local activities, and low perceptions of disease risk. Podoconiosis patients faced two additional barriers: difficulty finding large enough sizes, and fear of stigma and labeling.
In terms of wearing shoes, it appears that in Wolaita zone, like other rural parts of Ethiopia, people are moving from a ‘norm’ of going barefoot, to one where shoes are worn, and it is becoming ‘shameful’ to appear in public places without wearing shoes. Expansion of schools in rural communities and proliferation of the variety of shoes in local markets have contributed enormously to changing mindsets towards accepting footwear as a valuable commodity:
The overwhelming majority of respondents were positive about wearing shoes. Both adults and children emphasized that, despite the impediments to securing footwear, everyone in the community was in favor of having shoes:
The following excerpts demonstrate that social pressures (and not just issues related to disease prevention) are important in driving the community norm towards wearing shoes:
Even young children communicate their wish for shoes to parents: attempting to wear their parents' shoes, nagging their parents to buy them shoes, and refusing to attend school barefoot. In many families, it is the children who press their parents into buying their first shoes. As one parent said,
The following excerpts also illustrate this very well:
In this specific area, although the MFTPA has worked to circulate messages about podoconiosis prevention for more than ten years, this does not appear to be an important reason for wearing shoes in the wider community. Adults emphasized using shoes to participate in social settings and public gatherings, while children emphasized the protective value of shoes against pain of walking on stones and other sharp objects. Respondents in many groups also mentioned that shoes protected them from cold and injuries, enabled walking and looked attractive on the feet. In some cases, shoes were worn simply because they saw others wearing them.
Since podoconiosis patients had been advised to wear shoes by the MFTPA, we found that the practices of treated patients differed from those of the general community:
Among patients, primary prevention may beneficially be linked with disease treatment, and patients using shoes for secondary prevention of complications may not only model behavior changes but also encourage them in children.
If I leave home without shoes, I immediately get sick. I can't step even a short distance without shoes. So, shoes are important to protect us from the painful feeling. (Affected female, age 28)
This study, which aimed to explore shoe wearing practices in a podoconiosis-endemic setting in rural Ethiopia, brings to light several issues relevant to other foot-related NTDs. We discovered that, despite a clear wish to wear shoes, and to wear them regularly, many practical and social barriers prevent these wishes being translated into practice. Many of the barriers cited will be relevant to those considering distribution of shoes to prevent snakebite, tetanus or helminthiases. We also witnessed inconsistency between reported and actual shoe wearing behavior, confirming the complexities that exist in relation to recording shoe use. We suspect that these complexities may not have been adequately addressed in earlier studies on risk factors for a range of NTDs.
Shoe wearing was intermittent, with adults more likely to say they wore shoes for social events and gatherings including market attendance, church services, weddings and funerals. Farmers, both male and female, rarely wore shoes while working in the fields, and many householders did not wear them while gathering wood or fetching water. Although children were usually encouraged to wear shoes at school, they were often dissuaded, sometimes forcibly, from wearing them for housework or play.
More consistent use of shoes was reported by podoconiosis patients than the wider community, several patients referring to advice received from the MFTPA. Perception of risk appeared to be an important contributor to this difference in behavior: patients reported changing their own shoe wearing behavior and influencing that of their children, while non-affected community members wore shoes less or not at all. Several articles have linked risk perception with actions related to health-seeking behavior, people with higher perceived vulnerability to illness being more likely to engage in protective behavior [30]. Research on foot care and footwear practices of peoples with diabetes [31], [32] has demonstrated similar links between use of shoes and perceived risk of disease to those presented here. This suggests that any future NTD interventions based on shoe distribution to individuals with disease must be accompanied by messages that appropriately convey mechanisms by which diseases occur and individual and community levels of risk.
While patients viewed shoes as a means of protection from disease, non-affected adults indicated they were beginning to have more general social value. Shoe wearing was seen as a mark of dignity, while going barefoot was seen as ‘shameful’, particularly by the younger generations. Some participants suggested that shoe wearing norms were in the process of change, and one directly ascribed this to education - “the advancement of education has changed the minds of the people… today”. Clearly, drivers of change in this norm are acting at many levels, and though some may be harnessed in intervention programs, others will be beyond easy reach. Children are also aware of the ‘shame’ of going to school barefoot, but also mentioned the role of shoes in preventing injuries from stones, thorns and other sharp objects. Future programs will need to address all these motivations for shoe use and highlight the range of benefits that shoe wearing is likely to bring.
Recurring barriers mentioned by study participants that are likely to be relevant in other NTD-endemic communities, were those of financial constraint and poor suitability of shoes for the most common activities. Financial constraints were reported to influence possession of shoes, type of shoe bought, age at which a child starts wearing, which children get shoes within families, consistency of use, frequency of replacement and activities for which they were worn. As with many health interventions whose benefits will only become apparent in the longer term, families naturally prioritized more immediate concerns such as food. Currently, in this area, shoes are being distributed free of charge, but this is unlikely to be sustainable in the long term or scalable to all rural populations exposed to NTDs. However, if shoes are to be considered health interventions rather than pure commodities, subsidies or micro-credit strategies that bring shoes within the reach of very-low income families must be contemplated. Social protection strategies like these may bring the necessary empowerment for individuals to realize the behavior changes they may desire to make.
Several participants gave highly practical reasons for preferring not to wear shoes while farming, saying that the shoes available in the market quickly became heavy with mud and failed to grip in the rainy season, and became uncomfortable when rough soil particles slipped inside during the dry season. Long Wellington-type boots might prevent these problems, but are more expensive than the shoes currently available. Clearly, promoting footwear that is appropriate to local activities and effective against the specific NTD is essential. For example, the prevention of snakebites in rice paddies will require different footwear than those required for the prevention of chronic larva migrans on the beach.
Some barriers to use of footwear were patient-specific. Swelling of the feet and lower limbs, nodules and wounds may make use of normal-shaped shoes impossible. Molla and colleagues [33] have documented similar challenges faced by podoconiosis patients in northern Ethiopia. Custom-made shoes might overcome these difficulties, and have been developed for patients with leprosy in similar resource-limited rural communities. Legs to Stand On, an initiative to prevent disabling disease of the lower limb in resource-poor settings, is leading cross-disease efforts to increase capacity to manufacture custom-made shoes in these communities.
However, custom-made shoes bring with them the possibility of stigma through labeling as ‘diseased’. This was raised as a very real barrier by a number of patient participants in our study. Some had developed tactics to mitigate the stigma they faced, by removing their shoes in certain situations, while others had abandoned them completely. Stigma has been documented, against podoconiosis patients and their families [26], [34] against patients with leprosy [35] and lymphatic filariasis [36], and interventions against these or other NTDs must not risk increasing stigma. In the future, much more attention must be directed to the design of custom-made shoes so that they do not increase stigma in relation to any NTD.
Several investigators have suggested that lack of association between footwear and disease in observational studies reflects poorly refined measurement of shoe wearing behavior. In most studies, there is no clear definition of the length of time spent wearing shoes or the activities for which they are worn. For example, while individuals may state they wear shoes ‘most of the time’, they may remove them to plough, sow, harvest or fish. These activities may represent the time of greatest exposure to infective or other agents. Many quantitative studies investigating the link between incidence or prevalence of NTDs and footwear have used simple questions such as ‘Do you wear shoes?’ with binary response options [8], [11], [13]–[15]. Our participants describe complex behaviors, wearing shoes in certain settings (including in church services and at school) but not in others (often those where exposure is more likely, such as farming). Clearly, more nuanced questions must be asked of study participants if a true picture of shoe wearing is to be captured. Better designed questions, informed by qualitative research, will allow identification of potential points of behavioral intervention that take into account the structural barriers posed by rural poverty.
We have explored behaviors related to use of shoes in a low-income rural setting where several NTDs including podoconiosis are prevalent [37]. We used a range of qualitative techniques among multiple target groups. Although the study included a large sample, all respondents were drawn from the same rural highland community in Ethiopia, and so we suggest caution in generalizing the reported outcomes to other cultural settings. Although we hoped to reduce social desirability bias by collecting data through individuals not linked to the MFTPA organization, it is likely that the information given by some respondents was still influenced by their wishes for perceived social conformity.
Although shoes are desired, they are either not worn or not worn sufficiently consistently to prevent disease. Consistent with well established conceptual models of health behavior adoption, we identified several barriers to shoe wearing that are amenable to intervention [25]. Moreover, several of these barriers will arise in other settings in relation to other NTDs, and we encourage program developers to consider each of these before developing theory-based interventions to encourage shoe wearing.
|
10.1371/journal.pcbi.1005159 | Long-Range Signaling in MutS and MSH Homologs via Switching of Dynamic Communication Pathways | Allostery is conformation regulation by propagating a signal from one site to another distal site. This study focuses on the long-range communication in DNA mismatch repair proteins MutS and its homologs where intramolecular signaling has to travel over 70 Å to couple lesion detection to ATPase activity and eventual downstream repair. Using dynamic network analysis based on extensive molecular dynamics simulations, multiple preserved communication pathways were identified that would allow such long-range signaling. The pathways appear to depend on the nucleotides bound to the ATPase domain as well as the type of DNA substrate consistent with previously proposed functional cycles of mismatch recognition and repair initiation by MutS and homologs. A mechanism is proposed where pathways are switched without major conformational rearrangements allowing for efficient long-range signaling and allostery.
| We are proposing a new model for how long-range allosteric communication may be accomplished via switching of pre-existing pathway as a result of only minor structural perturbations. The systems studied here are the bacterial mismatch repair enzyme MutS and its eukaryotic homologs where we identified strong communication pathways connecting distant functional domains. The functionally-related exchange of nucleotides in a distant ATPase domain appears to be able to switch between those pathways providing a new paradigm for how long-range allostery may be accomplished in large biomolecular assemblies.
| Allostery is a fundamental part of many if not most biological processes. It is classically defined as the induced regulation at one site by an event at another distal site. Venerable models for allostery, such as the MWC (Monod-Wyman-Changeux) [1] and KNF (Koshland-Nemethy-Filmer) [2] models emphasize a mostly static picture of induced conformational changes. The MWC model proposes coupled conformational changes via a population shift while the KNF model highlights the induced-fit of a binding of a ligand via common communication routes. A broader view of allostery [3–6] emphasizes communication pathways via protein motions but without requiring actual conformational changes. The idea of this model is that relatively minor perturbations may shift communication between multiple pre-existing pathways. Such a mechanism has been demonstrated by nuclear magnetic resonance (NMR) experiments for the binding of cyclic-adenosine monophosphate (cAMP) to the dimeric catabolite activator protein (CAP) [7] as well as for allosteric regulation in Pin1[8]. Recent work based on Markov state models that integrate energetics and kinetics has added further nuances to the discussion by emphasizing both conformational and kinetic selection as the main mechanism of allostery in signaling proteins protein kinase A [9] and NtrC [10]. The idea of kinetic selection is consistent with a pathway selection mechanism without significant conformational changes. Recent reviews have attempted to integrate the different ideas into a unified view [11, 12] with the main question being to what degree conformational dynamics plays a role. Likely, the degree of dynamics will depend on a given system and the economics of achieving allosteric signaling within the thermodynamic and functional constraints in the biological environment. One particular question that is central to this work is how long-range allostery can be achieved in very large systems where larger conformational changes and global selection mechanisms that are conceptually straightforward in smaller proteins could be more challenging to realize.
It is difficult to obtain detailed insight into allostery from experiments, especially for larger and more complex systems, because NMR spectroscopy is generally limited to small and soluble proteins that can be easily labeled and expressed in large quantities. On the other hand, crystallography is not well-suited for studying allosteric effects due to its inherent dynamic nature.
Computational approaches such as statistical coupling analysis (SCA) [13], normal mode analysis (NMA) [14, 15], dynamical network analysis [16], and Markov state model analysis based on extensive molecular dynamics simulations [9, 10] offer complementary means for exploring allosteric mechanisms in biological systems. SCA, a bioinformatics-based method, obtains allosteric information by identifying coevolving residues from multiple sequence alignments, while NMA, a structure-based approach, suggests induced movements from a few robust low-frequency normal modes. Allosteric pathways obtained from these two methods would be encoded in the sequence and/or structure, but sensitivity to minor perturbations with this type of analysis is lacking. Dynamical network analysis [16] is based on molecular dynamics (MD) simulations and has been used to identify synchronous and/or asynchronous correlated residue motions in order to describe possible allosteric communication pathways. Examples of where this approach has been applied successfully to probe allosteric coupling include a tRNA-protein complex [16], the M2 muscarinic receptor [17], and cysteinyl tRNA synthetase [18]. Here, we used dynamical network analysis to develop a paradigm for allostery in very large multi-subunit complexes based on long-range signal propagating pathways in the MutS component of the methyl-directed DNA mismatch repair (MMR) system.
MMR is responsible for correcting errors that escape immediate proofreading during DNA replication and the mechanism is widely conserved from prokaryotic to eukaryotic organisms. MMR alone can increase the accuracy of DNA replication by 20–400 fold [19]. While several components, such as MutS, MutL, MutH, nuclease, and polymerase, are needed to work together to complete DNA repair [20], MutS is responsible for the initial recognition of DNA lesions, in particular mismatches and insertions or deletions (IDLs). MutS is a homodimer, but, structurally and functionally, it acts as a heterodimer because only one subunit (termed the ‘A’ chain in this paper) directly contacts the lesion sites [21]. MutS homologs (MSH) in eukaryotes are heterodimers with differing substrate specificities. MutSα (MSH2-MSH6) preferentially recognizes base pair mismatches and single base IDLs [22], whereas MutSβ (MSH2-MSH3) has a higher affinity and specificity for small DNA loops composed of 2–13 bases [23].
The crystal structures of prokaryotic MutS and its eukaryotic homologs, complexed with mismatched DNA heteroduplexes, feature a similar overall Θ shape [22, 24–26]. Each subunit of MutS and MSH is comprised of five distinct domains (see Fig 1): the mismatch-binding domain (MBD, domain I), the connector domain (domain II), the lever domain (domain III), the clamp domain (domain IV), and the nucleotide binding domain (ATPase, domain V) [24]. The MBD and clamp domains interact with the bound DNA directly. The MBD contains a conserved, mismatch-identifying Phe-X-Glu motif, forming specific interactions with mismatches. The phenylalanine forms an aromatic ring stack on the 3′ side of the mismatched base [24, 25] although there is also evidence for base flipping of the mismatched or neighboring base during the mismatch recognition process [27, 28]. MSHβ, which specializes in the recognition of longer insertions/deletions, lacks this motif. The lever and connector domains connect the MBD and clamp domains to the ATPase domain. The ATPase domain is a conserved domain in the ABC (ATP binding cassette) superfamily. Biochemical studies have provided evidence that ATPase activities are coupled with DNA scanning, mismatch recognition, and repair initiation [29–31]. The different functional states are assumed to involve different conformations of MutS. The major states are MutS without DNA with open clamps, MutS scanning DNA in search of a mismatch with the clamps closed, MutS bound to a mismatch in the tightly DNA-bound conformation seen in crystallography, and a sliding clamp configuration where MutS is able to move away from the mismatch without scanning or complete dissociation from the DNA [32]. Based on the biochemical data, nucleotide binding and exchange to the ATPase domain appear to be key allosteric effectors coupled to DNA mismatch recognition that at least in part trigger changes between those functional states. This implies allosteric coupling between the mismatch binding site and the ATPase site over a distance of 70 Å [33] is essential for the biological function of MutS. Mismatched binding promotes exchange from ADP to ATP based on kinetic measurements of ATP hydrolysis [30, 34, 35] and results in asymmetric activity of the two ATPase domains [30], whereas the sliding clamp state is supposed to be formed in ATP binding states [36–38].
Previous studies have examined MutS and eukaryotic analogs via molecular dynamics simulations [27, 32, 39–43], but many mechanistic questions remain. Here, we subjected previously generated simulations of MutS, MutSα (MSH2-MSH6) and MutSβ (MSH2-MSH3) to dynamical network analysis to elucidate allosteric communication pathways between the structural domains in MutS and MSHs. In particular, we addressed the questions of how intra-molecular signaling could be accomplished over very long distances via protein dynamics and how small perturbations could affect the signal propagation. Previous work has suggested coupling between the MBD and ATPase domains, but mechanistic details and in particular the role of exchanging NTPs still remain largely unclear [43, 44]. The dynamic network analysis applied here allowed us to probe for pathways connecting the domains in contact with the DNA and the ATPase domain. Furthermore, by comparing pathways in simulations with different nucleotides bound to MutS and different DNA substrates bound to MutSα and MutSβ we were able to develop hypotheses for how communication along those pathways may be shifted during the functional cycle of MutS and its homologs.
A number of very similar MutS and MSH crystal structures are available with different nucleotide bound states and different mismatches or IDLs. The structural variations that can be discerned primarily focus on the MBD, ATPase and clamp domains and involve mostly local side-chain displacements rather than larger conformational changes of the main chain. For example, the MutS crystal structures 1E3M (with a single ADP) [24] and 1W7A (with bound ATP) [33] differ by only 0.35 Å in the Cα coordinates after superposition. MD simulations paint a similar picture. In previous work from our group, MD simulations of MutS with all possible combinations of nucleotides bound to the ATPase dimer did not reveal large conformational changes of the overall MutS structure based on RMSD and clustering analysis [27]. A similar conclusion was found for Thermus aquaticus MutS in a recent study [44], although different nucleotides bound to the ATPase domain were not examined. Taken together, this information has suggested that allosteric communication in this system likely takes place via subtle changes in local dynamics to achieve signaling in MutS rather than via conformational selection or induced conformational changes [3, 44].
Average dynamical cross-correlation matrices (DCCM) were calculated from the MD simulations. Fig 1 compares the DCCMs between MutS simulations with different nucleotides. A comparison of the DCCMs after 50, 100, 150, and 200 ns generally shows little change after 50 ns (S1 Fig). This suggests that the correlations based on 200 ns trajectories are well converged consistent with a previous study [45]. In all cases, we found strong local correlation within domains but also weaker coupling between distant parts of the complex (Fig 1). Overall, different nucleotide bound states resulted in similar coupling patterns, but differences as a function of different nucleotide bound states can be discerned. For example, the positive MBD(I)-connector(II)A coupling is strongest in ADP-None, while the strongest positive MBD(I)-connector(II)B coupling is observed in None-ADP. Also in the case of ADP-None, the MBD and connector domains of subunit A are strongly negatively coupled with the lever and clamp domains of subunit B. The two clamp domains are positively coupled in cases of ATP-ADP and None-None, which are stronger than the others. The positive coupling between the two ATPase domains is strongest in ATP-ADP. Similar direct correlations between MutS domains have also been observed in other work based on MD simulations of Thermus aquaticus MutS [44]. However, while a direct correlation analysis suggests coupling, it does not provide complete information about the pathway(s) along which allosteric communication take place and it discounts the possibility of asynchronous communication via stochastic steps that would introduce a variable time delay between signal input and output along a given communication pathway.
Next, we turned to dynamical network analysis to allow for a more dynamic model of allostery where direct correlations between distant sites are not required. In this approach, pathways connecting residue pairs along the shortest path with the highest pairwise local correlations based on the converged DCCMs from 200 ns MD sampling are determined. We focused our analysis on the functionally most relevant signal propagation between the MBD, ATPase, and clamp domains using specific key residues as anchor points (S2 Table). A first set of pathways was determined between MutS-F36, the key residue in direct contact at the mismatch site, and MutS-K620, the key residue involved in binding the phosphate tails of NTPs in the ATPase domain. A second set of pathways was focused on the communication between the two ATPase domains connecting MutS-K620 in the A and B chains and a third set of pathways was constructed from MutS-K620 to MutS-N497, which is the contact point of the clamp domain with the DNA opposite the mismatch site in the B subunit of MutS.
Mapping of the resulting pathways onto the MutS structure is shown in Fig 2. The computational analysis suggests multiple major pathways that vary as a function of the nucleotides bound to the ATPase domain. Within each major pathway, there are ensembles of similarly optimal minor pathways. The variability in the pathways was greatest within a given structural domain, where strong coupling between many residues allowed for many alternate, equivalent routes. However, connections between domains were limited to certain key residue pairs (S3 Table) that presented bottlenecks in the respective pathways. When employing network analysis to group strongly coupled residues into communities (S2 Fig), these communication bottlenecks appear as critical inter-community edges that are hypothesized to correspond to switching points between major pathways when perturbed.
Tables 1–3 quantify the features of the optimal pathways in terms of the number of steps (hops) required to traverse a path from the beginning to the end, a weight reflecting the degree of correlation along the optimal path, and the minimum pairwise correlation for any residue pair along the path. This analysis was carried out for each of the three sets of pathways as a function of different nucleotides bound in the ATPase domains. The algorithm employed here is designed to always find an optimal path connecting two given residues. In order to identify paths that are functionally relevant we focused on paths that stand out by having significantly lower weights than other paths while also requiring that the minimum correlation along the path was at least 0.7. Our assumption is that even if a path has an overall low weight, it would not be an effective route of communication if it contained one or more links with poorly coupled residues.
The overall premise of this study is the development of a dynamic allostery model for MutS since structural and previous simulation data suggest little conformational change as a function of nucleotides bound to the ATPase domains. Such a model implies the presence of communication paths between key structural elements (MBD, ATPase, and clamp domains) and the main result of this work is the identification of such paths in a nucleotide-dependent manner. Integrating previous biochemical data with such a dynamic allosteric model allowed us to arrive at the mechanism depicted in Fig 7 and described in more detail in the following:
In the absence of DNA, the ADP-ADP state is presumed to be dominant [54]. The ADP-ADP state dissociates directly from DNA, while the binding of DNA induces the dissociation of one ADP molecule, more likely to be the one in B subunit. Therefore, the ADP-None state is presumed to be the mismatch scanning state [34, 55]. Crystal structures of MutS were also obtained mostly in the ADP-None state [56–59]. Our results also support this idea. The MBDA strongly couples with the connector and ATPase domains in the ADP-None state via the proposed pathway ② that consists of a broad ensemble of individual paths with a few bottlenecks at domain boundaries. This communication is then proposed to result in exchange of ADP for ATP in the ATPaseA domain upon mismatch recognition [34, 35]. In our model, the presence of ATP in the A site would abolish the communication between the MBD and ATPaseA domains because an optimal or suboptimal path via the connector domain is either absent altogether (ATP-ADP) or present with less favorable weights or low minimum pairwise correlations (ATP-none, ATP-ATP in path ①) that suggest inefficient coupling. At the same time, the ATPaseA -ATPaseB communication would engage the Walker B motif of chain B when ATP is present in the A site. We further hypothesize that ADP or ATP binding to the B site would follow, leading to the ATP-ADP state. Since the lifetime of the ATP-None state is believed to be short [54] this would occur quickly. Once the ATP-ADP state is reached, our model suggests that the ATPaseA domain connects to the lever instead of the connector. Because the connector primarily connects with the MBD domain while the lever domain provides a route to the clamp domain, we hypothesize that in the ATP-ADP state communication from the ATPaseA domain would be switched from the MBD domain to the clamp. At the same time, strong coupling between the ATPaseA -ATPaseB domains via the signature loop could disrupt the strong ATPaseB-clamp connection present in the ADP-None state and allow release of MutS from the mismatch site.
The mechanism above postulates communication routes within the context of a dynamic allosteric mechanism that could be tested further experimentally, e.g. via mutations of pathway residues. Based on the dynamics sampled in the underlying simulations we are able to propose a structural basis for how pathways are switched in the presence of different nucleotides, however, the model is still lacking a clear mechanism for how ADP would be exchanged for ATP following mismatch recognition, for how the clamp domains would respond to signaling resulting from nucleotide exchange as proposed here, and what role MutL binding plays in this process. We speculate that the altered correlated dynamics induces subtle shifts in the overall conformational landscape which would favor ADP-ATP exchange and lead to clamp opening. To address this idea in more detail, significant additional simulations are required to probe the DNA binding process and clamp dynamics leading to the sliding clamp conformation in excess of the scope of the present work. Such a model could also conceptually integrate recent conformational landscape-based ideas of allostery with the communication-focused analysis presented here into a complete model for allostery in a large, complex system such as MutS where simpler concepts of conformational selection or induced-fit may not be able to adequately describe the allosteric mechanism.
MutS can recognize a broad range of lesions, mismatches and IDLs, but MSHs have differentiated substrate specificities. MutSα (MSH2-MSH6) primarily recognizes mismatches and single base IDLs, whereas MutSβ (MSH2-MSH3) recognizes DNA loops composed of 2–13 bases. Based on previous simulations of MutSα and MutSβ with native and swapped substrates and no DNA at all [60], we also analyzed how different DNA substrates would shift the signaling pathways identified via our computational analysis.
In the MSH complexes we identified pathways analogous to paths ①, ②, and ③ in MutS (see Fig 8 and details in S8 Fig and S9 Fig) suggesting that the proposed communication pathways may be preserved in the eukaryotic homologs. There appears to be strong communication from the MBD through the connector domain when MutSα and MutSβ are bound to their native substrates (a G:T mismatch and a four-nucleotide insertion loop (IDL-4L), respectively). However, swapping the substrate would abolish that path in favor of coupling along the lever domain. Again, cancer-associated mutations in MSH6 and MSH2 map onto the paths, some at critical edges connecting different domains (see S7 Fig and S8 Fig). Interestingly, communication between the MBD and ATPase domain of MSH3/MSH6 through the connector would also be present in the absence of DNA. These findings expand our allosteric model where effective communication between the MBD and ATPaseA domains (and subsequent initiation of repair) would depend on both the nature of the DNA substrate and the nucleotides bound in the ATPase domains.
Long-range signaling and allostery is a key mechanistic component of many large biomolecular complexes. Here, we present a detailed analysis of E.coli MutS and MSHs where several long-range signaling steps are essential for initiating DNA repair following mismatch recognition. Using dynamic network analysis based on extensive molecular dynamics simulations we developed a model consisting of a number of communication pathways that depend on strong local pairwise residue dynamical coupling where signaling would be expected to progress stochastically along those paths. In this model, different combinations of ATPase-bound nucleotides would result in switching between different pathways to implement the functional cycle of MutS without significant conformational rearrangements. A signaling mechanism based on pre-existing pathways that are switched on or off by different nucleotides and/or different DNA substrates is consistent with previous crystallographic and simulation studies that show surprisingly little structural variations in mismatch-bound MutS and homologs. The benefit of such a mechanism could be energetic economy, especially when considering the very long range over which the pathways appear to operate. Experimental validation of the hypotheses presented here could involve mutations of key residues, but it will also be interesting to see whether similar mechanisms are at play in other large enzymes. However, further computational studies will also be necessary to develop a more complete mechanistic understanding of how exactly signaling along the proposed pathways would promote and depend on nucleotide exchange and how it would lead to sliding clamp formation and complex formation with MutL.
MD simulations of the E.coli MutS protein bound to a G:T mismatch DNA (PDB ID: 1W7A) [33] were previously performed [27]. Each ATPase site may have three states: ATP, ADP or no nucleotide. All combinations of the three states in either of the two ATPase domains were simulated. They are denoted as ATP-None, None-ATP, ATP-ATP, ADP-None, None-ADP, ADP-ADP, ATP-ADP, ADP-ATP and None-None (S1 Table). In this notation, the first nucleotide is present in the ATPase site of the mismatch-binding moiety (subunit ‘A’) and the second one in the ATPase site of the non-mismatch-binding moiety (subunit ‘B’). Additional new simulations were carried out for five mutants of the E. coli MutS system to test the mechanistic hypotheses developed in this study: E169P (ADP-None), L240D (ADP-None), and Q626A (ATP-ADP) in chain A as well as L558R in either chain A or B (ATP-ADP). These simulations were simulated using the same protocol as the previous simulations of the wild-type systems (see below).
MD simulations of human MutSα and MutSβ were started from the crystal structure 2O8B [22] and 3THX [26] (MutSα/G:T and MutSβ/IDL-4L) [61]. In MutSα and MutSβ structures, MSH6 and MSH3 are the mismatch-bound moieties (equivalent to the A subunit in MutS), while MSH2 interacts with the DNA non-specifically (equivalent to subunit B in MutS). Additional simulations were carried out for apo structures, where the DNA heteroduplex was removed (MutSα/Apo and MutSβ/Apo), and for MutSα and MutSβ where the respective substrates were swapped (MutSα/IDL-4L and MutSβ/G:T) [61].
In total, 15 previous simulations and five new simulations (S1 Table), each for at least 200 ns, were analyzed. All of the simulations were carried out with NAMD 2.8 [62] using the CHARMM27 force field [63], the latest force field available at the time those simulations were initiated. All systems were solvated in explicit solvent using the TIP3P water model and sodium counterions to neutralize the systems. Simulations were carried out under periodic boundary conditions with the particle-mesh Ewald method [64] to calculate electrostatic interactions at constant temperature (300K) and constant pressure (1 atm) using a Langevin thermostat and barostat. The fully solvated systems consisted of about 165,000 atoms for the MutS systems and about 600,000 atoms for the larger MutSα and MutSβ systems. All of the systems remained overall stable with RMSD values of 3–5 Å for Cα atoms with respect to the initial experimental structures. Additional details of the system setup and simulation results are described in our previous papers [27, 61]. VMD [65] was used to visualize and analyze simulations and generate structural figures.
Allosteric networks within the proteins were identified using the NetworkView plugin of VMD [16, 66]. The dynamic networks were constructed using data from our molecular dynamics simulations of the protein-DNA complexes described above, each sampled every 1 ps. For each molecular system, a network graph was constructed with two nodes for each nucleotide (at N1/N9 and Pα/P), while protein residues were represented with a single node at the Cα position. All of the conformations from a given trajectory were pooled to calculate the local-contact matrix. A contact between two nodes (excluding neighboring nodes) was defined as within a distance of 4.5 Å for more than 75% of MD trajectories. The resulting contact matrix was then weighed by the correlation values of the two end nodes in the dynamical network as wij = −log(|Cij|), where Cij are the elements of the correlation matrix calculated as Cij=⟨Δri⋅Δrj⟩/⟨Δri2⟩1/2⟨Δrj2⟩1/2. The correlation matrices, also called dynamic cross-correlation matrices (DCCM), were calculated using the carma software [67]. The length of a path is the sum of the edge weights between the consecutive nodes along this path. And the optimal (shortest) paths between two nodes in the network were obtained by the Floyd-Warshall algorithm [68]. The number of optimal paths that cross one edge is termed as betweenness of the edge. Suboptimal paths within a certain limit (offset) between the two nodes were also determined in addition to the optimal path. The number of suboptimal paths shows the path degeneracy. Communities were calculated based on the dynamical network by the Girvan–Newman algorithm [69]. The nodes in one community are more compactly interconnected than other nodes.
All pathways were determined between residues in the MBD (located within 10 Å of the mismatch site) and residues in the ATPase domain (located within 10 Å of a bound nucleotide) or between residues in the clamp domain (within 10 Å of DNA) and residues in the ATPase domain (within 10 Å of a bound nucleotide). The residue pairs with the shortest optimal path were finally selected as representative residues (S2 Table). Suboptimal paths between specific residue pairs were calculated with edge length offsets of 3, 5 and 10 for the MBD-ATPase, ATPase-ATPase, and ATPase-clamp interactions, respectively.
|
10.1371/journal.pntd.0000165 | Bioinformatics and Functional Analysis of an Entamoeba histolytica Mannosyltransferase Necessary for Parasite Complement Resistance and Hepatical Infection | The glycosylphosphatidylinositol (GPI) moiety is one of the ways by which many cell surface proteins, such as Gal/GalNAc lectin and proteophosphoglycans (PPGs) attach to the surface of Entamoeba histolytica, the agent of human amoebiasis. It is believed that these GPI-anchored molecules are involved in parasite adhesion to cells, mucus and the extracellular matrix. We identified an E. histolytica homolog of PIG-M, which is a mannosyltransferase required for synthesis of GPI. The sequence and structural analysis led to the conclusion that EhPIG-M1 is composed of one signal peptide and 11 transmembrane domains with two large intra luminal loops, one of which contains the DXD motif, involved in the enzymatic catalysis and conserved in most glycosyltransferases. Expressing a fragment of the EhPIG-M1 encoding gene in antisense orientation generated parasite lines diminished in EhPIG-M1 levels; these lines displayed reduced GPI production, were highly sensitive to complement and were dramatically inhibited for amoebic abscess formation. The data suggest a role for GPI surface anchored molecules in the survival of E. histolytica during pathogenesis.
| The causative agent of the infectious disease, amoebiasis, is the parasite Entamoeba histolytica, which targets human intestine and liver. Once in the host, this parasite attaches to human cells and matrix components via factors at its surface such as the Gal/GalNAc lectin and proteophosphoglycans (PPGs). These factors are themselves anchored to the parasite surface by a glycosylphosphatidylinositol (GPI) moiety. To synthesise the GPI, a cascade of enzymes are necessary including the mannosyltransferase 1 (PIG-M1). A homolog of the PIG-M1 enzyme was shown to be present in E. histolytica (EhPIG-M1). To study the role of EhPIG-M1 in E. histolytica, parasites were constructed that had a reduced amount of mannosyltransferase. These parasites displayed a diminished production of GPI molecules and a lower amount of PPGs at the cell surface. Interestingly, the parasites were highly sensitive to the host blood complement and the formation of liver abscesses in hamsters was dramatically impaired. These results suggest that molecules anchored to the cell surface with the GPI moiety have a pivotal role in the survival of E. histolytica during pathogenesis.
| Glycosylphosphatidylinositol (GPI) is a glycolipid required for anchoring many cell surface proteins and glycoconjugates to the surface of a wide range of human parasites including Trypanosoma brucei, the agent of sleeping sickness, Leishmania the causative agent of leishmaniasis, Plasmodium falciparum the agent of malaria and Entamoeba histolytica responsible for amoebiasis [1]. A common feature of the surface of these parasites is the presence of a large glycocalyx containing the GPI-anchored compounds that allow them to interact with their external environment. During invasion of human cells or tissues, the glycocalyx contributes to the adhesive mechanisms sustaining interaction of parasites with their target cells. GPI anchors are structurally complex glycophospholipids that are added to carbohydrate chains, as in the case of glycosylinositolphospholipids (GIPLs) and lipophosphoglycan (LPG) or post-translationally to the C-terminal end of many membrane proteins in the ER. Studies on the variant surface glycoproteins of T. brucei led to discovery of the role of GPI in anchoring proteins to the cell surface [2]. During parasitic infections, GPIs of various protozoan parasites, particularly those of P. falciparum and various Trypanosoma and Leishmania species, can activate host macrophages, triggering the production of proinflammatory cytokines and nitric oxide contributing to disease pathogenesis [1]. Recent studies have suggested that GPI and/or many GPI-anchored molecules could be secreted by the parasites during their invasive process. In the context of human infection by P. falciparum, it has been proposed that the secreted GPI of parasite origin functions as the dominant malarial toxin [3],[4]. GPIs of T. cruzi have the same function [5]. These data support the view that GPIs of the parasitic protozoa are dominant proinflammatory agents playing a role in the immunopathology of these parasitic infections. GPI-anchored molecules also play vital roles in amoebic pathogenesis. During dysentery, amoeba trophozoites bind to colonic mucins and to epithelial cells through the Gal/GalNAc lectin, an immunodominant protein complex containing a GPI-anchored subunit [6]. This lectin associates with another GPI-anchored protein, the intermediate IgL sub-unit. E. histolytica also expresses at its surface an abundant second class of GPI-linked molecules referred as GPI-anchored proteophosphoglycan (PPG) [7],[8],[9]. The GPI anchor of PPGs is unusual because it contains a highly acidic polypeptide backbone modified by 1-6 glucan side-chains and this core is also modified by heterogeneous galactose side-chains. Interestingly, the non-virulent E. histolytica strain Rahman synthesizes one class of PPGs containing short disaccharide side-chains [8] and no similar molecule was detected in the non-pathogenic species Entamoeba dispar. PPGs are important virulence factors during hepatic amoebiasis since monoclonal antibodies recognizing these compounds protect animals from abscess development [10]. The E. histolytica PPGs in addition diverge from the conserved sequence because they contain an anchor with the core structure Gal1Man2GlcN-myoinositol, where the terminal Gal residue replaces the 1-2 linked mannose residue of other eukaryotic GPIs. A large number of studies in yeast and mammalian cells allowed to conclude that steps in GPI biosynthesis are conserved in eukaryotes [11]. In general, biosynthesis of GPI begins in the Endoplasmic Reticulum (ER) with the transfer of N-Acetyl Glucosamine from UDP-N-Acetyl Glucosamine to ER membrane residing phosphatidylinositol (PI). This step is catalyzed by N-Acetylglucosamine transferase located in the ER membranes. This intermediate is then deacetylated to form GlcN-PI by another ER enzyme-GPI-deacetylase (PIG-L). It is thought that GlcN-PI is then flipped to the ER lumen by a set of flippases. Next, a set of mannosyltransferases acts on GlcN-PI to add on three mannose moieties successively to form (Man) 3-GlcN-PI [11]. This intermediate is recognized as a substrate by ethanolamine phosphotransferase to add on an ethanolamine phosphate group to the terminal mannose of the extending GPI glycan core, which is conserved in most eukaryotic cells. The mannose groups in the GPI core are all derived from dolichol-phosphate-mannose (Dol-P-Man). Thus, three Dol-P-Man dependent mannosyl transferases are required for independent addition of mannoses to the GPI-core. The first of these mannosyltransferases is PIG-M1 (Phosphatidylinositol glycan mannosyltransferase ) which transfers the first mannose to the growing GPI anchor from the luminal side [12]. The analysis of E. histolytica genome sequence allowed identification of genes involved in the GPI biosynthetic pathway [13]. Among a total of 22 genes in yeast and 23 in humans, 15 genes were identified in E. histolytica; these genes include all catalytic subunits of the enzymatic complexes sustaining GPI biosynthesis. Studies on the GPI biosynthetic pathway in E. histolytica are rather scarce. Nevertheless, it has been found recently that the antisense RNA-mediated inhibition of EhPIG-L, the GPI-deacetylase, has an important effect on cell growth, endocytosis and parasite adhesion to human cells [13].
In this report, we present the molecular identification in E. histolytica of PIG-M1. The analysis of the sequence of EhPIG-M1 reveals conserved residues that may play a role in the enzymatic catalysis and/or in the maintenance of the spatial structure of the protein. We demonstrated that anti-sense inactivation of mRNA encoding EhPIG-M1 leads to i) the accumulation of GlcN-PI intermediate; ii) a reduction of GPI contents on the amoeba surface; and iii) the inhibition of abscess formation in infected animals. The loss of parasite virulence phenotype may be due to an increase of complement sensitivity of the inactivated parasite strains.
UDP-[14C]N-acetyl glucosamine was obtained from Amersham, U.K. GDP-mannose, Dolichol mono phosphate (Dol-P), phosphatidylinositol, tunicamycin, tetracycline and mannosidase were from Sigma (USA). EN3HANCE spray for surface autoradiography was from Dupont, NEN, (France). HPTLC plates were from Merck. FLAER was obtained from Protox Biotech (Canada).
Pathogenic Entamoeba histolytica (strain HM-1: IMSS) and derivatives were cultivated axenically in TYI-S-33 medium at 37°C [14]. Transfected trophozoites were maintained in presence of hygromycin at 5 µg ml−1 and the drug concentration was raised to 30 µg ml−1 for 48 h, then tetracycline was added at 1 µg ml−1 for 5 days before harvesting parasites.
Bacterial strain Escherichia coli TG1 was used for amplification of the plasmid constructions. Bacteria were grown in Luria-Bertani medium. Bacteria bearing plasmids were grown in presence of 50 µg ml−1 ampicillin.
A cDNA fragment identified during EST program sequencing [15], was translated and the predicted peptide was compared by BLAST algorithm with the E. histolytica genome data base at TIGR. From growing trophozoites (107) of the HM-1:IMSS strain, messenger RNA (mRNA) was prepared using Trizol reagent. To identify the mRNA encoded by pigM1 gene a RACE experiment was performed according to manufacturer's protocol (Invitrogen, USA) using 5 g of total amoeba RNA. The reverse oligonucleotides used for 5′-RACE were designed on the XM_64 49 88 DNA sequence but are common to the two potential ORFs. For RT, ManT15 starting at base 650-5′AAA GCA GTT CCA AAT ACT GC; and for the nested PCR, ManT16 starting at base 581-5′AAA CAA AAG AAT AAT GGA AGT G. The forward oligonucleotide was in the anchored sequence added by RACE.
The first 708 base pairs from the E. histolytica pigM1 encoding gene (Accession number at NCBI: XM_644988) were amplified by PCR using the oligonucleotides 5′ GAG GAT CCA TGG GAA TAA AAG GTC AAG AAGG and 5′ CCA AAG AAA TTC GGT ACC ATA ACG ATA ATA AT that assure the insertion of a BamHI site at the 5′ end of the amplified fragment and a KpnI site at its 3′ end. The amplified DNA fragment was cloned into the KpnI-BamHI sites of the Tet plasmid resulting in an inversion of this fragment compared to the oriented sense of the endogenous gene [16]. The resulting construct was verified by DNA sequencing and the recombinant plasmid was introduced in a virulent E. histolytica HM-1: IMSS strain recently passed through a hamster liver. As a control, the Tet vector containing the cat gene was also transfected. After five days of Tet treatment mRNA was purified with Trizol. Quality and integrity of purified RNAs was checked by spectrophotometry at 230, 260, 280 and 320 nm, electrophoresis on 0.8% agarose gel and assay on Bioanalyzer 2100 (Agilent, USA). Purified RNAs were retrotranscribed with Superscript II Retro-Transcriptase (Invitrogen, USA) according to manufacturer's protocol with specific primers detecting the sense or the antisense species of mRNA. In the same reaction mix was retrotranscribed the L9 sense mRNAs (60S ribosomal protein L9-encoding gene), which was used as endogenous control.
We submitted ORF1 from E. histolytica, i.e., EhPIG-M1, to the GLOBE program (http://www.predictprotein.org/) for predicting the degree of globularity. GLOBE is based on figuring out the accessibility to the solvent. We then BLAST-searched the SwissProt and TrEMBL databanks for homologous sequences. From this output, we selected the sequences of six organisms (H. sapiens Q9H3S5; R. norvegicus Q9EQY6; M. musculus Q99J22/Q8C2R7; C. elegans Q17515; D. melanogaster Q9W2E4; T. brucei Q9BPQ5) for multiple alignments using the T-Coffee program [17]. Since these analyses indicated that the amino acid sequence of EhPIG-M1 corresponded to that of a membrane protein, we used the SignalP program [18] in order to predict the cleavage site of the sequence signal peptidase, i.e., the boundary between the signal peptide and the mature protein. To obtain the secondary structure content of the protein, we submitted the sequence to the PROFsec routine in the PHD suite of programs (http://www.predictprotein.org/).
We confronted the entire sequence to PROSITE, a database of protein families and domains (http://us.expasy.org/prosite/) in order to identify those sequence motifs that were conserved throughout the selected sequences and those that were proper to EhPIG-M1. For detection of the ER retention motif, we used the PSORT package [19].
We submitted the EhPIG-M1 sequence with and without the signal peptide to transmembrane prediction programs that have shown the highest accuracy in a recent comparative study [20]. These programs are TMHMM2 [21], HMMTop2 [22] and PHDpsihtm08 [23],[24]. The former two methods are based on hidden Markov models, whereas the latter is based on neural networks.
Once the transmembrane segments were predicted, we used the TopPred algorithm [25] for predicting the orientation of each transmembrane helix with respect to the membrane. The TopPred method makes use of the “positive-inside rule” which orients a transmembrane segment so that the difference in charge between the residues surrounding that segment and belonging to the intracytoplasmic and intraluminal media is positive. In addition, we calculated the hydrophobicity moment of each helix using the aWW scale [26], which takes into consideration the hydrophobic character of the amino acid residues, as well as the intra helical interactions between the side chains. Moreover, we also ran a prediction in which we imposed the following constraints: the N- and C-termini of the immature protein are intracytoplasmic, and the glycosylation site is intraluminal.
According to the amino acid sequence of EhPIG-M1, two peptides predicted to be located between two transmembrane domains were chosen for commercial antibody production (EUROGENTEC, Belgium). Two rabbits were injected with a mix of these peptides coupled to KLH. After four immunizations, sera from rabbits were tested by ELISA immunoreaction using the peptides and then specific IgG were purified by peptide affinity chromatography that uses the two peptides.
Amoeba extracts from 106 trophozoites were obtained after growing and washing the parasites with PBS and incubating the trophozite pellet in 100 µl of a protease inhibitor cocktail then boiled for 10 minutes. The extract was then mixed with Laemmli buffer and proteins were resolved in an 10% SDS-PAGE (equivalent to 104 amoeba by lane). Resulting gels were Comassie bleu stained or used for immuno-detection. The preimmune serum (1∶500 dilution) and the peptide purified anti EhPIG-M1 (dilution 1∶2000) were used along with a secondary antibody (1∶25000) recognizing rabbit IgG (Jackson, USA) according to ECL revelation procedure (Amersham, UK). The primary anti-actin antibody (clone 4, Chemicon, USA) was used to normalize the protein loading and the ECL western blot image using the ImageQuant v.1.2 software (Molecular dynamics, GE Healthcare, USA). The ratio of EhPIG-M1 to actin allowed us to accurately assess variations in the quantity of EhPIG-M1 in the protein extracts. The experiments were performed twice.
A million E. histolytica cells were labeled with 10−8 M concentration of FLAER (Protox Biotech, Canada) following Vats et al., 2005 [13]. For in situ epifluorescence labeling of PPG, 5×106 parasites were incubated as described [10] using monoclonal antibody EH5 (1∶100 dilution) kindly provide by Dr Michael Duchene (University of Vienna, Austria). Fluorescent samples were examined on a Zeiss confocal laser scanning microscope with the pinhole fixed to 83 µm for all samples. Observations were performed in nineteen planes from the bottom to the top of each cell. The distance between scanning planes was 1 µm.
Membranes were prepared from E. histolytica as described by Carver and Turco [27] and Vats et al, 2005 [13]. Crude membrane preparations (typically 1 mg protein) were used to analyse the GPI biosynthesis intermediates with 14C UDP N-acetylglucosamine (0.05Ci) in buffer [50 mM Hepes/NaOH (pH 7.5), 5 mM MgCl2, 5 mM MnCl2, 0.1 g/ml tunicamycin, 2M PI (phosphatidylinositol), GDP-Mannose (1.7M), Dolichol-P (20 g/ml), 1g/ml leupeptin, and 2mM PMSF], in a total volume of 200 l. Purification and analysis by thin layer chromatography (TLC) of intermediates was done according to Vats et al., 2005 [13]
Male Syrian golden hamsters (Mesocricetus auratus), aged 6 weeks, with an average weight of 100 g, were inoculated intraportally, intrahepatically or intraperitonial with 5×105 HM-1:IMSS trophozoites, anti-sense EhPIG-M1 and control strains [28]. Drinking water provided to these animals was supplemented with 50μg/ml of tetracycline 48 hours before inoculation and during the entire infestation procedure. Hamsters were sacrificed 7 days after inoculation. The livers were removed, inspected for the presence of amebic abscesses and photographed.
Lysis of 4×106 parasites by hamster complement was determined on wild type pathogenic HM-1:IMSS strain, PIG-M1-AS or control transfected strains cultured in presence of tetracycline. Hamster serum was used as the complement source and heat-inactive serum (56°C, 30 min.) was prepared as a control. Serum concentration was fixed at 30% of the final volume; a concentration that is not toxic for virulent strains. Three different times of contact were experimented: 15, 30, 60 min. Twenty l of trophozoite suspension were mixed in 96 wells plate with 30 l of serum and 50 l of PBS containing 2,5 mM MgCl2, 10 mM EGTA in a final volume of 100 µl; presence of Mg/EGTA activate the alternative pathway. Control, heat-inactive serum, was assayed in the same conditions. The assay mixtures were incubated at 37°C. Samples were removed at each defined time interval and resting intact trophozoite were counted using Malassez chamber and phase contrast microscopy.
The percentage of lysis resistance was calculated as = % of complement resistance = Ns/Ndc×100. Samples were analyzed in duplicate and the experiments were repeated three times. (Ns; number of viable trophozoites after contact with fresh serum, Ndc: number of viable trophozoites after contact with decomplemented serum). Statistical parameters were determined using a two-way analysis of variance (ANOVA). Inter-strain differences were tested using Fisher's multiple mean comparison method.
Large scale sequencing of transcripts from E. histolytica [15] allowed us to identify a cDNA fragment that when translated predicted a polypeptide sharing a significant level of sequence similarity with α1,4 mannosyltransferases (PIG-M) from humans. BLAST-based sequence similarity search with the E. histolytica genome revealed two homologues of PIG-M. The locus 1 (XM_644988) displays 38% identity to human PIG-M across the entire length of 412 amino acids. The second locus (XM_647339) encodes a protein of 384 amino acids, out of which a block of 379 amino acids show 97% identity when compared with locus 1. Both putative proteins have the DXD motif. This motif is thought to be involved in binding a manganese ion necessary to interact with the nucleotide sugar substrate, a characteristic of glycosyltransferases [29],[30]. It is generally present in mannosyltransferases. However, the two putative genes differ in their overall organization, such as sequence divergence at 5′ UTR regions and presence of an intron in ORF2. There is a stop codon at amino acid position 226 which falls within the predicted short 9bp intron of ORF2. The predicted intron, in fact, removes the stop codon giving a longer ORF. In order to determine whether or not these two genes were expressed, we used primers allowing for amplification of the 5′ end of both the putative mRNAs by a RACE experiment. Purified mRNA from growing parasites was converted to cDNA following a primer extension approach. The amplified DNA fragment was purified and the nucleotide sequence was determined. The results indicated that only the DNA fragment encoding the ORF1 was transcribed during parasite culture; the product of ORF 1 was then named EhPIG-M1. The data suggests that EhPIG-M1 gene may be expressed by exponentially growing E. histolytica. It is difficult to say at present if locus 2 is a pseudogene or not as it may get expressed under conditions not tested by us.
The two-dimensional structural organization of EhPIG-M1 was deciphered by using a number of bioinformatics tools as described in experimental procedures. The overall predicted organization of the protein with respect to the lipid bilayer membrane is shown in Figure 1A. EhPIG-M1 is a polytopic membrane protein likely to be associated with the ER/Golgi system. A signal peptide is present at the N-terminus. The model produced by HMMTop2 in absence of the signal peptide sequence gave equivalent results whether the topographic constraints (N- and C-terminal extremities intra-cytoplasmic and glycosylation site intraluminal) were introduced or not. EhPIG-M1 is composed of 12 transmembrane spans -the signal peptide and 11 transmembrane helices with two large intraluminal loops, O1 and O4 (Figure 1A and Figure S1). The N-terminus of the immature protein is intra-cytoplasmic, in agreement with the known mechanisms of protein translocation in the ER [31],[32]. The signal peptide transmembrane fragment is predicted to be in a helical conformation. The first loop (O1, residues M31-I98) in the immature protein contains the DXD motif. The release of the N-terminal anchor point upon release of the signal peptide should facilitate the folding of this fragment. Surprisingly, the EhPIG-M1 sequence presents no Carbohydrate Recognition Domain (CRD) that could interact with mannose. In addition, we find no motif consisting of two glutamates separated by seven residues, motif whose importance for the enzymatic activity of glycosyltransferases has been recognized. Two conserved amino acids surround the DXD motif–a threonine and a tyrosine. The extended 45TDIDY53 motif corresponds to a potential phosphorylation site, possibly involved in regulating the catalytic act. According to the aWW scale, the transmembrane helices in EhPIG-M1 are in general hydrophobic, rather than amphipathic. In our 2D model, the following pairs of helices are linked by small loops, mostly hydrophilic and, being geometrically constrained must closely pack: 1-2, 2-3, 3-4, 4-5, 5-6 and 7-8, 8-9, 9-10, 10-11. Thus because the O4 loop (E229-K267) between transmembrane helices T6 and T7 is rather large, the transmembrane helices in EhPIG-M1 may consist of two structural domains, one composed of helices 1-6 and the other of helices 7-11 (Figure 1A). The topology of the proposed model in the usual notation is as follows: (cytosol) Nter-(single TM)-GD-(multiple TM)-GD-(multiple TM)-Cter, TM representing the transmembrane passages and GD the globular domains. The overall hydrophobicity of most of the passages leads us to believe that thay will have a tendency to conglomerate
The 2D model of EhPIG-M1 inferred (Figure 1A), predicts that the two internal peptides 64VNGESPYRRATYRYTPL80 (peptide I) and 238TYLYHGTRTDHRHNL252 (peptide II) do not belong to any membrane domain. Taking into account this sequence information, peptides I and II were synthesized and used to prepare an anti-EhPIG-M1 antibody (precisely amino acids 62 to 74 and 232 to 244). For that purpose; rabbits were immunised with a mixture of these peptides in order to raise specific antibodies. The purified antibody was used in immunoblots to investigate the presence of EhPIG-M1 in E.histolytica protein extract. A unique polypeptide of apparent molecular mass of 47 kDa was identified (Figure 2A). This matched the expected mass of the EhPIG-M1 protein. To investigate the potential role of EhPIG-M1 in E. histolytica pathogenesis we constructed a parasite strain transcribing an antisense RNA of the first 708 nucleotides of the EhPIG-M1 encoding gene using a tetracycline (tet) inducible gene expression system [16]. We first determined whether the plasmid construct induced transcription of an antisense EhPIG-M1 RNA in parasites treated by tet. RT-PCR analysis showed the presence of a specific band indicating the presence of AS-RNA (Figure 2B). The level of EhPIG-M1 protein in transfected cells was analysed by a western blot assay and the results showed that the abundance of EhPIG-M1 in the antisense expressing parasites was reduced to 60% (Figure 2C) of that from control parasites. These data suggest that induction of anti-sense RNA leads to reduction of EhPIG-M1 cell content.
To investigate a consequent potential reduction in GPI content, FLAER labeling was done on living PIG-M-AS trophozoites, incubated with or without tet and examined by confocal microscopy. There was roughly 4-5 fold decrease of GPI- labeling by FLAER in tetracycline induced PIG-M-AS cells against uninduced cells (Figure 2D). Differential labeling was particularly seen at the level of the number of fluorescent vesicles within a cell. Induction of antisense RNA by tet significantly reduced the number of fluorescent vesicles suggesting a reduction of the amount of GPI within the cell. Since one of the most abundant GPI-anchored molecule at the E. histolytica surface is proteophosphoglycan (PPG), we investigated whether changes in overall PPG content appeared subsequently to reduction of EhPIG-M1 level. Using confocal microscopy and monoclonal anti-PPG antibody EH5 (mAb5), we observed a low but significant reduction of PPG on the amoeba surface (Figure 2D). However, there was no detectable change in the level of Gal/GalNAc lectin light subunit (anchored by GPI on the amoebic surface) in tet induced EhPIG-M1 antisense cells (data non shown). There was also no significant difference in the level of phagocytosis of red blood cells or surface receptor capping activity on tet induction in these cells (data non shown) suggesting that these two phenomena were not modified by the reduction of EhPIG-M1 levels.
If EhPIG-M1 encodes a functional α1,4 mannosyltransferase, then inhibition of its expression by antisense RNA should lead to an accumulation of the GlcN-PI intermediate, a substrate for PIG-M. A decrease in the synthesis of Man-GlcN-PI (product of PIG-M activity) is also expected. The first two intermediates of GPI-biosynthesis have been previously characterized [13] as GlcNAc-PI and GlcN-PI, stressing on the functional existence of the GPI biosynthesis pathway in E. histolytica. Therefore, to validate the functional aspect of EhPIG-M1 we set up an in vitro reaction with crude membrane preparation and labelling with [14C] UDP-N-acetylglucosamine the intermediate products of GPI biosynthesis. The radiolabeled GPIs synthesized by the parasite extract were isolated and quantified in TLC as previously described [13]. Previously isolated and characterized GlcNAc-PI and GlcN-PI were taken as control and migration references. Addition of GDP-Mannose and Dolichol-P resulted in the formation of a new species X, which migrated slower than GlcNAc-PI and GlcN-PI (Figure 3A). Since no other radiolabeled spots were obtained besides GlcN-PI and X, we concluded that GlcN-PI was processed to X. Densitometric analysis showed that there was a decrease of roughly 50% in the production of X compound when membranes from tet-induced antisense parasites were used and compared to the non induced set (Figure 3B). The data suggests that EhPIG-M1 may be using GlcN-PI as a substrate producing X, which is probably the mannosylated form of GlcN-PI. In correlation, GlcNAc-PI accumulates in the extract from EhPIG-M1-AS strain.
The virulence phenotype of EhPIG-M1-AS cell line was examined by using the liver abscess model in hamsters with and without tet induction. Animals were infected with PIG-M-AS cells that were cultivated for five days in the presence of tetracycline. A control was performed by injection of animals with amoebas carrying the vector plasmid, with the cat gene treated in the same manner. Infected hamsters drank water containing tetracycline for the period of infection and then they were sacrificed for pathophysiology analysis. A dramatic reduction of pathogenicity was observed (Figure 4A). In contrast to the controls, abscesses were not seen in animals infected with EhPIG-M1-AS parasites in spite of infection carried out using different routes, that is, intrahepatical, intraperitonial or intraportal. A histological observation of hepatical lesion of tissue in the area of injection demonstrated, by comparison with the controls, that EhPIG-M1-AS trophozoites did not survive in the living tissue. We then used live imaging by two-photon microscopy to follow PIG-M1-AS strain during infection in real time. We observed that tet-induced PIG-M1-AS cells were completely lysed as soon as these trophozoites were in contact with the animal tissue. The sensitivity of PIG-M1-AS strain to living tissue factors precluded any further video-microscopy analysis and prompted us to measure the behaviour of these parasites in the presence of hamster serum, rich in complement.
To test the parasite survival in the presence of serum, trophozoites from the wild type strain, the control and EhPIG-M1-AS cells grown in the presence and the absence of tet were incubated with 30% of hamster serum for different periods of time. A large destruction of EhPIG-M1-AS strain expressing the antisense construct was observed within 15 minutes of incubation (Figure 4B) whereas the WT and the non-induced PIG-M1-AS cells were resistant (p<0.001). All the parasite cells were insensitive to heat inactivated serum with no complement activity. Overall the results suggest that reduction of EhPIG-M1 by anti-sense technology is able to reduce cell surface GPI molecules and accounts for inhibition of important pathogenicity features such as resistance to complement and ability to cause liver abscess.
In this work we have identified an E. histolytica protein homolog of mammalian PIG-M, an endoplasmic reticulum–localized α1,4 mannosyltransferase required for synthesis of the glycosylphosphatidylinositol (GPI) anchor. PIG-M is the enzyme that incorporates the first mannose into the GPI core [11]. The EhPIG-M1encoding gene (XM_644988) is transcribed in growing parasites as assessed by RACE-PCR experiments and by cDNA sequencing. Bioinformatics analysis of EhPIG-M1 secondary structure reveals that it is a transmembrane protein probably residing in the ER. The C-terminus containing the retention motif is intra-cytoplasmic indicating that it can be recognized by COP I, a specific cytoplasmic protein that retrieves proteins from the Golgi after interaction with the ER retention motif [33],[34]. However, the potential ER retention motif -403LRKQKQLKLN412- is atypical. Since the organization of the ER and Golgi of this early branching eukaryote E. histolytica is different from other eukaryotes it is possible that the ER retention signal may be different [33],[34]. The cellular location and trafficking of EhPIG-M1 are matters under investigation.
There are a few successful examples of functional characterization of E. histolytica genes using antisense RNA [13]. This approach is particularly useful in absence of a method available to carry out targeted gene deletion. In this study, the antisense RNA approach has been used to understand the role of EhPIG-M1 in amoebic virulence. This strategy generated parasites with 40% reduction in EhPIG-M1 content that leads first to the accumulation of GlcN-PI, the substrate of EhPIG-M1, indicating that there is a loss of activity of this enzyme. These parasites also displayed a radical phenotype shown by dramatic changes in their survival when incubated in the presence of complement and are non virulent in the hamster model of hepatical amoebiasis. The reduction of GPI and/or GPI-anchored molecules at the amoeba surface observed by staining should account for these phenotypical changes leading to loss of virulence due to their enhanced susceptibility to complement action (Figure 4).
The human complement system, as part of the humoral innate immune system, is essential for recognition of microbes, opsonization followed by intracellular killing by phagocytes, or direct lysis. Previous works have shown that the extracellular cysteine proteinases of E. histolytica activate the complement pathway by specifically cleaving C3 leading to a modified C3b-like protein thus preventing terminal membrane attack complex (MAC) formation [35]. This interesting mechanism of complement activation (which should appear as a disadvantage for the survival of the parasite), leads to passive lysis of non-pathogenic strains; whereas pathogenic strains are resistant to complement attack [36]. Moreover, the Gal/GalNAc lectin, present on the amoeba surface, inhibits the assembly of lytic MAC [37]. Complement binds to E. histolytica surface and it is believed that receptors for complement molecules exist in this parasite. However, the mechanism by which amoeba resist complement action is not well known and whether GPI-anchored molecules participate in this feature remains to be molecularly established. However, examples for the involvement of GPI-anchored molecules in complement resistance have come from Leishmania. This parasite resists by prevention of binding at their surface of the attack complex. LPG, a GPI-anchored glycoconjugate, is a major C3 binder [38]. MAC is formed on the top of the LPG coat and does not reach the membrane bilayer leading to its spontaneous elimination. Interestingly, a Leishmania major lpg1- mutant, which lacks LPG, shows attenuated virulence and is highly susceptible to human complement, indicating that LPG can act as a biological barrier thus protecting parasites from complement attack [39]. The experiments reported here suggest one important role of GPI and of GPI-containing molecules for E. histolytica pathogenesis. Increasing our knowledge of GPI biosynthetic pathway in this pathogen will open opportunities for the discovery of alternative treatments against amoebiasis. |
10.1371/journal.pgen.1007580 | Host genetics and the rumen microbiome jointly associate with methane emissions in dairy cows | Cattle and other ruminants produce large quantities of methane (~110 million metric tonnes per annum), which is a potent greenhouse gas affecting global climate change. Methane (CH4) is a natural by-product of gastro-enteric microbial fermentation of feedstuffs in the rumen and contributes to 6% of total CH4 emissions from anthropogenic-related sources. The extent to which the host genome and rumen microbiome influence CH4 emission is not yet well known. This study confirms individual variation in CH4 production was influenced by individual host (cow) genotype, as well as the host’s rumen microbiome composition. Abundance of a small proportion of bacteria and archaea taxa were influenced to a limited extent by the host’s genotype and certain taxa were associated with CH4 emissions. However, the cumulative effect of all bacteria and archaea on CH4 production was 13%, the host genetics (heritability) was 21% and the two are largely independent. This study demonstrates variation in CH4 emission is likely not modulated through cow genetic effects on the rumen microbiome. Therefore, the rumen microbiome and cow genome could be targeted independently, by breeding low methane-emitting cows and in parallel, by investigating possible strategies that target changes in the rumen microbiome to reduce CH4 emissions in the cattle industry.
| Methane is a potent greenhouse gas and ruminant livestock contribute a substantial amount of total methane from human activities. Variation between cows’ methane production has been found partly due to their genetics (heritable), making genetic selection a promising strategy for breeding low methane emitting cows. We hypothesized that the total methane production by a cow is affected by rumen microbes which are directly responsible for production of methane, as well as the cows’ own genetics and their interaction. We sampled the rumen contents of 750 dairy cows and found the relative abundance of some bacteria and archaea to be heritable and associated with methane production, but the majority of variation in relative abundance of rumen bacteria and archaea is due to non-genetic factors. We compared the amount of variation in methane production associated with host genetics as well as rumen bacteria and archaea and found the host genetics to explain 21% and rumen microbes 13%. Importantly, the two were largely independent of each other, so breeding for low methane emitting cows is unlikely to result in unfavorable changes in the rumen microbiome. However, further functional annotation of rumen microbiota is needed to confirm this. Strategies that target each source of variation can be conducted in parallel to optimize reduction in methane production from dairy cows.
| Methane (CH4) is a potent greenhouse gas (GHG) with a climate change potential ~32 times greater than carbon dioxide (CO2)[1] and an atmospheric half-life of 12 years, which is substantially shorter than CO2 (> 100 years)[2]. Therefore, reducing CH4 emissions from anthropogenic-related sources has been identified as a key area for mitigating climate change with immediate effects[2,3]. Livestock accounts for 14.5% of anthropogenic-related GHG emissions and enteric CH4 emissions from ruminants accounts for 5.8%[3]. Furthermore, CH4 emissions from livestock is predicted to markedly increase due to an expected doubling in the global milk and meat demand by 2050[4].
Ruminants, the most widespread livestock species, can digest a wide variety of high fiber feedstuffs due to the distinct microbiome in their rumen. Methane is a natural by-product of gastro-enteric fermentation of high fiber plant biomass by microbial enzymatic activity in the rumen [5]. Bacteria, protozoa, and fungi in the rumen produce CO2 and hydrogen (H2), which are converted to CH4, primarily by archaea known as methanogens. Approximately 99% of CH4 emitted from cattle is released in the breath by eructation and respiration[6]. The emission of CH4 is also a crucial pathway for maintaining H2 balance and ruminal pH, as the optimal conditions for anaerobic fermentation by the rumen microbial community is limited to a narrow range of partial pressure of H2 and pH [7]. Hydrogenase-expressing bacteria convert metabolic hydrogen from anaerobic fermentation into H2 which is then converted to CH4 via methanogenesis [7]. Furthermore, emitted CH4 has a caloric value and represents a 2–12% net loss of a cow’s gross energy intake[8,9]. Consequently, cattle and other ruminants with increased efficiency to digest high fiber feedstuffs but reduced CH4 production could in principal benefit the global climate and concurrently improve the profitability and sustainability of cattle production.
Mitigation to decrease CH4 production by cattle to date has been largely unsuccessful, as the available measures are temporary and not cumulative. Large international research approaches target the rumen microbial communities through feed additives (chemical or biological), feed formulations, and anti-methanogen vaccines[10]. However, rumen microbial species rapid adaptation to changes in the substrate results in resistance to treatments and CH4 production returns to pre-treatment levels[11]. Conversely, rumen transplantation studies (transfaunation) show that the rumen bacterial community recovered to near pre-transfaunation composition after a short period of time[12]. This indicated the existence of a degree of host influence on rumen microbial composition[12]. Host genotype in cattle was reported to explain inter-animal differences in CH4 production[13,14] and the rumen microbial community influenced CH4 production[15]. However, empirical evidence linking the host’s genetic influence over the rumen microbial community and CH4 production is rather limited[15].
A promising strategy is genetic selection for low CH4 emitting cows, as it is sustainable, persistent, and cumulative over subsequent generations. Whether the host influences the rumen microbial community, and consequently CH4 production, or the two interact to affect CH4 production is currently unknown. If reduced CH4 production in cows is a consequence of poor symbiosis with rumen microbes and thus fiber digestibility, there is a risk selection for reduced CH4 production will act against the very symbiosis which has aided ruminants and rumen microbes’ coexistence. Thus, the extent to which the rumen microbiome is under the host genetic influence needs elucidation. If host genetics impose a strong influence on rumen microbial composition, traits influenced by rumen microbes could be improved by using rumen microbial composition as indicator traits in selection. However, should host genetics impose a strong influence on rumen microbial composition and selection for CH4 production proceed without cognizance of rumen microbial composition, there is a risk of unfavorable correlated responses in rumen microbial composition.
We hypothesized that: 1) the relative composition of the microbiome in the rumen is heritable i.e. controlled by host genome and 2) variation in methane emission from rumen is influenced by both the cow genome and rumen microbial content.
Methane concentration in the exhalation-breath of 750 lactating Holstein dairy cows from farmer herds in Denmark was measured individually during automated machine milking for one week. Within-week methane measurements had a high repeatability coefficient of 0.70 ± 0.02 (estimate ± SE). Estimated average daily methane emission was 395.8 ± 63.5 g/d (mean ± SD), which was consistent with reports from the literature[16]. Considerable variation in estimated CH4 emission among cows was observed. The top 10% methane emitting cows (519.28 ± 28.5 g/d) had a 41% mean difference from the low 10% emitting cows (303.8 ± 11.9 g/d) (S1 Fig). Results from linear mixed model with pedigree records indicated methane emission was moderately heritable, 0.19 ± 0.09 (heritability coefficient, h2 ± S.E), which was consistent with previous findings in lactating Holstein cows in Denmark[13].
We identified 3,894 bacterial operational taxonomic units (OTUs, ≥ 97% identity) and 189 archaeal OTUs, which were present in a minimum of 50% of the cow samples (50% threshold maximizes the variation in a binary trait i.e. presence or absence). Taxonomic classification revealed generic bacterial and archaeal composition. The predominant bacterial phylum found was Bacteroidetes 72.2% ± 6.5 (mean ± SD), followed by Firmicutes (18.3% ± 5.6) and Tenericutes (2.8% ± 1.0). Absconditabacteria, Spirochaetes, Fibrobacteres, and Proteobacteria each comprised less than 2%, and another 20 phyla constituted 1% of all sequence reads.
The archaeal community was dominated by two families, Methanobacteriaceae and Methanomassiliicoccaceae (35% ± 22.1) of the orders Methanobacteriales (64.2% ± 22.2; mean ± SD) and the recently proposed order Methanomassiliicoccales and class Thermoplasmata[17], respectively. The remaining archaeal community was comprised of 10 families, which were low in abundance, cumulatively accounting for less than 1% of all archaeal sequence reads.
OTU abundance and OTU abundance collapsed at genus and family levels were used as microbial phenotypes. The heritability thereof was estimated using a linear mixed model with pedigree records (known as ‘animal models’), which partitions total variance into additive genetic and environmental variance[18]. We calculated 95% confidence intervals for OTU h2 estimates and found for 6% of bacterial and 12% of archaeal OTUs, the estimates were significantly higher than zero (P < 0.05), ranging from 16–44% (Fig 1) and 18–33% (Fig 2), respectively. Due to the high number of independent tests, we calculated false discovery rate (FDR) corrected P—values for h2 estimates with a FDR threshold of 15% (S1 Table).
Heritability of bacterial and archaeal abundance was further estimated at the genus level. In total eight bacterial genera out of 144 showed significant h2 estimates ranging from 0.17 to 0.25 (Table 1). Only a single archaeal genus, Methanobrevibacter, had a h2 estimate significantly different from zero (0.22 ± 0.09). However, Methanosphaera and Methanomicrococcus might also be under host additive genetic control with heritability estimates approaching significance thresholds (Table 1).
Associations between relative bacterial and archaeal OTUs, genera abundance, and host CH4 emissions were tested, while simultaneously controlling for environmental factors and familial structures common in livestock due to relatedness among study samples [19,20]. The OTU or genera log-transformed abundance present in > 50% of cows were fit as an explanatory variable in a linear mixed model for CH4 production. Numerous significant OTUs were detected but failed to pass the threshold for multiple testing (FDR ≤ 0.15) (Supplementary Table 1). This was a hypothesis-generating analysis and not directed at specific hypothesis testing therefore we reported the significance and FDR corrected values (S1 Table). Seven genera in total were detected, which exceeded the significance threshold at FDR of 15%. The -log10 P-values are plotted in Fig 3.
Traditionally, dimension reduction techniques such as principal coordinates analysis (PCoA) are used to summarize community composition differences between individuals (Beta diversity) into clusters, which are further examined for associated biological or explanatory variables. Differences in bacterial and archaeal community structures were estimated for the entire sample population at OTU level using the Bray-Curtis[21] dissimilarity metric (PCoA, Fig 4A and 4B). Briefly, the Bray-Curtis dissimilarity is the sum of minimum counts of shared species in two animals divided by the sum of counts of all species in each animal, where 0 indicates the same composition and 1 indicates no shared composition. Analysis revealed clustering of cows into ‘ruminotypes’ for both bacterial and archaeal community composition, which both associated significantly with high and low CH4 emitters at opposing polar regions (Mann-Whitney test, P < 0.001) but failed to cluster distinctly from the intermediate CH4 emitters. Analysis of community structures using ANOVA revealed bacterial PCo1 was partly explained by non-genetic factors: parity (i.e. lactation number) (3.6%), sequencing batch (2%) and lactation stage (1%). A genetic analysis controlling for these factors showed PCo1 was likely heritable (0.20 ± 0.10) and thus influenced by the host additive genetics. Bacterial PCo2 was partly explained by the herd of origin (< 1%) and parity (< 1%) and was not heritable (0.02 ± 0.05). Similar findings were observed for archaea, with the variation in PCo1 partly explained by herd (< 1%), parity (19.9%), sequencing batch (5%) and lactation stage (< 1%). The genetic analysis controlling for these factors exhibited moderate heritability (0.39 ± 0.05). Archaeal PCo2 variation was partly explained by herd (< 1%) and parity (< 1%), which were likely not heritable (0.05 ± 0.05).
The relative proportion of variation in CH4 emissions due to rumen microbial composition and host additive genetic components was estimated individually and jointly using linear mixed models. Likelihood ratio tests revealed that fitting either random effect of rumen microbial composition or individual cow’s polygenic component fitted the data significantly better than the null model i.e. including only fixed effects (P < 0.001). The model fitting both random effects (microbial composition and polygenic component) was significantly better (P < 0.001) than models including only one random effect. The proportion of variance in CH4 production explained by the microbiome, here defined as microbiability (m2), was calculated in analogy to the heritability (h2)[22,23]. The contrast between the two intra-class correlation coefficients h2 and m2 with their respective standard errors for all models are depicted in Fig 5. The m2 of CH4 emission estimated individually was 0.15 ± 0.08 (estimate ± S.E) and the h2 estimated individually was 0.19 ± 0.09. Simultaneous estimates of both effects indicated slightly lower microbiability (0.13 ± 0.08), whereas h2 exhibited a corresponding increase (0.21 ± 0.09) as compared to the preceding models fitting only one of the random effects. The combined microbial abundance and additive genetic effects were responsible for ~ 34% of the total phenotypic variation in CH4 emissions.
The results of this study show that estimated CH4 emissions from a dairy cow were partially under the influence of host (cow’s) additive genetics, which explained 19% of the total variation. Of the rumen bacterial OTUs, a modest ~ 6% were associated with host additive genetics exhibiting significant heritability estimates (16–44%) (Fig 1). Similarly, only ~ 12% of archaeal OTU abundance was influenced by host additive genetics, with heritability estimates ranging from 18–33% (Fig 2). However, bacterial and archaeal heritability estimates failed to pass the threshold for multiple testing. Our test was conservative as a large number of taxa were analyzed with many OTUs having little or no influence by the host genome. Studies with larger sample sizes would give more reliable estimates of the heritabilities, especially for lower heritable OTUs. The h2 estimates observed in this study were consistent with findings of intestinal microbiota in mice[24,25] and humans[26,27] and confirm that the majority of variation in rumen microbial abundance is due to factors other than host additive genetics [28]. Interestingly, the patterns of h2 with phylogeny differed between the bacteria and the archaea (Fig 1 and Fig 2). Heritable OTUs were distributed throughout the bacterial microbiome whereas archaea showed increased heritability within the Thermoplasmatales. This highlights the value of collating phylogeny with heritability estimates to focus research into possible mechanisms which predispose differential relative abundance of certain taxa across genetically related cows. The method employed to sample rumen contents is high-throughput and less invasive than surgical procedures, making it better suited to sampling large numbers of cows under commercial farm conditions. Large sample size is critical in genetic evaluations. However, it is important to note that the floral rumen scoop is inserted into an undefined portion of the rumen and likely samples the liquid phase. Recognizing that rumen microbial communities differ between liquid, solid and epimural phases[29], studies testing the repeatability and representativeness of sampling are needed.
We utilized linear mixed model analysis to test for associations between bacterial and archaeal OTUs, genera and families with estimated CH4 emissions, while concurrently accounting for effects such as parity, lactation stage, herd of origin and familial structure from the pedigree. Several bacterial genera associated with CH4 emission were detected. Out of these, four were found either to be affected by methane inhibitors or related to H2 production and other methanogenesis substrates. Three were moderately heritable (0.17–0.25) (S1 Table). One of the identified bacteria, Sporobacter, with a mean relative abundance of 0.01% (Ruminococcaceae, Clostridiales, Firmicutes), belongs to a group with only a single cultured representative, Sporobacter termitidis, isolated from the intestine of wood-feeding termites (Nasutitemes lujae), also known for producing large amounts of CH4. However, when this isolate was co-cultured with an archaea species, Methanospirillum hungatei, CH4 was not produced. S. termitidis was found to generate acetate and methylsulfides, but not H2 or CO2, therefore interspecies H2 transfer did not occur and facilitate CH4 production[30]. The recent discovery and proposed archaeal order Methanomassiliicoccales species found to utilize methylsulfides and H2 in methanogenesis[31], provides a possible mechanism for methylsulfide producers to contribute to CH4 production when H2 producers are present. Methanomassilicoccales was prevalent in our samples (mean relative abundance 35%); therefore, Sporobacter could potentially be contributing to CH4 production via a similar pathway.
We also detected Sphaerochaeta with a mean relative abundance of 0.01%, associated with estimated CH4 production. Genomes from cultured Sphaerochaeta isolates revealed acetate, formate, ethanol, H2, and CO2 were potential fermentation end products[32], many of which are methanogenic archaea substrates[33]. Furthermore, seed extracts from Perilla frutescens (Lamiaceae), a medicinal herb, decreased CH4 production in vitro from rumen samples of lactating dairy cows and decreased Sphaerochaeta abundance[34]. Interestingly, Caro-Quintaro et al.[32] reported up to 40% of the genes from Spaerochaeta species were exchanged with members of Clostridiales (Firmicutes) and this inter-order-species horizontal gene transfer was most extensive in mesophilic anaerobic bacteria, such as the conditions found in termite and ruminant guts[35]. Here 16S rRNA gene sequencing is used as a proxy for metabolic activity but cannot account for inter-order-species horizontal gene transfer. Therefore, full metagenome sequence may have an advantage over the 16S rRNA gene to describe rumen microbial contents.
One bacterial genus detected in the present study, which is positively associated with estimated CH4 production, is classified in the yet uncultured BS11 gut group of the Bacteroidales (mean relative abundance 1.4%). The relative abundance of the BS11 group reportedly decreased concomitantly with CH4 production by dietary methanogenic inhibitors, such as P. frutescens seed extract, mentioned previously[34], monesin and essential oil supplementation in dairy cattle[36,37], and bromochloromethane in Japanese goats[38]. Thus, supporting our finding of a positive association between BS11 and CH4 production. Solden et al.[39] employed metagenomics sequencing and shotgun proteomics approaches to phylogenetically and metabolically resolve the BS11 gut group. They resolved two genera within the group and both exhibited multiple pathways to ferment hemicellulose, a capability previously unknown for BS11. The resulting fermentation end products included acetate, butyrate, propionate, CO2, H2[39] the latter two being methanogenesis substrates. Genes encoding ‘fucose sensing’ pathways were found for only one of the proposed BS11 genera, offering a possible mechanism for interaction between genes in the BS11 group and the host[15]. However, further studies are needed to elucidate the links between CH4 inhibitors, host genes and CH4 production.
Due to the absence of cultured rumen bacteria isolates, an understanding of the metabolic function in many bacterial genera remains in its infancy. However, from the isolates discussed above, results suggested CH4 emissions depend on abundance of bacterial taxa that produce substrates for methanogenesis, such as H2. Remarkably, associations between archaeal relative abundance and estimated CH4 production were not detected in the present study, despite the knowledge that archaea are directly responsible for CH4 production. A meta-transcriptome study in sheep found archaeal transcription pathways and not simply abundance, contributed to inter-animal differences in CH4 production[40]. This study was congruent with conclusions reached in two recent reviews, which examined results from dairy cattle and other ruminant studies employing 16S rRNA[41] and ‘meta-omics’ approaches[42], where bacteria abundance produced and utilized H2 or stabilized pH, which affected CH4 emissions and feed efficiency and archaeal activity matched substrate availability.
The combined effects of the bacterial and archaeal community structure (beta diversity) on estimated CH4 emissions were investigated by conducting PCoA on the archaeal and bacterial communities, which revealed 2–3 clusters for archaea (Fig 4A) and two clusters for bacteria (Fig 4B). Beta diversity is a non-parametric distance measure used in microbiology and ecology to assess the differences between environments or samples (in this case cows) as opposed to alpha diversity which takes into account the diversity within cows. Clusters of a similar nature were first reported in intestinal bacterial community types in humans[43,44], chimpanzees[45], mice[46] and pigs[47], referred to as “enterotypes”, and found associated with specific host phenotypes. This concept was extended to sheep rumen bacterial communities and referred to as “ruminotypes”[48]. The ruminotypes observed herein followed a continuous gradient and did not form discrete clusters, which is consistent with the latest findings in microbiome stratification. [49]. Importantly, we found that animal and farm factors like herd of origin, parity and lactation stage, as well as technical factors, i.e. sequencing batch, contributed to the observed variation and stratification in ruminotypes. Similar findings were reported in rumen bacterial richness at different lactation stages and over different parities[50], suggesting later parities (higher parity cows are older) decreased bacterial richness and increased production[51]. We detected a moderate heritable genetic component acting along PCo1 axis, with h2 of 20% for bacterial and 39% for archaea, when controlling for lactation stage and parity, demonstrating the first evidence of host additive genetic influence on rumen bacterial and archaeal community structure (beta diversity). All the above-mentioned factors contribute to microbiome structure and associations with host phenotypes.
An association was detected between the highest and lowest CH4 emitters and bacterial and archaeal ruminotypes along PCo1, however, ruminotype cluster memberships were not exclusive to high and low emitters. This suggested ruminal bacterial and archaeal community structure provided a modest contribution to CH4 emission. Kittlemann et al.[48] surveyed microbial community composition in multiple sheep cohorts with low and high CH4 yield (methane emission per kg dry matter intake, CH4/DMI). A ruminotype “S” associated with low CH4 yield and enriched with Sharpea azabuensis was reported. A follow up study in sheep also found low CH4 yielding sheep to be associated with ruminotype “S”, enriched with Sharpea spp. It was hypothesized a smaller rumen size and higher turnover rate promoted faster growing bacteria, such as Sharpea, which favor hetero-fermentative growth on soluble sugars, resulting in lower H2 production and subsequently decreased CH4 formation by hydrogenotrophic methanogens[52]. Smuts et al.[53] reported passage rate (and consequently turnover rate) in sheep was heritable, indicating a possible mechanism for host genetics to influence ruminotypes. Methane emission phenotypes differed between the sheep and the present study. Kittlemann et al.[48] assessed the amount of CH4 production per unit of DMI but not CH4 production directly. DMI measurements are not currently recorded on dairy cattle under commercial farms due to the high costs and therefore, CH4 emissions in the present study could not be corrected for feed intake. In light of the differences in phenotype definitions and similarities in ruminotypes between studies, it would be of interest in future work to obtain DMI records on cows and test if the ruminotypes observed show an increased relationship with CH4 yield. The heritability estimates for PCo1 and PCo2 indicates these measures could potentially be used as indicator traits in genetic selection should they be highly correlated to a trait of interest, however PCo1 and PCo2 (beta diversity) does not account for the total rumen microbial variation within and between individuals.
The method employed to measure CH4 production in the present study is high throughput and non-invasive, making it practically viable for measuring large numbers of animals under commercial farm conditions. However, the cost trade off of this method is that it makes use of milk yield and body weight in the estimation of CH4 production. Validation of this method with the ‘gold standard method’ (climate respiration chambers) has yielded highly correlated (r = 0.8–0.89) and concordant (concordance correlation coefficient = 0.84) results in dairy cattle [54,55]. However, the effects of body weight and milk yield on estimation of CH4 cannot be discounted and further research into the relationships between these variables and the rumen microbiome would be of value.
In this study, we quantified the combined effects of all rumen bacterial and archaeal OTUs simultaneously on estimated host CH4 emissions using a microbial relationship matrix among cows. This is a parametric approach similar to assessing both alpha and beta diversity, as total rumen microbial variation within and between individuals is taken into account simultaneously. We expressed the combined effects as the variance ratio due to microbial composition to the total variance in estimated CH4 emissions (m2, microbiability), an analogy to h2. Estimated CH4 emissions had 15% m2, indicating the combined rumen bacteria and archaea abundance of dairy cattle was associated with a considerable amount of variation in estimated CH4 emissions among animals. Ross et al.[56] first proposed the generation of metagenomic relationship matrices in dairy cattle and reported a CH4 emission prediction accuracy of 0.47, explaining 22% of the total variation in CH4 production [57]. However, Ross et al. [57] did not have sufficient data to estimate h2 or microbiability (m2) in CH4 production. A study with 207 pigs employing 16S rRNA sequencing of gut microbes, found eight of the 49 bacterial genera to be heritable and estimated m2 and h2 for feed intake (m2 = 0.16, h2 = 0.11), daily gain (m2 = 0.28, h2 = 0.42) and feed conversion ratio (m2 = 0.21, h2 = 0.19) [23]. Only daily gain had higher h2 compared with m2. These findings suggest agreement with holobiont theory, where variation in the genome and microbiome can cause variation in some complex traits, on which artificial, natural selection and genetic drift can act [58,59]. However, the aforementioned study did not have adequate numbers of animals to estimate m2 and h2 simultaneously to assess the relative interactions between additive genetics and the microbiome. Thus, it was unable to assess if host additive genetics co-influences the microbiome and variation in phenotypes.
In contrast, we estimated m2 and h2 concurrently to examine the shared information between the two effects. Microbiability of estimated CH4 production decreased by two percentage points to 13% and h2 exhibited a corresponding increase from 19 to 21%. This result indicated host genetic effects do interact with the microbial community composition but are not the primary mechanism for host genetic effects on estimated CH4 emissions. A possible explanation for the negligible amount of shared influence between the two relationship matrices might be the small percentage of heritable bacterial and archaeal OTUs. This implies that the rumen bacterial and archaeal communities affected estimated host CH4 emissions independently and host genetics influenced a small portion of these bacteria and archaea. The combined host additive genetics and rumen microbial community composition explained ~ 34% of the total variance in estimated CH4 emissions in dairy cattle. Thus, breeding for low CH4 production can be expected to result in limited correlated genetic responses to shape the rumen microbiome and breeding can likely proceed without taking cognizance of the rumen microbiome for this trait. However, larger studies estimating genetic correlations between rumen microbiota and CH4 emissions and better functional annotation of rumen microbiota are needed to confirm this.
Microbiability estimates can be used as a tool for quantifying the cumulative effects of microbial abundance on phenotypes, e.g. complex diseases and quantitative traits. However, further research is required to elucidate the biological mechanisms shaping microbiability. For example, animal factors known to affect CH4 production and rumen microbial populations, such as passage rates or individual differences in feed intake might influence microbiability estimates. Human intestinal microbiome studies find that numerous disease phenotypes are associated with microbial richness, species abundance, and microbial community structure[60,61]. Subsequent work using stool consistency and opaque markers as proxies for colonic transit time found all three metrics and disease phenotypes are partially confounded with colonic transit time[62,63]. Similarly, in sheep studies, low CH4 yielding sheep are associated with lower retention time and smaller rumens[64], relationships with specific rumen microbial clusters[48] and different bacterial and archaeal species[52]. Therefore, studies are needed to determine if microbial differences among subjects associated with phenotypic differences are causative or are consequences of unknown extraneous factors. It is also necessary to clarify the mechanisms which allow rumen microbes to be passed on to successive generations, to assess the efficacy of perturbations of the rumen microbiome such as probiotics and rumen transplants aimed at desired changes to the rumen microbiome and associated changes in phenotypes[65]. Regardless of the underlying biology, quantifying the relative contribution of rumen microbes and additive genetics to complex phenotypes helps characterize whether the host genome and microbiome are acting jointly as a holobiont and highlights the merits of targeting microorganisms to achieve a specific change in a phenotype or selective breeding. Furthermore, providing additional information, such as relative abundance of rumen fungi and protozoa, or ‘meta-omics’, including meta-transcriptomics or meta-proteomics data can be readily adopted and incorporated into this methodology, offering insights into economically important livestock and disease traits in humans.
Methane production by dairy cows is not only influenced by factors such as feed intake and composition among others, but also the cow’s individual genetic composition and rumen microbial composition. Each cow’s additive genetic effects influence a modest amount of variation in the abundance of a small percentage of rumen bacterial and archaeal taxa, and thereby contribute to variation in rumen microbiome composition and function. We detected associations between CH4 emissions and rumen bacteria abundance, which are known to produce methanogenesis substrates, suggesting bacteria driven CH4 production pathways. Although we detected a heritable component to ruminotypes, the association to CH4 production was weak. Concurrently, host additive genetic effects and rumen microbes contributed to inter-animal differences in CH4 production, however negligible interaction was observed between microbiability and heritability. Consequently, cow additive genetic effects on CH4 emissions were largely unmodulated by cow additive genetic effects on rumen bacteria and archaea abundance. Strategies to reduce CH4 emissions in ruminants can be optimized by a multifaceted approach, for instance, selective breeding to unlock host’s genetic potential and strategies which may effect desired changes in the rumen microbiota like rumen transplantation, and probiotics.
Methane emissions from 750 lactating Holstein cows in five commercial herds were recorded using a portable Fourier Transform Infrared unit (FTIR; Gasmet DX-4000, Gasmet Technologies, Helsinki, Finland)[13,66] and one research herd using a permanently installed non-dispersive infrared (NDIR; Guardian NG/Gascard Edinburgh Instruments Ltd., Livingston, UK)[67]. Briefly, the FTIR and NDIR equipment were installed within the feed bins of automated milking systems (AMS) in each commercial herd with the FTIR for seven consecutive days and the NDIR were permanently placed in the research herd. The FTIR and NDIR device inlets were installed in the AMS feed bins and methane (CH4) and carbon dioxide (CO2) gas concentrations (ppm) sampled continuously every 5 s and 1 s, respectively[66,67]. Cows were milked individually in the AMS and milked on average (18.2 ± 3.4) times during the seven-day period, for durations ranging from five minutes to 12.2 minutes. Mean CH4 and CO2 gas concentrations were corrected for environmental factors, including diurnal variation and day to day differences using a linear mixed model following Difford et al.[67] to approximate daily averages. Measurement stability was assessed by model repeatability and used as data quality control. All herds practiced indoor feeding strategies with ad libitum access to feed and water. A total mixed ration (TMR) was provided, consisting primarily of rolled barley, corn silage, grass clover silage, rapeseed meal, soybean meal and up to 3 kg of concentrate supplement given during milking. Although all commercial herds employed a standardized TMR recipe, ingredient-specific differences among farms were expected to contribute to differences in TMR dietary values over herds.
Weekly mean values for milk yield and body weight were combined with weekly gas concentrations, as described in Lassen et al. [66] and applied to predict cow heat production[68]. During each week of CH4 and CO2 recording at different herds, milk samples were collected to estimate milk fat and protein percentages. Cow fat and protein corrected milk yield (FPCM) was estimated following the national recording scheme (RYK, Skejby, Denmark)[69]. Methane production (L/day) was estimated using the CH4 to CO2 ratio and predicted CO2 emission[70] from the conversion of cow heat production units to CO2 production, following Madsen et al.[71] and then converted to (g/d) using CH4 density at standard temperature and pressure.
Holstein cow pedigree records were traced in the Danish national database (NAV, Skejby, Denmark) as far back as 1926 to construct a pedigree-based relationship matrix for the quantitative genetic analysis.
Immediately following the CH4 recording period, rumen content samples were drawn from individual cows by oral insertion of the probe “Flora Rumen Scoop” [72]. Approximately 40 mL of the liquid fraction containing particulate matter was drawn from the rumen using this method. Trained technicians conducted the sampling to ensure correct probe insertion into the rumen following a previously established protocol [72], recognizing that the location of the flora rumen scoop may differ somewhat from sampling to sampling. The entire “Flora Rumen Scoop” was rinsed vigorously between animal sampling to minimize cross-contamination. Samples were labeled, immediately placed on ice, and transferred to the laboratory within two hours for further processing. Each 40 mL sample was mixed vigorously, a subsample of 1.2 mL rumen fluid was collected, and transferred to a 1.5 mL vial, then snap frozen in liquid nitrogen, before storing at -80°C, until shipped on dry ice to a commercial sequencing company (GATC Biotech, Constance, Germany) for analysis.
DNA extraction, sequencing library construction and sequencing were conducted by GATC Biotech (Constance, Germany). Rumen samples were defrosted at 4°C overnight and vortexed until homogenous. A representative sample (500 μl) containing rumen liquid and solids was used for DNA isolation using the Qiagen QIAamp stool kit (Valencia, United States of America) following the manufacturer’s instructions, modified for the larger sample size[73].
Two primer sets were used to create 16S rRNA libraries, one set for all bacteria and one set for all archaea. Universal bacterial 16S rRNA gene primers (covering the V1-V3 variable regions) 27F: 5’-AGAGTTTGATCCTGGCTCAG-3’ and 534R: 5’-ATTACCGCGGCTGCTGG-3’ were used to generate the bacterial amplicon libraries (expected amplicon size 508 bp)[74]. Universal archaeal 16S rRNA gene primers (covering the V4-V6 variable regions) S-D-Arch-0519-a-S-15 5’-CAGCMGCCGCGGTAA-3’ and S-D-Arch-1041-a-A-18 5’-GGCCATGCACCWCCTCTC-3’ were used to generate the archaeal amplicon libraries (expected amplicon size 542 bp)[75]. Following protocols standardized by GATC Biotech, PCR amplifications were conducted with GoTaq Green polymerase (Promega, Madison, USA) with 30 PCR cycles and a 60°C annealing temperature for the archaeal amplicon libraries and 25 PCR cycles with a 60°C annealing temperature for the bacterial amplicon libraries. The 16S rRNA amplicons were purified using the Axyprep Fragment Select bead purification system (Axygen Biosciences, New York, USA), according to the manufacturer’s instructions. The size and purity of the PCR product was verified on a Fragment Analyzer using a High Sensitivity NGS Fragment Analysis Kit (Advanced Analytical Technologies, Ankeny, USA). Multiplex indices and Illumina overhang adapters were added to both amplicon libraries in a second PCR amplification round (six cycles), followed by Fragment Analyzer analysis to confirm the correct size of the amplicons (Advanced Analytical Technologies, Ankeny, USA). Ninety-six libraries were pooled in equimolar concentrations and sequenced with an Illumina sequencing instrument using the 300 bp paired-end read mode, according to the manufacturer’s specifications. Approximately half the samples were run using the illumina MiSeq platform and half with the HiSeq platform. The 300 bp paired end protocol was adapted to HiSeq by GATC Biotech. The specific samples entered into sequencing batches within each sequencing platform were recorded for subsequent significance testing to examine possible differences between sequencing batches and sequencing platforms in statistical analyses.
Bacterial and archaeal sequence reads underwent quality control, processing and were clustered into operational taxonomic units (OTUs) using the LotuS pipeline[76] with the following options: Sequence truncation length and minimum sequence length after barcode and primer removal was 230 bp. Minimum average sequence quality score was 27, the maximum number of ambiguous bases was 0, maximum homonucleotide run was set to 8. Sequences were filtered away if any of the 50 bp segments in a sequence had average scores below 25 or if the expected number of errors exceeded 2.5 in the binomial error model. The low-quality sequence ends were trimmed by applying a sliding window quality filter with a width of 20 bp and a minimum average quality score within the window of 25. Sequences were truncated if the probabilistic accumulated error exceeded 0.75. The reads were de-replicated and sequences with a minimum of 10 replicates were retained for OTU clustering within the Lotus pipeline. Sequence pairs were merged with Flash[77] and clustered into OTUs based on sequence similarity (97%) with UPARSE[78] and chimeric sequences removed with UCHIME reference-based chimera detection[79]. Representative sequences from each OTU were aligned with ClustalO[80] and a phylogenetic tree built with FastTree2[81]. Representative sequences, the OTU table, and phylogenetic trees were transferred to QIIME (version 1.9.0)[82], where further analyses were performed. Taxonomy was assigned to each OTU using the RDP classifier with a confidence level of 0.8[83] using greengenes (gg_13_8_otus) as the reference database. Unclassified OTUs and OTUs classified to non-target kingdoms were filtered from the OTU tables, i.e. only OTUs classified as k_Bacteria were maintained for the bacterial primer set and similarly OTUs classified as k_Archaea maintained for the archaeal primer set. Finally, samples with < 50,000 sequences were removed and OTUs containing < 10 sequences were filtered out of the OTU table.
All handling of animals were conducted according to 'Metagenomics in Dairy Cows' protocol. The protocol and study were approved by The Animal Experiments Inspectorate, Danish Veterinary and Food Administration, Ministry of Environment and Food of Denmark (Approval number 2016-15-0201-00959).
|
10.1371/journal.pcbi.1003631 | A Novel Cell Traction Force Microscopy to Study Multi-Cellular System | Traction forces exerted by adherent cells on their microenvironment can mediate many critical cellular functions. Accurate quantification of these forces is essential for mechanistic understanding of mechanotransduction. However, most existing methods of quantifying cellular forces are limited to single cells in isolation, whereas most physiological processes are inherently multi-cellular in nature where cell-cell and cell-microenvironment interactions determine the emergent properties of cell clusters. In the present study, a robust finite-element-method-based cell traction force microscopy technique is developed to estimate the traction forces produced by multiple isolated cells as well as cell clusters on soft substrates. The method accounts for the finite thickness of the substrate. Hence, cell cluster size can be larger than substrate thickness. The method allows computing the traction field from the substrate displacements within the cells' and clusters' boundaries. The displacement data outside these boundaries are not necessary. The utility of the method is demonstrated by computing the traction generated by multiple monkey kidney fibroblasts (MKF) and human colon cancerous (HCT-8) cells in close proximity, as well as by large clusters. It is found that cells act as individual contractile groups within clusters for generating traction. There may be multiple of such groups in the cluster, or the entire cluster may behave a single group. Individual cells do not form dipoles, but serve as a conduit of force (transmission lines) over long distances in the cluster. The cell-cell force can be either tensile or compressive depending on the cell-microenvironment interactions.
| Adherent cells sense, transduce and respond to their microenvironment by generating traction forces on their surroundings. To accurately understand these mechanotransduction processes, it is critical to have a robust and reliable method for traction force visualization and quantification. However, most cell traction force microscopy methods are limited to only single cell traction force analysis. Considering that most physiological processes are essentially collective multi-cellular events, there is a need for traction force microscopy methods capable of analyzing traction forces resulting from multiple cells. We have developed a novel and robust multi-cellular traction force microscopy method for computing cell traction on soft substrates, and applied it to compute traction field generated by both multiple cells and cell clusters. We verified the accuracy, robustness, and efficiency of the method by theoretical, numerical and experimental approaches. Our method provides a powerful toolset to pursue the mechanistic understanding of collective biological activities, such as cancer metastasis and neuromuscular interactions.
| Recent research has demonstrated that cells communicate with each other as well as with their microenvironments through mechanical signaling [1], [2], [3], [4], [5], [6], in addition to biochemical ones [7], [8], [9], [10], [11], [12], [13], [14]. Many physiological processes, including cell adhesion [15], [16], [17], cytoskeleton polarity [13], [18], cell proliferation [19], [20], cell differentiation [12], [21], [22], embryogenesis [23], [24], cancer metastasis [7], [25], and wound-healing [26], [27], can be significantly influenced by the transmission and sensation of physical forces between the cells and their microenvironments. For example, exposure of HCT-8 human colon cancer cells to soft substrates results in a profound stable cell state transition from an epithelial phenotype to a metastasis-like phenotype (MLP) [7], [8], [28], [29], [30], [31]. Adherent cells actively sense the local anisotropy of their microenvironment [2], [18], [32], [33] as well as the forces applied by neighboring cells [1], [4], [11], [34], [35], followed by polarization of stress-fibers and synergetic cell functions. Hence, accurate estimation of the traction forces exerted by the cells on their substrates under various physiological conditions can provide important insight on many fundamental questions regarding the mechanical interactions between various cell types and their microenvironment [36], [37], [38]. Over the past few decades, several seminal techniques to assess the cellular traction forces have been developed (see reviews [14], [39], [40], [41], [42], [43], [44]). However, most of them are limited to computation of traction forces exerted by single, isolated cells.
Efforts at visualizing cellular traction forces may be traced back to 1980s when Harris et al. used thin polymeric silicone substrates for cell culture, and observed the wrinkling phenomena caused by the traction of migrating cells [45]. However, quantitative estimation of the traction from the wrinkling of silicone substrates is challenging due to the inherent non-linearity of the problem. From 1995 on, Lee, Jacobsen and Dembo et al., as well as other groups, developed several traction force microscopy techniques (TFM) to quantify the cellular traction produced by migrating or stationary cells on soft substrates [46], [47], [48], [49], [50], [51], [52], [53], [54]. TFM computes the cell traction forces from the deformation of a soft substrate with known elastic properties, such as polyacrylamide (PA) gel, on which cells are cultured. The deformation is measured from the displacements of micro-fluorescent markers embedded in the substrate. The motion is measured from two images. First image is taken with the cells adhered to the substrate. Here, the cells have generated traction force on the substrate, and the image gives the deformed configuration of the soft substrate. Then cells are removed from the substrate through trypsinzation, and a second image is taken. Subsequently, the substrate is relieved of cell traction, and the image shows the un-deformed configuration of the substrate. A comparison of the two images gives the displacement field of the substrate's top surface due to cell tractions. Digital image correlation method (DICM) is used to quantify the displacement field. The traction field is estimated from the displacement field. Several methods have been proposed for force estimation ranging from analytical methods, i.e. the Boussinesq formulation (either using Bayesian likelihood regularization method [51], [55] or Fourier transformed approach [49]), to computational methods like finite element analysis (FEA) [56]. The Boussinesq formulation approach, which assumes the substrate as a semi-infinite elastic half space [57], was first adopted by Dembo and Wang, et al., to compute the traction forces from the displacement fields followed by regularization [51], [55], [58], [59]. Since the Boussinesq formulation involves solving an inverse problem, the solution demands computational regularization schemes to predict the approximate traction solutions. Importantly, Butler, Trepat and Fredberg, et al. [49], [60], [61], [62], [63] made significant progress in mitigating some pitfalls of the regularization scheme by solving the Boussinesq equation using Fourier transform. Later Schwarz et al. introduced a new method to compute traction forces only at the focal adhesion site of the cell by assuming that the cell force transfer occurs only through these sites[50]. Some novel platforms, such as the photobleaching-activated monolayer with adhesive micro-patterns developed by Scrimgeour et al. [64] and the elastic substrates with micro-contact printing demonstrated by Stricker et al. [65], were also used to characterize the cell traction force. Furthermore, a FEA-based technique was also developed by Yang et al. to greatly improve the accuracy of traction force calculations [56]. The FEA method no longer depends on the Boussinesq formulation and thus is not limited by the semi-infinite elastic half space assumption [66], [67]. Recently additional contribution has been made in traction force computation in three dimensions [19], [68], [69], [70], [71], [72], [73]. 3D TFM techniques compute the 3D traction force fields from the cell induced 3D displacement and strain fields obtained using laser scanning confocal microscopy (LSCM) and digital volume correlation (DVC). However, it is challenging to obtain the Z-dimension displacement field and the technique can only be applied to single cell cases, rather than multiple cells or cell clusters.
The above studies focused on traction force computation for single cells far from their neighbors, i.e. cells that do not interact mechanically with each other. However, live cells do interact with their neighbors chemo-mechanically and form cell clusters [7], [29], [37], [74], [75], [76]. In this paper we present a novel finite-element-based TFM technique to compute the traction fields generated by multiple cells and clusters. We first present a theoretical proof showing that the 3D traction field computed from prescribed displacement field of the substrate is unique. We verify the uniqueness by considering a 2-cell case. We test the accuracy of the computational technique by applying a known force on PA gel substrate using a micro-needle, and by comparing the experimental force with the computed one. Finally, we compute the traction fields generated by multiple cancerous and fibroblast cell clusters, and reveal that cells might be under compression in such 2D clusters. We believe that the present technique may enable better examination and understanding of a variety of biological phenomena involving homotypic and heterotypic cells and cell cluster interactions [77], [78], [79].
Consider a 3D linear elastic solid with volume V in static equilibrium. Its boundary, S, consists of Su and Sσ (S = Su+Sσ) where displacements and traction are prescribed respectively.
Proposition: Given displacement field at Su and traction at , the corresponding traction at Su is unique. (Note: indices i, j = 1,2,3 correspond to x,y,z Cartesian coordinates respectively; all equations follow standard tensor notation and summation convention). Supporting material Text S1 presents the proof of the proposition.
We illustrate our computational scheme as follows. Consider two separate cells on a soft elastic substrate. The substrate is adhered to a rigid surface (such as glass) at the bottom. The lateral boundary of the substrate is far from the cells. In the finite element scheme, the substrate is modeled as a rectangular pyramidal solid body. It is discretized as a collection of small cubes with common nodes. We need to prescribe three boundary conditions, namely any combination of forces (Fx, Fy, Fz) and displacements (ux, uy, uz), at each of the surface nodes. For example, (Fx, uy, uz) can be a boundary condition at a surface node. To ensure that the body is at rest (no rigid body translation or rotation), at least two of the nodes are prescribed with ux, = uy, = uz = 0. Given the boundary conditions, finite element scheme calculates the deformation of the solid body such that the total energy is minimized. Thus the displacements at each node within the body, and at the surface nodes where forces are prescribed are evaluated. This leads to the evaluation of strains and stresses using the elastic properties of the solid (Young's modulus and Poisson's ratio for the isotropic gel). Surface traction is calculated from the stress near the surface and normal vector to the surface (), as shown in Supplementary Materials S1. Surface nodal forces are calculated from an area integral of traction at the vicinity of the node. Thus, the analysis provides the forces at nodes where displacement is prescribed, and displacements where forces are prescribed. If (Fx, uy, uz) is prescribed at a surface node for example, one gets (ux, Fy, Fz) at that node. Even though the solution is unique in principle, errors are introduced if the discretization is coarse. With finer discretization, the solution converges to the correct one. This convergence test is often employed to gage the accuracy of the solution.
In our problem with two cells, we prescribe zero displacement boundary conditions at the bottom surface and at the four vertical sides of the body (Fig. 2). Thus all the nodes on the bottom and the vertical sides are fixed. For simplicity of illustration, consider that there are a few nodes on the top free surface outside the cell boundary, and a few nodes within (Fig. 2). Our objective is to calculate the traction on these nodes. We can experimentally measure displacements (ux, uy, uz) at all the nodes on the surface. They are generated by cell forces, although we do not know the precise locations of these forces. We also know that the surface outside the cells has no traction, and that each cell or cell cluster produces a traction field that is self-equilibrated, i.e., the sum of forces applied by the cell or the cell cluster on the substrate is zero. Cell traction can be evaluated by prescribing either of the two boundary conditions:
Remarks. (1) The mixed boundary scheme applies exact boundary condition (zero force) at nodes outside the cells. Hence none of the displacements (ux, uy, uz) need to be prescribed at these nodes. Thus, it is not necessary to measure the displacements of the beads outside the cells. Due to the exact boundary conditions outside the cells, the traction solution is expected to be more accurate. However, errors will be introduced if the cell boundary is incorrectly defined and there are nodes that fall outside the cell boundary where cells apply traction. In cases where the cell boundaries cannot be identified due to imaging conditions (Fig. 3), displacements should be prescribed for regions nearby the cells.
(2) Displacement uz and Poisson's ratio: It is shown in the supplementary material (Supplementary materials text S3, Fig. S1b and c), that if the Poisson's ratio of the gel approaches 0.5, then the in-plane displacements, (uy, uz), on the surface of the gel are independent of the out-of-plane component of traction (Fz). That is, (ux, uy) are determined by (Fx, Fy) on the surface. Similarly, uz is determined by Fz on the surface only. Thus, in order to evaluate the in-plane traction only, one needs to measure and prescribe in-plane displacements only at the surface nodes, and prescribe arbitrary boundary condition in z direction (i.e Fz = 0 or uz = 0) at all surface nodes, when Poisson's ratio is close to 0.5. We experimentally measured the Poisson's ratio of our gel as 0.47±0.02 (Fig. S3b, n = 5). In order to estimate the in-plane traction only, we have prescribed Fz = 0 for all nodes within the cells in the rest of the paper. This results in an error of less than 2% in the calculation of in- plane forces Fx and Fy (Supplementary materials text S3 and Fig. S3b). If Fz is desired, one needs to measure and prescribe (ux, uy, uz) at the surface nodes. Also, if Poisson's ratio is much less than 0.5 (e.g., 0.35), (ux, uy, uz) must be prescribed at the nodes within the cells even when only in-plane traction is desired.
In this section, we demonstrate computationally that the traction solution from finite element simulation is unique as long as the full 3D boundary conditions are prescribed. We define two circular boundaries representing two cells with half-cell distance apart on a soft gel surface. The diameter of each boundary is chosen as 20 µm, close to real cell size. A three-dimensional finite-element (FEM) block model is generated (ANSYS 12.0 Workbench Package) to represent the PA gel substrate [79]–[98]. The gel is presumed linear elastic, isotropic, and homogeneous in their mechanical properties for a wide range of deformations [78], [99]. The Elastic modulus, E, of the gel is 1KPa (our experimental value is 1.05±0.17 kPa, measured by AFM indentation (n = 15; Fig. S3a), [99]–[101]]). The model height is 70 µm, same as the thickness of PA gel used in experiments. We first apply an in-plane force field (Fig. 4a) within each boundary, and compute the corresponding displacement field, ux, uy, uz (Fig. 4b). Second, we use the computed ux, uy and uz within the cell boundary on the surface (Fig. 4c), and zero-traction conditions outside the boundaries to calculate the traction within the cells (Fig. 5d). A comparison between the prescribed and the calculated forces from the two steps shows close quantitative agreement (within 1%) (Fig. 4e-f). Note that individual cells or cell clusters generate self-equilibrated traction on the substrate. Hence, we use a measure of accuracy of the traction solution by defining the error ratio, (2)where Fxi and Fyi, are the nodal force components within the individual cells, and i = 1, n, the number of nodes within the cell or cell cluster boundary. For exact solution, ε = 0.
In this section, we demonstrate the applicability of the method by evaluating the traction induced by two neighboring cells. Here, two monkey kidney fibroblasts were plated on PA gel (1 kPa) with Poisson's ratio of 0.47 (Fig. 5a). Two different regions (two sets of Su and Sσ) were selected to prescribe the displacement boundary conditions: (1) displacement field underneath the two cells were prescribed in the model (the white parts in Fig. 5b), whereas the traction-free condition was applied outside the cells (the black part in Fig. 5b); (2) the displacement field within a region enclosing both cells was prescribed (the white part in Fig. 5c), whereas the traction-free condition was applied outside this region (the black part in Fig. 5c). The out-of-plane force, Fz, was prescribed as zero within the cellular regions in (1) and (2). The traction fields were calculated for both cases (Fig. 5d,e,g,h), and compared (Fig. 5f and i). The RMS of node-by-node traction difference inside 2-cell region (superscripts indicate regions 1 and 2) was 21.7 Pa, which shows close match with only 5.1% of maximum traction inside the cells (426. 8 Pa).
In this section, we compare our mixed-boundary condition method with traditional whole-field displacement boundary condition method, which requires iterative calculation and has been successfully used by Fredberg, et al [49], [102]. Briefly, the iteration calculation proceeded as follows: (a) we assigned the complete 2D DICM (digital image correlation method) displacement data (ux,uy) for all nodes of the top surface of the gel (both intracellular and extracellular regions; Fig 6a-b). We prescribe Fz = 0 within the cluster for both the mixed boundary condition and iterative methods. (b) The traction field was solved using FEM. Then all the forces in the extracellular region were replaced by Fx = Fy = Fz = 0 to satisfy the traction-free condition, while the forces in the intracellular region were retained intact. (c) The new traction field was used to generate a new displacement field using FEM. Thus a new displacement field was computed within the intracellular region. (d) The computed intracellular displacement field was replaced with the DICM displacement field (ux and uy), while the computed extracellular ux, uy, and uz from previous step were retained intact. (e) The steps (b), (c), (d) were repeated until the solution converged, i.e., the difference between the root mean square (RMS) of surface nodal forces in two consecutive cycles became less than 5% (Fig. 6c-e).
Our computational results showed that the solutions from mixed-boundary and iterative methods converge (Fig. 6c-e). We found, the difference between the root mean square (RMS) value of traction from the two methods was 1.6×10−1 kPa (Fig. 6f), less than 3.8% of the maximum computed cell traction. The difference between the RMS of the nodal forces was 0.2 nN, which is 0.25% of the maximum nodal force at cell cluster - substrate interface (Fig. 6g). The distribution of traction |t| and forces at nodes (Fig. 6h-6j) shows good agreement between the two methods. We used ε (Eqn 2) as a measure of accuracy of the traction solution.
In FEM, convergence test is required to determine the optimal mesh size needed to obtain the accurate solution. Three mesh sizes, 3.23 µm, 4.84 µm, and 6.45 µm were tested, as shown in Fig. 7a-c, and used to calculate the traction field of the same cell cluster by mixed-boundary condition method. The distribution of nodal traction and forces showed minor difference between the three mesh sizes (Fig. 7a-c and 7e). The values of ε were 4.74%, 6.69%, and 6.12% for mesh size of 3.23 µm, 4.84 µm, and 6.45 µm respectively (Fig. 7d). Therefore, in the following computations, mesh size of 4.84 µm was used for analysis. The upper limit of mesh size is dependent on the specific cell size and the gradient of the traction field produced by the cell. A starting point on mesh size can be <20% of cells size.
A key attribute of the present method is the computation of traction fields generated by multiple cell clusters interacting with each other. Each cluster may consist of multiple cells, and the cluster size might be similar to or larger than the thickness of the soft substrate. Hence the effect of the glass-gel interface needs to be considered, and the gel may not be treated as half space. In the following, we study several cell clusters (Figs. 8–10) and outline the main biological findings. The mixed-boundary condition method was used to compute the traction fields.
The majority of fundamental physiological processes in tissue development, health, and disease are coordinated by the collective activities of multiple cells [60], [62], [76], [102], rather than single cells[10], [103]. To understand how mechanical traction applied by neighboring cell cluster groups could specify or mediate the tissue functionalities [7], [8], [11], [75], [104], [105], [106], robust cellular traction evaluation method is indispensable. In the present study, we developed a finite element element-based traction force microscopy (TFM) to accurately compute and visualize the traction maps resulting from multiple cell clusters. The uniqueness, convergence, and correctness of traction solutions are substantiated. We showed that as the gel Poisson's ratio >0.4, the in-plane traction can be obtained with minimal error from the in-plane displacement field alone. For Poisson's ratio <0.4, both in and out of plane traction depend on both in and out of plane displacement boundary conditions, and it is essential to measure these displacements to compute any of the traction components. The method presented is applicable to substrates with any value of the Poisson's ratio. It calculates the full 3D traction field given the 3D displacement boundary condition within cells or cell clusters. Moreover, unlike the classical TFM methods that are based on Boussinesq solutions [39], [40], [48], [49], the FEM takes into account the effect of substrate thickness and nearby environment. It is now known that cells can sense the substrate depth within the cellular length scales by showing distinct morphological variation on the gel substrate with same Elastic modulus but with varying thickness[22], [107].
We applied the method to compute the traction generated by multiple cell clusters. Some of the clusters were more than 100 µm in size consisting of many cells, while others were in close proximity to each other. The computational scheme presented here is ideal for studying such clusters, since the domain of traction field is much larger than the thickness of the gel, and one needs to account for the finite thickness of the substrate. A few interesting biological insights emerge from these analyses. First, the cluster may behave as a single contractile unit where the peripheral cells serve as anchorage sites. Force is transmitted between distant peripheries by the cells inside the cluster. Thus the cells are subjected to tensile intercellular forces, as if the peripheral cells are pulling the interior cells outward. It needs to be seen whether there are specific cells within the cluster that generate the force, or all the cells behave as contractile actuators. In any case, the cells probably use cell-cell junctions and cytoskeleton to transmit the force through the cluster.
We also found instances where traction is limited to small regions well within the clusters. These regions can have locally balanced traction (forming dipoles), leaving the rest of the clusters nearly traction free and weakly adhered to the substrate. These clusters are spherical in morphology, as expected. The traction free regions tend to minimize the surface area by being circular, just as a free-standing cell cluster takes a spherical shape. It is plausible that the cells within the circular clusters are under compression due to the surface tension of the peripheral cells. In any case, the interior traction maps can be highly dynamic. When cell clusters merge, the traction map can change their orientations, and the net force can increase by an order of magnitude over short times.
It is known that cells generate contractile forces. Thus, it is expected that the cells in a 2D cluster will be under intercellular tension. We found evidence to the contrary. If the cells are on soft substrates where they do not spread much, but they adhere to the substrate, then some of the cells in the cluster may be subjected to compression. We found regions within such clusters where the neighboring cells apply repulsive forces on the substrate, i.e., the cells are pushing against each other while being adhered to the substrate. One possible explanation might be that the neighboring cells are growing, but their adhesion sites are stationary.
In conclusion, we developed a robust FEM-based cell traction force microscopy technique to estimate the traction forces produced by multiple cells and clusters. The utility of the technique is exemplified by computing the traction force fields generated by multiple monkey kidney fibroblast (MKF) and pre-MLP human colon cell (HCT-8) clusters in close proximity. The developed technique is user-friendly and computationally inexpensive. Our FEM-based traction force microscopy provides a powerful tool to probe multi-cell questions involving assembly/disassembly dynamics of cell ensembles, tissue network formation, and wound healing. Future work is needed to determine the subcellular processes involved in mechano-sensing and regulation, and their respective timescales.
Polyacrylamide (PA) gel substrates with 1 kPa stiffness used in present study were made by mixing 12.83% (v/v) of acrylamide (Sigma-Aldrich, Inc.), 1.54% (v/v) of N, N-methylene-bisacrylamide (Sigma-Aldrich, Inc.), 2% (v/v) of 1 µm diameter fluorescent micro-beads (Invitrogen, Inc.) and 10 mM Hepes (Gibco., Inc.) [7], [11]. Solution was vortexed thoroughly for 5 min to obtain uniform distribution of beads. TEMED and ammonium persulfate (Fisher Scientific, Inc.) were used to initiate PA gel crosslinking. Chemical modification of glass slides and preparation of PA gels were carried out following the procedures described previously [50], [80], [81], [82], [83], [84], [85]. Briefly, a circular glass coverslip (Fisher Scientific, Inc.) of 1.2 cm in diameter was placed on an acrylamide solution drop on activated coverslip and placed on the bottom of a petri dish. Capillarity spreads the drop and fills the space between the circular coverslip and the activated coverslip. The gel was cured at room temperature and reached to the stabilized thickness of 70 µm [82], [85], [86]. The circular glass coverslip was peeled off from the gel that remained on the activated cover slip. The surfaces of the air dried PA gels were activated by incubating in 97% hydrazine hydrate (Acros Organics.) for 12 h followed by a complete rinsing with DI water and 30 minutes incubation along with gentle shaking in 5% acetic acid (Avantor Performance Material, Inc.) [7], [8], [11], [13], [81]. Solution of human fibronectin (25 µg/ml, BD Biosciences) was prepared by dissolving in phosphate buffer saline (PBS) and the carbohydrate groups of fibronectin were oxidized by sodium periodate (Sigma-Aldrich, Inc.). To minimize the displacement noise and rigid body motion during imaging, the glass slides was firmly adhered to the bottom of 30 mm petri dish using adhesive glue (Henker Consumer Adhesive, Inc.). Full experiment procedures and sample characterization are provided in Supporting Materials Text S4-S9 and Figures S1–S5.
|
10.1371/journal.ppat.1000793 | Tsetse EP Protein Protects the Fly Midgut from Trypanosome Establishment | African trypanosomes undergo a complex developmental process in their tsetse fly vector before transmission back to a vertebrate host. Typically, 90% of fly infections fail, most during initial establishment of the parasite in the fly midgut. The specific mechanism(s) underpinning this failure are unknown. We have previously shown that a Glossina-specific, immunoresponsive molecule, tsetse EP protein, is up regulated by the fly in response to gram-negative microbial challenge. Here we show by knockdown using RNA interference that this tsetse EP protein acts as a powerful antagonist of establishment in the fly midgut for both Trypanosoma brucei brucei and T. congolense. We demonstrate that this phenomenon exists in two species of tsetse, Glossina morsitans morsitans and G. palpalis palpalis, suggesting tsetse EP protein may be a major determinant of vector competence in all Glossina species. Tsetse EP protein levels also decline in response to starvation of the fly, providing a possible explanation for increased susceptibility of starved flies to trypanosome infection. As starvation is a common field event, this fact may be of considerable importance in the epidemiology of African trypanosomiasis.
| In Africa, tsetse flies transmit the trypanosomes causing the devastating diseases sleeping sickness in man and nagana in domesticated animals. These diseases are major causes of underdevelopment in Africa. Paradoxically, most, but not all, flies are resistant to infection with trypanosomes, but we do not have a clear picture of how flies fight off trypanosomes. Here we show that a particular, tsetse-specific immune responsive protein called tsetse EP acts as a powerful antagonist of trypanosome establishment in the fly midgut. It is known that starvation of flies leads to an increase in their susceptibility to trypanosomes and this may be a considerable factor in the epidemiology of the disease in Africa. Here we demonstrate that starvation leads to a decrease in tsetse EP levels, which may explain how starvation of the fly works to increase its susceptibility.
| African trypanosomes are protozoan parasites that cause sleeping sickness in humans and nagana in domestic livestock in sub-Saharan Africa. An epidemic involving several hundred thousand people that spread through Sudan, the Central African Republic, DRC and Angola in the 1990's, demonstrated how socially and economically devastating these diseases are [1]. Trypanosomes kill more than 3 million cattle annually and those animals that survive display low productivity due to the wasting effects of the disease [2]. The annual losses from trypanosomiasis in cattle amount to more than US $4.5 billion [3]. Trypanosomes, by influencing food production, natural resource utilization and the pattern of human settlement, are thus seen by the African Union as one of the greatest constraints to Africa's socio-economic development [4].
African trypanosomes are cyclically transmitted by tsetse flies (Glossina spp.). Trypanosoma brucei and T. congolense undergo a complex cycle of development in the tsetse beginning almost immediately after ingestion of an infected bloodmeal when trypanosome bloodstream forms (BSF) differentiate to the procyclic form in the fly midgut lumen [5],[6],[7]. For the first three days following infection all flies contain trypanosomes. Between days 4 and 5 trypanosome infections are eliminated from most flies [7] through a process we term self-cure. The identified factors that influence vector competence (the ability to transmit parasites) include the age of the fly, the number of bloodmeals taken and the activation of fly immune processes, with both antimicrobial (host defense) peptides [8], and lectins [9],[10],[11] implicated in parasite-vector interactions. More recently, antioxidants have been shown to increase fly susceptibility when administered to flies in an infective bloodmeal [12]. Most mature tsetse are resistant to trypanosome infection although the mechanisms involved in elimination of trypanosomes from the fly midgut (self-cure) are not understood [13].
As T. brucei BSF trypanosomes transform in the tsetse midgut the trypanosome surface coat changes from variant surface glycoproteins (VSG) to procyclins. At first the procyclins are a mixture of GPEET and EP forms and then expression of GPEET becomes repressed [14]. Our attention has been drawn to a fly protein called tsetse EP (accession number CAC86027), named for the extensive glutamic acid-proline dipeptide repeats that in Glossina morsitans morsitans comprise more than 40% of its length. The repeat section of this molecule shows remarkable sequence identity to the repeat section of the EP form of procyclin surface coat molecules of T. b. brucei [14]. These repeats are very rare in the protein databases and their co-incidence in two species showing such a close biological relationship is remarkable. Our knowledge of tsetse EP is limited although we do know that it is strongly up regulated following fly challenge with Gram-negative bacteria [15] suggesting a possible function in the insect immune response. In addition up regulation of the immune response by injection of E. coli also leads to a significant reduction in trypanosome prevalence [8],[16]. For these reasons we have undertaken a series of experiments to see if these observations are connected. We provide evidence that tsetse EP protein has a powerful role in protecting the tsetse fly midgut from trypanosome infection.
Tsetse midguts were carefully dissected into distinct structural regions (Figure 1, Panel A) to determine the location of tsetse EP mRNA and protein. Tsetse EP transcripts were detected in all sections tested. However, lower levels were consistently observed in the proventriculus (PV) (Figure 1, Panel B). The Western blot (Figure 1, Panel C) overlay of the nigrosine-stained PVDF with the autoluminogram revealed the strong presence of tsetse EP protein in all tissues except for the PV. Given the presence of tsetse EP transcript in the PV we conclude that tsetse EP protein was either not produced in PV, was rapidly turned over in that organ or was rapidly translocated from there into the anterior gut. Similarly, tsetse EP transcript was detected in salivary glands from teneral and fed flies [17] but tsetse EP protein was only weakly detected by immunoblotting suggesting it may be rapidly translocated to midgut.
Tsetse EP protein appears to be ubiquitous in Glossina spp. as its presence was confirmed in eight species of tsetse previously examined by Western blotting with the anti-EP repeat antibody (mAb 247) [15]. Using the available sequence analysis of the amino acid sequence of tsetse EP protein [15],[17] we designed effective double stranded RNA for knockdown experiments (Figure 2). A protein sequence comparison (87% similarity) between two species (G. m. morsitans and G. p. palpalis) revealed that the outstanding sequence difference was in the length of the C-terminal EP repeat region [15]. The tsetse EP protein is probably a preproprotein containing a short (19 mer) hydrophobic, N-terminal signal sequence as predicted by SignalP 3.0 [18]. Amino acids 20–48 appear to be removed from the remaining peptide during an undefined maturation process as determined by mass spectrometry and N-terminal sequencing [17],[19]. The EP rich domain is extremely hydrophilic, and thus almost certainly is highly soluble in aqueous solvents. It is interesting that all 8 of the cysteine residues are situated up stream of the EP rich C-terminus, suggesting that this region may be highly folded. For our experiments, we designed dsRNA to target in RNA interference the homologous region 23 residues downstream from the N-terminus of the mature protein (Figure 2, red highlighted region: GKFASDKCAQEGQ). The dsRNA target varies only slightly between G. m. morsitans and G. p. palpalis (4/39 nucleotides differ and these are shown in yellow lettering in Figure 2). Consequently the same dsRNA construct was used to achieve gene knockdown in both species.
During RNAi experiments mRNA levels are often extrapolated to predict protein expression levels. However, this is often misleading as the correlation between transcript abundance and protein expression levels can often vary as much as 30 fold or more, leading to a grossly distorted analysis of a biological system [20] and this may be especially true in the midgut of blood sucking insects where post-transcriptional regulation may be a common phenomenon [21]. Consequently, we measured tsetse EP levels at both the mRNA and protein levels. We show that injection of dsRNA leads to significant reductions in transcript levels compared to controls (Figure 3A). In addition, immunoblot analysis using the anti-EP repeat monoclonal antibody (mAb247) to detect the tsetse EP protein in midguts of knockdown flies showed complete elimination of the endogenous protein following a single injection of 4 or more µg of dsRNA (Figure 3B).
We employed a reverse genetics approach to determine if tsetse EP influences parasite establishment in the midgut of the fly. We injected double-stranded RNA (dsRNA) into the thoracic haemocoel of male flies of different ages. Typically the flies were allowed to recover for 36–48 h after injecting dsRNA. This provides enough time for the dsRNA to start silencing tsetse EP protein transcription and for endogenous protein levels to decline [22]. After this point, flies were offered an infective bloodmeal containing virulent strains of either T. b. brucei (TSW196) or T. congolense (1/148) BSF. Seven days after the infectious meal the midguts were dissected, examined microscopically, snap frozen, and the number of infections was recorded (Table 1). A complicating feature of this insect system is a natural decrease in susceptibility in older flies termed the teneral phenomenon. Typically more than 50% of flies establish midgut infections when fed trypanosomes in the first bloodmeal. However, if infected in the second bloodmeal, this susceptibility declines to ∼30% of the population. By the third bloodmeal, tsetse populations are predominantly refractory to infection with typical midgut establishment rates of 10% or less (Table 1). So, we investigated flies with differing feeding histories. Tsetse EP knockdown flies, infected at all feeding time points investigated, showed statistically significant increases in susceptibility to T. b. brucei establishment in the midgut when compared to the controls (Table 1).
To determine if this phenomenon was present in other tsetse species we also investigated G. p. palpalis. Table 1 shows there are statistically significant increases in T. b. brucei establishment in the midgut of tsetse EP knockdown G. p. palpalis. We also conducted experiments to determine if the phenomenon extending to other trypanosome species. T. congolense also establishes higher midgut infections in EP knockdown flies (Table 1). Based on our current and previous [15] observations, the increase of vector competence to midgut inhabiting trypanosomes in tsetse EP knockdown flies is possibly a genus-wide phenomenon in Glossina.
Male tsetse received 5 bloodmeals prior to starvation. Flies were killed at 24 h time points, up to 7 days after the last blood meal, and individual midguts were assayed by immunoblotting using an anti-EP antibody (mAb247) (Figure 4). After 3 days of starvation a clear decline in tsetse EP protein levels is evident (Figure 4, asterisks). Fat body atrophy was also apparent in these flies when viewed with a dissection microscope. Tsetse EP protein levels increase again in flies 24 hours after feeding following a previous starvation period of 7 days (Figure 4, lane 8). We have no data to show if the starvation-induced decrease in tsetse EP protein is specific or part of a general lowering of protein levels in the midgut in response to starvation.
Although RNA interference is an exquisite genetic technique to knockdown target genes, the success in achieving this post-transcriptional silencing appears to be gene-specific with variability due, in part, to the half-life of endogenous target protein and unexpected lethal secondary effects from depletion of gene specific product [23]. Our unpublished observations in Glossina reveal that, for some genes, transcript knockdown cannot be achieved regardless of the construct designed. This may relate to the lack of a spreading mechanism in Diptera and the difficulty of dsRNA reaching cells in complex organs [22]. We have previously shown and confirm here that thoracic injections of microgram amounts of specific dsRNA can effectively depress tsetse EP transcription in the tsetse midgut for up to 2 weeks [22]. Thus, the effect of persistent tsetse EP knockdown on trypanosome midgut establishment (7–10 day experiment) could be confidently measured by microscopic examination. The data we present here shows that a tsetse molecule, tsetse EP protein, plays a role in protecting the midgut from infection with trypanosomes.
Computer analysis of the translated protein sequences from both G. m. morsitans (CAC86027) and G. p. palpalis (AAL82540), using multiple alignment tools and protein prediction algorithms, revealed that these proteins are highly conserved [15]. Including its signal peptide, tsetse EP protein from G. m. morsitans has a mass of 35.7 kDa and appears to form dimers and trimers and potentially larger oligomeric aggregates within the fly [15],[17]. Apart from the EP sequence the tsetse EP protein has no currently defined protein domains [17]. However a possible clue to function may be suggested by the preliminary observation of weak agglutinating activity of the large molecular complex towards freshly collected, washed rabbit red blood cells, suggesting tsetse EP putatively has some lectin activity [17]. In addition it has been demonstrated that tsetse EP protein is strongly up regulated following immune stimulation with E. coli [15] providing good evidence that it is part of the immune response system. Given this it is interesting to note that the Imd immune regulatory pathway mainly responds to gram negative organisms [24] and the Imd pathway has been implicated in the response of dipterans to parasite infections [16],[25].
Although all species of tsetse studied to date express tsetse EP protein [15], orthologues are not found in the Anopheles, Aedes, Apis or Drosophila genomes. A search of non-redundant databases revealed only two eukaryotic protein hits (apart from the procyclins): gi|94390895 [Mus musculus] and gi|109464874 [Rattus norvegicus]. These hypothetical proteins contain significant continuous EP repeat regions: e.g. 115 dipeptide repeats, representing 75% coverage of the rat protein. Unfortunately, no further functional information is available for these proteins. Remarkably, extensive regions of EP repeats (also varying in length) are contained in several procyclins that form the surface coat of procyclic trypanosomes of the T. brucei group [14],[26],[27]. Given the scarcity of EP repeats in organisms the chances of this happening coincidentally in trypanosomes and tsetse flies seem remote.
To examine the possibility that the tsetse EP protein and the EP procyclins from T. b. brucei were involved in antigenic mimicry we investigated another trypanosome species that lacks EP procyclins. The procyclic coat of T. congolense contains no extensive dipeptidyl EP repeats although similar anionic motifs are present [28]. Despite the absence of EP repeats, establishment of T. congolense is similarly affected by tsetse EP protein knockdown (Table 1). Our experiments demonstrate that tsetse EP protein can partially protect against the midgut establishment of trypanosomes from both the Trypanozoon and Nannomonas group trypanosomes and thus, strictly sequence-specific interactions in tsetse and trypanosome are not likely at play.
To assess if trypanosome establishment is altered by tsetse EP gene knockdown in tsetse species other than our G. m. morsitans laboratory model, we tested our RNAi protocol on G. p. palpalis. Knockdown of tsetse EP protein in G. p. palpalis also led to an increase in midgut infections (Table 1), confirming that tsetse EP protein influences trypanosome midgut establishment in both of these major vectors of trypanosomiasis. Given that tsetse EP has been demonstrated in a wide variety of Glossina species [15] this data suggests it may be a genus wide phenomenon.
It has been demonstrated that up regulation of the immune response by injection of E. coli leads to a significant reduction in the ability of trypanosomes to establish in the tsetse midgut [8],[16]. We have already demonstrated that tsetse EP protein is strongly up regulated upon introduction of Gram-negative bacteria into the fly [15]. Our demonstration here that knockdown of tsetse EP leads to increased fly susceptibility suggests that upregulation of tsetse EP protein upon injection of E. coli may be one explanation for the subsequent decrease in the susceptibility of the fly to trypanosomes.
It is interesting to note that older flies in field populations show unexpectedly high levels of susceptibility compared to laboratory reared flies where susceptibility rapidly declines following eclosion [29],[30] (Table 1); the reasons remain unexplained. We have demonstrated here that starvation reduces tsetse EP levels in flies (Figure 4). It has already been demonstrated that starvation of mature flies results in an increase in parasite survival in the midgut [31],[32],[33]. Consequently, starvation, which is likely to be a common phenomenon in the field, could explain the differences in susceptibility seen between field and laboratory populations of flies. The observed reduction of tsetse EP protein expression and loss of parasite resistance upon starvation may have considerable epidemiological significance in African trypanosomiasis.
In summary, this paper provides direct evidence for a tsetse-specific midgut molecule (tsetse EP), which is an antagonist of trypanosome survival in the vector. RNAi-induced knockdown of the midgut-associated, immunoresponsive tsetse EP protein increased the frequency of trypanosome establishment in the fly midgut up to more than six fold. The precise mechanism by which tsetse EP protein influences the refractorial capacity of the midgut remains to be elucidated.
Tsetse (G. m. morsitans) were maintained in laboratory colony at the Liverpool School of Tropical Medicine (LSTM) at 26°C and 65–70% relative humidity. Glossina palpalis palpalis were supplied as puparia from the International Atomic Energy Agency (IAEA) Entomology Laboratories, Siebersdorf, Austria. Every 48 hours, male flies were fed horse blood through silicone membranes. For infectious bloodmeals blood stream forms (BSF) of Trypanosoma brucei brucei TSW196 MSUS/CI/78/TSW196 [CLONE A], which is a fully fly-transmissible clone and able to undergo genetic exchange [34], and T. congolense 1/148 (Lister 1/148; isolated from a Zebu ox, Dongo River, Nigeria, Godfrey, 1960) were added to sterile defibrinated horse blood (TCS Biosciences Ltd., Buckingham, UK). Typically 200 µL of mouse blood (containing 4×106 parasites) were diluted in 5 mL of horse blood. Flies were dissected 6 days after the infectious bloodmeal. Midguts were dissected in saline on a glass slide and infection status determined by searching 10 random fields by light microscopy (125× magnification).
Double stranded RNA was transcribed using a MEGAscript High Yield T7 Transcription kit (Ambion, Huntingdon, UK). tsetseEP templates were available as clones from the tsetse EST program [35]. A double stranded fragment of the ampicillin resistance gene (dsAMP) was generated using pBluescript II SK+ as template. Template DNA was removed from the transcription reaction by DNase treatment and dsRNA was purified using MEGAclear™ columns (Ambion) and eluted in nuclease free water. Eluates were concentrated in a Christ (Osterode, Germany) 2–18 rotational vacuum concentrator to approximately 5 µg per µL. Primers were designed with the 20 base core T7 promoter sequence at the 5′ end. Primer sequences used were: AmpT7A TAATACGACTCACTATAGGGTTGCCGGGAAGCTAGAGTAAGTA; AmpT7B TAATACGACTCACTATAGGGAACGCTGGTGAAAGTAAAAGATG; EPT7A TAATACGACTCACTATAGGGTTCTGGCAAACCCTCAAT; EPT7B TAATACGACTCACTATAGGGCTACGATAAATATGTCCCTCTAAT.
Borosilicate glass capillaries (2.00 mm outside diameter) were formed into a fine point using a needle puller (PC10; Narishige, Japan). To generate tsetse EP knockdowns, male flies were anaesthetized by chilling and intrathoracically injected with 10 µg (2 µL volume) of dsRNA buffered in nuclease-free water.
The primers used in semi-quantitative RT-PCR reactions for determination of transcript abundance in tsetse tissues were: Gm GAPDHA CTCAGCTTCTGTGCGTTG (Tm°C 67); Gm GAPDHB AGAGTGCCACCTACGATG (Tm°C 67); GmmEPA ACCGTTCGTTCGCTTTACTAC (Tm°C 47); GmmEPB ACCCGCAGCCGTTTGACTTTC (Tm°C 51).
Total RNA was extracted from individual tissues using Trizol (Invitrogen, Paisley UK) and treated with RQ1 RNase-Free DNase. RNA was quantified using a Nanodrop ND-1000 (Wilmington, DE) spectrophotometer. A Promega Access RT-PCR System (Promega, Southampton, UK) was used for amplification of transcripts. G. m. morsitans GAPDH (Accession number DQ016434) was used to normalize samples. PCR cycling conditions were: 48°C for 45 minutes, 94°C for 2 minutes, followed by 30 cycles of 94°C for 30 seconds, 57°C for 1 minute, 68°C for 2 minutes and a final extension of 68°C for 7 minutes. TsetseEP gives a product of a larger size when genomic DNA (indicative of a putative intron) was used as template (approximately 365 vs 315 bp. respectively) and was used to ensure genomic DNA was removed from experimental templates.
Immunoblotting using Hybond™-P polyvinylidene difluoride (PVDF) transfer membrane (Amersham Biosciences, Amersham, UK) was performed as previously described [36]. In brief, the primary antibodies used were either a 1∶20 dilution of anti-EP repeat mouse mAb TRBP1/247 [37]. The secondary (detecting) antibody was a 1∶50,000 dilution of horseradish peroxidase conjugated goat anti-mouse IgG/IgM (H+L) (Caltag Laboratories, South San Francisco, CA). Kodak Biomax MR film (Eastman Kodak Company, Rochester, NY) was used to detect chemiluminescence. After development of the autoluminograms, proteins were stained on the PVDF membrane with 0.2% (w/v) nigrosine in PBS. The exposed film was superimposed on the stained PVDF membrane to reveal the precise location of the immunoreactive protein bands in relationship to the entire protein profile and to ensure equivalent protein loading.
Tsetse, which had fed twice, were injected on day 4 post emergence with 7 µg of gene-specific dsRNA (2 µl injection volume). Flies in the control group were injected with nuclease free water. Injected flies were fed again on day 5 and midguts dissected on day 7 were snap frozen in liquid nitrogen in pools of 5. The NorthernMax® formaldehyde-based system for Northern Blots (Ambion) was used. Total RNA (20 µg per lane) was loaded on a 1% formaldehyde-agarose gel. The Strip-EZ™ PCR probe synthesis and removal kit (Ambion) was used to synthesize single stranded DNA probes, which were labeled with [α32P] dATP (MP Biomedicals, Stretton Distributors, UK). Membranes were hybridized overnight at 42°C, given 2×5 minute low stringency washes and 2×15 minute high stringency washes before exposure to Kodak BioMax MR film.
|
10.1371/journal.pgen.1003637 | The Cohesion Protein SOLO Associates with SMC1 and Is Required for Synapsis, Recombination, Homolog Bias and Cohesion and Pairing of Centromeres in Drosophila Meiosis | Cohesion between sister chromatids is mediated by cohesin and is essential for proper meiotic segregation of both sister chromatids and homologs. solo encodes a Drosophila meiosis-specific cohesion protein with no apparent sequence homology to cohesins that is required in male meiosis for centromere cohesion, proper orientation of sister centromeres and centromere enrichment of the cohesin subunit SMC1. In this study, we show that solo is involved in multiple aspects of meiosis in female Drosophila. Null mutations in solo caused the following phenotypes: 1) high frequencies of homolog and sister chromatid nondisjunction (NDJ) and sharply reduced frequencies of homolog exchange; 2) reduced transmission of a ring-X chromosome, an indicator of elevated frequencies of sister chromatid exchange (SCE); 3) premature loss of centromere pairing and cohesion during prophase I, as indicated by elevated foci counts of the centromere protein CID; 4) instability of the lateral elements (LE)s and central regions of synaptonemal complexes (SCs), as indicated by fragmented and spotty staining of the chromosome core/LE component SMC1 and the transverse filament protein C(3)G, respectively, at all stages of pachytene. SOLO and SMC1 are both enriched on centromeres throughout prophase I, co-align along the lateral elements of SCs and reciprocally co-immunoprecipitate from ovarian protein extracts. Our studies demonstrate that SOLO is closely associated with meiotic cohesin and required both for enrichment of cohesin on centromeres and stable assembly of cohesin into chromosome cores. These events underlie and are required for stable cohesion of centromeres, synapsis of homologous chromosomes, and a recombination mechanism that suppresses SCE to preferentially generate homolog crossovers (homolog bias). We propose that SOLO is a subunit of a specialized meiotic cohesin complex that mediates both centromeric and axial arm cohesion and promotes homolog bias as a component of chromosome cores.
| Sexual reproduction entails an intricate 2-step division called meiosis in which homologous chromosomes and sister chromatids are sequentially segregated to yield gametes (eggs and sperm) with exactly one copy of each chromosome. The Drosophila meiosis protein SOLO is essential for cohesion between sister chromatids. SOLO localizes to centromeres throughout meiosis where it collaborates with the conserved cohesin complex to enable sister centromeres to orient properly – to the same pole during the first division and to opposite poles during the second division. In solo mutants, sister chromatids become disconnected early in meiosis and segregate randomly through both meiotic divisions generating gametes with random (and mostly wrong) numbers of chromosomes. In this study we show that SOLO also localizes to chromosome arms where it is required to construct stable synaptonemal complexes that connect homologs while they recombine. In addition, SOLO is required to prevent crossovers between sister chromatids, as only homolog crossovers are useful for forming the interhomolog connections (chiasmata) needed for homolog segregation. SOLO collaborates with cohesin for these tasks as well. We propose that SOLO is a subunit of a specialized meiotic cohesin complex and a multi-purpose cohesion protein that regulates several meiotic processes needed for proper chromosome segregation.
| Meiosis is a specialized type of cell division that generates haploid gametes from diploid germ cells. It encompasses a single round of DNA replication followed by two rounds of chromosome division in which first homologous chromosomes then sister chromatids segregate. During prophase of the first division (prophase I), homologous chromosomes pair, synapse and recombine with their partners. The resulting crossovers, stabilized by cohesion between sister chromatid arms, serve as chromatin linkers known as “chiasmata” that enable homolog pairs to bi-orient on the first division spindle. At anaphase I, resolution of sister chromatid arm cohesion leads to homolog segregation. Sister chromatids remain attached at their centromere regions until anaphase II, when resolution of centromere cohesion allows them to segregate [1]–[6].
Cohesion between sister chromatids is essential for several key steps in meiotic segregation and is mediated by ring-shaped cohesin complexes that embrace sister chromatid pairs [2], [7]. The subunits of cohesin are two SMC (structural maintenance of chromosomes) proteins, SMC1 and SMC3, and two non-SMC subunits, a “kleisin” subunit, which can be either the mitotic SCC1/RAD21 protein or its meiosis-specific paralog REC8, and a SCC3/SA-family subunit. SMC1 and SMC3 are long intramolecular coiled-coil proteins that form extended hairpin structures with N- and C-terminal globular ATPase domains at one end and a globular hinge domain at the other. SMC1 and SMC3 bind to each other at their hinge domains and to opposite ends of the kleisin subunit at their ATPase domains, forming a tripartite ring that embraces pairs of sister chromatids. The SA subunit binds to the kleisin subunit and regulates cohesin chromosome binding. Cohesin is loaded on chromatin prior to or during S phase and establishes cohesion during DNA replication. Although cohesin can be removed by other means and at other times in the cell cycle, cleavage of RAD21 or REC8 by the protease Separase at anaphase leads to release of sister chromatids and triggers segregation [7]–[10].
In meiosis, cohesion has a dual role, to keep homologs connected by stabilizing chiasmata on chromosome arms until anaphase I, and to keep sister chromatids connected at their centromere regions until anaphase II. The same cohesin complex, REC8 cohesin, is responsible for both arm and centromere cohesion and the same protease, Separase, is responsible for cleaving both arm cohesin at anaphase I and centromere cohesin at anaphase II. Since cohesin must be loaded prior to the first division, the centromeric cohesin complexes require protection from cleavage during anaphase I. This function is carried out by the centromeric guardian protein Shugoshin and its effectors (including the PP2A phosphatase) [11], [12]. REC8 and Shugoshin and the two-step cohesin release mechanism appear to be widely conserved [2], [7], [12]. REC8 and other cohesins are also required for several other essential steps during the first meiotic division, including homolog pairing, synapsis and recombination [2], [8], [9]. However, it is not clear to what degree these roles involve cohesion. In yeast and C. elegans, mutations in rec8 and smc3 can disrupt recombination, DSB formation and DSB repair without affecting cohesion [13], [14].
Another crucial meiosis-specific centromere modification, mono-orientation, is needed at the first division to prevent sister centromeres from connecting to opposite poles (bi-orienting) as they do at all other divisions. Instead, sister centromeres must collaborate in forming a single microtubule-binding surface and orient toward the same pole (mono-orient) so that their counterparts on the opposite homolog can orient to the opposite pole. This coordinated orientation of centromeres is essential to ensure that they segregate reductionally, with both sisters co-segregating to the same pole during the first meiotic division, rather than equationally as in mitosis or the second meiotic division. The mono-orientation process is not well understood. In S. cerevisiae, mono-orientation is mediated by a specialized Monopolin complex that clamps sister centromeres together, and a different specialized monopolin protein Moa1 is required for mono-orientation in S. pombe. However, these yeast proteins are not conserved. In several higher eukaryotes including C. elegans and Arabidopsis, cohesin is required for mono-orientation but what role it plays is not known [15]–[19].
Proper homolog segregation requires recombination to generate the crossovers that serve as chiasmata. Meiotic recombination is initiated by programmed double-strand breaks (DSBs) induced by the conserved Spo11 endonuclease [20]. Breaks are then repaired by a meiosis-specific version of the ubiquitous homologous recombination pathway modified to ensure that the repair products include adequate numbers of homolog crossovers (at least one per chromosome pair) [21], [22]. A crucial modification, known as “homolog bias”, involves preferential use of homologous over sister chromatids as repair templates, a reversal of the sister chromatid bias that prevails in somatic DSB repair [23], [24]. Understanding of the mechanism of homolog bias is rudimentary but studies in yeast have identified two groups of proteins that play key roles: the meiosis-specific recombinase DMC1, a paralog of RecA and RAD51, which preferentially mediates invasion of homologous rather than sister strands [25], [26]; and the SC proteins RED1, MEK1 and HOP1 that seem to function mainly by inhibiting sister chromatid exchange (SCE) [23], [24], [27]–[29]. The few proteins outside of yeast that have been identified as being important for homolog bias, including ORD in Drosophila, HIM-3 in C. elegans, and SYCP-2 and SYCP-3 in mammals are also SC proteins, pointing to a possible conserved function of the SC in homolog bias [11], [30]–[32].
Either before or coincident with the early stages of meiotic recombination (depending on organism), homologs pair and “synapse”, a process that culminates in assembly of a tripartite structure called synaptonemal complex (SC) [5]. SC consists of two parallel lateral elements (LEs) that encompass the axes of the homologs, connected by densely packed transverse filaments that span a central region of about 100 nm, and a central element that lies parallel to and midway between the LEs. Transverse filaments are homo-dimeric coiled-coil proteins that bind to each other at their N-termini and to the LEs at their C-termini [33], [34]. In many eukaryotes, the LEs are clearly visible prior to synapsis when they are called axial elements (AEs), but in Drosophila no AEs have been observed. Instead, the LEs and central regions of the SCs assemble simultaneously during synapsis [5], [6].
Synapsis initiates during zygotene as short stretches of SC assembled at axial association sites, accompanied or preceded (depending on species) by alignment of homologs [5], [6]. In some eukaryotes, axial association sites correspond to DSB sites where the early stages of interhomolog recombination take place [6]. However, in Drosophila, DSBs are delayed until pachytene when homologs are fully synapsed, and synapsis is initiated and completed independent of the recombination apparatus [35], [36]. The initial SC patches are extended by a poorly understood process that leads eventually, at pachytene, to fully aligned and synapsed homolog pairs. Recombination is thought to be completed during pachytene and after it is complete, the SCs are disassembled and homologs disassociate except at chiasmata, which keep them connected throughout the first division. [5], [6].
AE/LEs are prominent, meiosis-specific versions of chromosome axes that develop in early prophase I [5], [6]. They encompass the paired sister chromatid axes that anchor the chromatin loops and are built on a condensed “chromosome core” of densely packed cohesin complexes that serves as a scaffold for assembly of additional meiosis-specific AE/LE proteins that promote homolog interactions, mostly by mechanisms that remain to be defined [37], [38]. The best understood AE/LE proteins are RED1 and HOP1, mentioned above as yeast proteins involved in homolog bias. RED1 is also required for synapsis and SC formation but some other AE/LE proteins are dispensable for SC formation although they are often required to stabilize chromosome cores and SCs [27], [30]–[32], [39]–[43]. In mammals and Drosophila, homologous chromosome cores can synapse with each other in the absence of the non-cohesin AE/LE components although the resulting SCs tend to be unstable and to disassemble prematurely [37], [38], [43]. Many eukaryotes have additional meiosis-specific kleisin family members or other cohesin paralogs and many of these are found primarily or exclusively in cores [40], [44]. One such paralog is C(2)M, a kleisin family member in Drosophila that is present only during prophase I in cores and is required for LE assembly, synapsis and normal levels of recombination but is dispensable for cohesion [45], [46]. Thus current evidence points to a fundamental role of the cohesin-based chromosome cores in synapsis and SC structure. However, although cores are cohesin-based, the role of cohesion in chromosome core and SC assembly remains to be clarified.
Cohesion is essential for chromosome segregation in Drosophila meiosis as well, but the way in which cohesion is mediated appears to differ from most other eukaryotes. No true REC8 homolog has been identified. The aforementioned C(2)M is the only known meiosis-specific kleisin, but its role is much more specialized than REC8. It is an essential component of the chromosome cores and required for synapsis and recombination but it is not enriched at centromeres and has no apparent role in either arm or centromere cohesion [45], [46]. Orientation Disruptor (ORD) is a cohesion protein that seems to carry out many of the functions of REC8 but it is not, on the basis of primary sequence homology, a cohesin. ORD localizes to centromeres and is required for centromere cohesion in both male and female meiosis. ORD also localizes to LEs and although not required for assembly of LEs or SCs, it is required to prevent their premature fragmentation and dissolution. Finally, ORD is required for normal levels of homolog recombination and is the only Drosophila protein known to suppress SCE. Although not a cohesin by sequence homology, ORD localizes along with the SMC cohesin subunits both at centromeres and on LEs and likely carries out some or most of its functions in collaboration with cohesin. The case is particularly clear for centromere cohesion where ord mutations lead to depletion of centromeric SMC cohesins in both male and female meiosis [30], [43], [47]–[52].
We have recently described a second meiosis-specific Drosophila cohesion protein, SOLO [53]. SOLO is required for centromere cohesion in Drosophila male meiosis and its loss leads to failure of mono-orientation and random chromatid assortment. SOLO and SMC1 are both enriched near centromeres throughout meiosis until both proteins disappear at anaphase II. In a mei-S332 (Shugoshin) mutant [11], both SMC1 and SOLO dissociate from centromeres simultaneously at anaphase I. In solo mutants, like ord mutants, centromeric SMC1 foci are absent at all stages of meiosis. Together these data indicate that SOLO functions in very close collaboration with the SMC1 cohesin subunit. However, like ORD, SOLO shows no sequence homology with cohesins, or with any other proteins in the database [53].
The previous study was limited to male meiosis in which homologs segregate by a unique mechanism that does not involve SCs, recombination or chiasmata. Instead a specialized conjunction complex holds homologs together in place of chiasmata [54]. SOLO is not required for any step in homolog segregation in males except for centromere mono-orientation [53]. In this paper we describe the roles of SOLO in Drosophila female meiosis and show that SOLO, like ORD, carries out a broad spectrum of meiotic functions that include cohesion, pairing and clustering of centromeres, regulation of chromatid orientation and segregation at both meiotic divisions, stable assembly of LEs and SCs, achievement of normal levels of homolog exchange, and suppression of sister chromatid exchange. We also show that SOLO and SMC1 reciprocally co-immunoprecipitate from ovarian protein extracts, further underlining the close cooperation between SOLO and cohesin. The very similar mutant phenotypes and lack of synergism between solo and ord mutations suggest that SOLO and ORD function together with cohesin in the same molecular processes. Overall, our data indicate that SOLO has essential roles in centromere cohesion, AE/LE stability and recombination. SOLO joins ORD as the second such protein to be identified in Drosophila. Analysis of the multiple functions of SOLO in meiosis should further insight into the roles of cohesion in meiotic segregation.
Errors in meiotic chromosome segregation, referred to here as nondisjunction (NDJ), generate aneuploid gametes that can be detected and quantified in genetic crosses. X chromosome NDJ generates diplo-X and nullo-X eggs that yield distinctive progeny classes (matriclinous daughters and patriclinous sons) (Figure S1) in standard crosses. X NDJ frequencies were found to be highly elevated in females hemizygous for three different solo alleles, averaging 58.4% compared to 0% in the sibling wild-type (WT) control crosses (Table 1). Because the X chromosomes carried markers adjacent to and flanking the centromeres, the progeny that developed from diplo-X eggs could be analyzed for whether both X centromeres came from a pair of sister chromatids (referred to as sister chromatid (S) NDJ) or from homologous chromatids (referred to as homolog (H) NDJ). The relative frequencies of S and H NDJ were similar for the three alleles, averaging approximately 21% S NDJ. This figure may underestimate %S because of reduced viability of the homozygous S NDJ classes relative to the heterozygous H NDJ class.
NDJ of the autosomal 2nd chromosome pair was also assayed (Table 2). Because of the inviability of 2nd chromosome aneuploids, progeny derived from NDJ gametes are not recovered in crosses to chromosomally normal males. However, by crossing females to males carrying an attached-2 chromosome (C(2)EN), which generate only diplo-2 and nullo-2 sperm, NDJ eggs can be recovered when fertilized by reciprocally aneuploid sperm. This assay allows detection of NDJ but does not permit calculation of a NDJ frequency as no regular gametes are recovered. Crosses of solo females to C(2)EN males yielded 2.5 and 3.0 progeny/female for two different alleles, indicating the occurrence of chromosome 2 NDJ. Heterozygous solo/+ controls yielded no progeny in similar crosses. Since two maternal 2nd chromosomes were recovered in half of the progeny, the relative frequencies of S and H NDJ could be measured. After correcting for viability differences, %S NDJ was estimated to be 32%, very near the expected frequency (33.3%) if chromatids segregate randomly at both meiotic divisions. These results indicate that solo causes NDJ of both sex chromosomes and autosomes and suggest that the NDJ mechanism might involve random chromatid assortment.
Crossover frequencies were measured in three euchromatic intervals, two (pn-m and m-f) that together encompass 80–85% of the recombinational length of the X chromosome and one (cn-bw) that encompasses about 90% of chromosome arm 2R, and in one mixed euchromatic/heterochromatic interval (f-y+) on the X chromosome (see Figures S1 and S2). For the X chromosome, exchange was measured in females hemizygous for each of the three solo alleles, using heterozygous (solo/+) siblings as controls to minimize background variation (Table 3). The chromosome 2 crosses were conducted similarly except that a null allele was used in place of the Df chromosome (Table 4). As the results for the three alleles did not differ significantly in either set of crosses for any of the intervals, combined results are also presented. Crossover frequencies decreased in all four intervals in the mutants, very substantially and uniformly (7.5- to 7.6-fold) in the three euchromatic intervals, and more moderately (26%) in the f-y+ interval that encompasses the X centromere.
The 7.6-fold reduction in crossovers between the distal (pn) and proximal (f) euchromatic X markers in our experiments falls within the fairly wide range of reported results for strong alleles of ord (6 to 20-fold reductions) and are in reasonable agreement with the reported 6.1-fold reduction for an ord-null genotype [47]–[49]. Based on very limited data, both solo and ord mutants cause similar reductions (6 to 10-fold) in frequencies of crossovers in euchromatic autosomal intervals as well (Table 4) [47], but have much weaker effects on exchange in intervals near or encompassing centromeres [47]–[49]. However, existing data do not reveal whether ord and solo function independently of each other in controlling exchange. To determine whether a solo ord double mutant would reduce exchange any further, we generated females that were trans-heterozygous for null alleles of both genes. Crossover frequencies in the X euchromatin (pn-f interval) were reduced 6.5-fold in the double mutants relative to solo ord/+ sibling controls (Table 3), a fold-reduction value intermediate between those of ord or solo single mutants. This result suggests that solo and ord function in the same recombination pathway, one that controls about 85–90% of crossovers along the X euchromatin and probably in autosomal euchromatin as well.
One way solo might function to promote homolog crossovers is by preventing recombination intermediates from being repaired by SCE. If so, solo mutations should increase SCE. Crossovers between sister chromatids cannot be detected in conventional recombination assays, but single (or other odd number of) crossovers between the chromatids of a circular (or “ring”) chromosome, generate double-ring dicentric chromosomes. In Drosophila females, the dicentrics generated by exchange between sister chromatids of a ring-X chromosome become trapped in unresolved bridges on the anaphase II spindle and are not transmitted. Since exchanges between sister chromatids of normal “rod” chromosomes have no consequence, the ratio of ring-X recovery to rod-X recovery among progeny of a ring-X/rod-X heterozygote is a rough measure of the SCE frequency. In previous studies, the ring-X/rod-X recovery ratio in WT control females ranged between 0.7 and 0.9 [30], [35], [45], [55], [56]. This likely reflects the normal background activity of the SCE pathway since in the absence of DSBs (i.e. in a mei-P22 mutant), the ring-X chromosome is transmitted as efficiently as the rod-X [35], [57]. These results also show that the meiotic apparatus in Drosophila can transmit ring chromosomes efficiently as long as they are not dicentric. Several meiotic mutants have been analyzed by this assay but to date, mutations in only one gene, ord, have significantly reduced ring-X recovery [30], [35], [45], [55], [56].
To estimate meiotic SCE frequencies in soloZ2-0198 and soloZ2-3534 females, we measured the ring/rod recovery ratio in progeny of solo/Df or +/+ females heterozygous for the ring-X chromosome Ring(1)2 (R(1)2). The ring/rod recovery ratios were 0.83 in the WT controls but only 0.35 and 0.36 in the solo crosses (Table 5). This result indicates that roughly 65 out of every 100 ring-X chromosomes were eliminated in solo meiosis. These results may actually underestimate the frequency of SCE because double ring-X crossovers, which might be quite frequent in solo mutants, yield normal mono-centric ring chromosomes which would not be detected in this assay. We conclude that solo mutations dramatically upregulate the SCE pathway, reversing the normal homolog bias to a sister bias.
The recovery of both S and H NDJ progeny suggested that sister chromatid cohesion might be lost prior to the first meiotic division, as in solo males [53]. To test this idea, we used an antibody against Centromere IDentifier (CID), a centromere-specific histone H3 variant [58], [59] to examine centromere behavior during the first meiotic division in WT and solo ovaries (Figures 1B and 1C). The maximum number of CID spots during the first meiotic division would be 16 if all centromeres were separate. However, sister chromatid cohesion and homolog alignment, which are essentially complete in all WT pachytene nuclei, reduce the expected number of CID spots to a maximum of four. Moreover since non-homologous centromeres tend to cluster in prophase I, observed numbers are usually even fewer [43], [60], [61]. As expected, in WT ovarioles, C(3)G-positive nuclei from both region 2a germaria (early-mid-pachytene) and stage 5–7 egg chambers (late pachytene) exhibited 1–4 CID foci, averaging 2.3 at both stages (Figures 1B, 1D and 1E). In contrast, CID signals were much more numerous in solo pro-oocytes at all stages. In solo germaria only about 10% of pro-oocytes exhibited 4 or fewer spots, the remainder exhibiting 5–8 (mean = 6.3 (Figures 1C and 1D)). This suggests that both homologous centromere pairing and centromere clustering were disrupted by early-mid pachytene in solo mutants but that sister chromatid cohesion remained intact at this stage. However, by late pachytene (stage 5–7 egg chambers) more than half of the oocyte nuclei from solo ovaries exhibited more than 8 CID spots (Figures 1C and 1E), while the remainder exhibited 5–8 spots (mean = 8.5). Thus, in most oocyte nuclei, some sister centromere pairs had separated prematurely by the latter stages of pachytene. Very similar results were reported for an ord mutant [61]. Since prematurely separated sister centromeres are unlikely to establish mono-orientation on the spindle of the first meiotic division, these results may help explain the NDJ data.
To explore the expression pattern of SOLO in the female germline, we made use of two different transgenes expressing full-length SOLO cDNAs tagged with the enhanced yellow-fluorescent protein Venus. UPS-SOLO::Venus (UPS-SOLO) is driven by native regulatory sequences carried in a 2.7 Kb fragment of upstream genomic DNA. UASp-Venus::SOLO (UAS-SOLO) is controlled by GAL4-responsive UAS sequences [53]. Both transgenes were able to complement the NDJ phenotype of a null solo allele but the UAS-SOLO construct did so more robustly (Table S1). A single copy of the UAS-SOLO transgene, when expressed under control of the germline-specific driver nos-GAL4::VP16 in a solo background, fully suppressed X chromosome NDJ. However, solo females carrying two to four copies of UPS-SOLO still underwent NDJ at modest but significant frequencies (7–11%). This difference cannot be explained by the location of the Venus tag because the C-terminally tagged SOLO protein completely rescued NDJ when expressed under control of nos-GAL4::VP16 (Table S1) so may reflect a deficiency in expression level or pattern.
In whole-mount ovarioles prepared from females lacking any functional copies of native solo, UPS-SOLO and UAS-SOLO exhibited overlapping but non-identical localization patterns (Figures 2A and 2D). Both proteins were expressed only in germ cells and in all regions of the germarium except for the anteriormost segment of region 1. The only really striking difference between the UPS-SOLO and UAS-SOLO expression patterns in whole-mount preparations was the considerably higher level of UAS-SOLO expression in a broad anterior domain that encompassed most of region 1 (except for the anterior tip) and anterior region 2a. As this domain coincides with the domain of highest expression of nos-GAL4, this is probably an ectopic over-expression effect.
In nearly all germ cells, both UAS-SOLO and UPS-SOLO exhibited small numbers of prominent bright nuclear foci and a broad diffuse pattern that appeared to encompass both cytoplasm and nucleus. In addition, some nuclei exhibited much fainter fibrillar or linear staining (discussed below). A distinctive aspect of the bright focal and diffuse staining patterns was the uniformity of expression level within cysts, indicating strong expression in both nurse and meiotic cells. Similar expression patterns were previously reported for ORD and the SMC cohesins [30], [43].
Bright foci of both UPS-SOLO and UAS-SOLO were observed in all germ cell nuclei in regions 2a, 2b and 3 of germaria and in egg chambers through at least stage 5. (UAS-SOLO signals have been detected as late as stage 8 (data not shown)). Fainter foci were also seen in some pre-meiotic nuclei in the posterior half of region 1. Most nuclei exhibited one to four SOLO foci per nucleus, suggesting that the foci may correspond to centromeres. This idea was tested by staining UPS-SOLO-expressing ovarioles with an antibody against CID. As shown in Figures 2B and 2C for germarial region 2b and a stage 5 egg chamber, all of the bright UPS-SOLO foci aligned with anti-CID signals, confirming that SOLO is enriched in the vicinity of centromeres in female germ cells. However, at higher magnification, the overlap between UPS-SOLO and CID foci sometimes appeared only partial (Figure 2C, inset) suggesting that SOLO may be enriched at pericentromeric domains as well as centromeric domains. UAS-SOLO foci aligned with anti-CID foci as well (data not shown).
SMC1 and SMC3 have been shown to be highly enriched on centromeres of female germ cells at similar stages [43]. To confirm co-enrichment of SOLO and SMC1 in females, we stained germaria expressing UAS-SOLO with an antibody against SMC1. As expected, the SMC1 signals formed bright nuclear foci throughout the germarium from posterior region 1 through region 3 in both meiotic cells and nurse cells (Figure 2D). As reported previously [43], and like SOLO, SMC1 signals were absent from the anterior tip of the germarium where germ line stem cells and cystoblasts reside. It is evident from Figure 2D that the bright SMC1 and UAS-SOLO foci overlap very extensively in germaria. They also overlap in later stages (data not shown). Thus, SOLO and SMC1 are co-enriched on meiotic centromeres in females as well as in males.
To test whether the centromeric SMC1 foci depend on solo, WT and solo germaria were stained with anti-SMC1 antibody. Whereas prominent SMC1 foci were present throughout the WT germarium, no SMC1 foci were detected in any nucleus in the solo germarium (Figure 2E). SMC1 foci were also absent from solo oocyte nuclei in later stages (data not shown). However, SMC1 staining did not disappear in solo germ cells. Diffuse staining was apparent in many germ cells in both WT and solo germaria, and appeared to be associated with chromosome arms (Figure 2E, arrowheads, insets). This staining pattern is explored further below. Thus, in female meiosis as in male meiosis, enrichment of the SMC1 subunit of cohesin at centromere regions is dependent on solo. However, SMC1 can localize to chromosome arms in the absence of solo.
The findings that SOLO and SMC1 are co-enriched on centromeres and that SOLO is required for SMC1 localization to centromeres suggest that they may interact physically. In order to address this issue, we generated transgenic flies that express a full-length SOLO cDNA with tandem 3XFLAG and 3XHA tags at its N-terminus regulated by UAS sequences. One copy of this transgene completely reverted the NDJ phenotypes of solo males and females (Table S1) when induced by the germline-specific driver nos-GAL4::VP16, indicating that the FH::SOLO fusion protein is fully functional. Western blots revealed high level expression of FH::SOLO in ovaries. The absence of signal in the lane derived from y w (control stock lacking transgene) ovary extracts confirms the specificity of the anti-FLAG antibody (Figure 3A). In the co-immunoprecipitation experiment, FH::SOLO was pulled down from extracts of transgenic ovaries by the anti-SMC1 antibody (Figure 3B) used in immunofluorescence experiments in this and previous studies [53], [54] but not by host control serum (Figure 3C). To rule out the possibility that FH::SOLO could be precipitated by cross-reactivity from the anti-SMC1 antibody, the reciprocal immunoprecipitation, i.e., using anti-FLAG antibody to immunoprecipitate SMC1, was carried out and the result showed that SMC1 was co-immunoprecipitated by anti-FLAG antibody (Figure 3D). Our results demonstrate that SOLO associates in vivo with SMC1, one of the core components of the cohesin complex.
The whole-mount preparations of germaria in Figures 2D, 2E and 4A show prominent linear signals of SMC1 and C(3)G in a subset of germ cell nuclei throughout regions 2–3. Based on previous studies, these structures are presumed to correspond to the LEs and central regions, respectively, of SCs [30], [33], [43], [45], [60]–[63]. Although it was less obvious in whole mount preparations, SOLO also localized to linear structures in pro-oocytes and oocytes (Figure 2D, arrowheads, and Figure 4A, arrows). To permit detailed comparisons of these patterns, chromosome spread preparations from UAS-SOLO germaria were stained with antibodies against C(3)G or SMC1. Linear UAS-SOLO signals, presumed to represent staining of chromosome arms, could be clearly seen in meiotic cells (Figures 4B and 4D, arrows), as identified by C(3)G or SMC1 linear structures, but were not confined to the meiotic cells. Thinner linear signals could be discerned in many pro-nurse cells in the same cysts (arrowheads). The same was true for SMC1 (Figure 4D, arrowhead), as previously reported [43], but not for C(3)G (Figure 4B, arrowheads), which is expressed in a meiosis-specific pattern. The thin linear UAS-SOLO and SMC1 signals in pro-nurse cells (Figure 4D, arrowhead) appeared to co-align extensively, similar to ORD and SMC1 [43].
Detailed comparisons of the ribbon-like localization patterns of UAS-SOLO with those of C(3)G and SMC1 in pro-oocytes were possible from magnified images such as those in Figures 4C and 4E. It is apparent from these images that the ribbon-like UAS-SOLO signals overlap quite extensively with the corresponding structures of SMC1 and C(3)G. The overlap is nearly complete for UAS-SOLO and SMC1. Although there were a few prominent segments that exhibited stronger SMC1 signals than UAS-SOLO signals (Figure 4E, arrowheads) and other segments with the reverse pattern, there were no segments of significant length that stained with SMC1 but not UAS-SOLO or vice versa. The overlap between UAS-SOLO and C(3)G was also very substantial but with more segments in which staining was quite unequal (Figure 4C). These results suggest that SOLO is widely distributed along SCs during pachytene and closely aligned with the cohesin SMC1, a pattern consistent with a possible role of SOLO as a component of Drosophila LEs.
To be sure that the results with the ectopically-driven UAS-SOLO were physiologically meaningful, we also carried out chromosome spread experiments using UPS-SOLO germ cells stained with anti-C(3)G (Figure S3). Like UAS-SOLO, UPS-SOLO localized to chromosome arms in pro-nurse cells (lower panels) and along C(3)G ribbon-like structures in pro-oocytes and oocytes (upper panels). However, UPS-SOLO signals were weaker than UAS-SOLO signals, and staining of the LEs was patchy and discontinuous rather than continuous. It is unclear at this point which pattern is correct. The fact that UPS-SOLO failed to fully rescue the X NDJ phenotype may indicate that its expression level is lower than the native gene. However, we cannot rule out the possibility that the more continuous SC labeling pattern of UAS-SOLO is due to overexpression and is therefore misleading. A transgene that expresses SOLO at native levels and fully rescues solo mutants will be required to resolve this question. Overall, these data indicate that SOLO localizes along chromosome arms in a pattern largely parallel to that of SC proteins, suggesting it may have a role in SC formation.
To assess the effects of solo mutations on SC formation, we stained dissected ovaries with antibodies against C(3)G and ORB. ORB is a cytoplasmic protein that is present in all cells in most pachytene cysts, but substantially enriched in pro-oocytes and oocytes [64]. Synapsis phenotypes were analyzed for two different solo alleles (soloZ2-0198 and soloZ2-3534), both of which are genetic null alleles for the NDJ phenotypes [53] (Table 1 and unpublished data).
In solo mutant germaria, both ORB-staining and C(3)G staining were significantly reduced relative to WT germaria. The reduction in staining resulted from two distinct phenotypes: first, a substantial reduction in the numbers of germ-cell cysts per germarium; and second, reduced and/or morphologically abnormal C(3)G staining in many pro-oocytes and oocytes (Figures 5 and S4). However, no defect in oocyte specification was observed. Cysts in region 3 and later stages nearly always had only one cell with enriched ORB staining and no more than one cell with C(3)G staining, (e.g., Figures 5B, 5C, S5B and S6B) although C(3)G staining could be completely absent (e.g., Figure 5C), as described below.
Further analysis revealed that the first phenotype is due not to loss of solo function but instead to an unexpected and, as yet, unexplained inhibitory effect of the solo alleles on expression of vasa, a gene with an overlapping transcription unit that is required for early germ-cell development [53], [65]. Expression of a GFP-VAS transgene in solo/Df females substantially improved the germ-cell cyst number phenotype (Figure S5) and nearly doubled fertility (Table S2) but did not improve either the abnormal C(3)G staining patterns (second phenotype) (Figure S6) or the fidelity of chromosome segregation (Table S2). This shows that the abnormal C(3)G staining patterns are due to loss of solo function, not to reduced vasa function, and will be our focus in the following sections.
The C(3)G staining defects caused by the solo mutations were observed in cells with enriched ORB staining, marking them as pro-oocytes or oocytes, and fell into three main phenotypic categories: i) cells with partial or fragmentary staining; ii) cells with no linear segments at all but only C(3)G foci (spotty staining); and iii) cells that should have exhibited C(3)G staining based on ORB-staining but did not (no staining) (Figures 5B, 5C and S6B). A fourth category consisted of cells with nuclei that appeared to be fully stained and did not exhibit any obvious fragmentation; these were referred to as “normal-like” even though the staining patterns in these cells were often less clearly defined than in WT. Quantitative analysis showed that the three abnormal patterns, fragmentary, spotty, and no staining, were present at highly elevated frequencies, compared to WT, at all pachytene stages in solo germaria (Figure 5D). The quantitative analysis also revealed a progressive deterioration in C(3)G staining with increasing age of cyst. 30–40% of ORB-enriched cells in regions 2a or 2b exhibited normal-like C(3)G staining but that frequency declined to less than 10% by region 3. Some C(3)G staining persisted in some late pachytene oocytes (e.g., Figure S4), but many lacked staining altogether. Staining defects were not limited to C(3)G. SMC1 staining patterns exhibited a similar spectrum of defects (Figure 5E) with very similar frequencies of staining categories (data not shown). Moreover, when the C(3)G and SMC1 staining patterns were compared in the same cells by dual immunostaining, the patterns were very similar, as illustrated by the solo panel series in Figure 5E. Overall, these data indicate that solo mutants cause fragmentation and degeneration of LEs and SCs from the onset of pachytene and that these phenotypes worsen as cysts age.
The phenotypes caused by mutations in solo and ord are very similar in most respects, including the progressive fragmentation and disintegration of both SCs and chromosome cores during pachytene [30], [43]. However, there is a significant difference in the time of onset of abnormalities between solo and ord mutants. Whereas the phenotype is already present at high frequency in region 2a in solo mutants, it doesn't manifest to a significant degree until late stage 2a/stage 2b in ord mutants. To better understand the relationship between these phenotypes, we constructed solo ord double mutants and compared the C(3)G staining patterns to those in ord and solo single mutants. Whereas ord germaria exhibited normal C(3)G staining in region 2a, abnormal C(3)G staining patterns were seen in solo ord germaria at all stages (Figure S7) and did not differ significantly from the pattern in solo mutants (Figure 5D). Why solo mutants disrupt synapsis earlier in pachytene than ord mutants remains to be determined.
The effect of solo on homolog exchange could reflect a defect either in formation or repair of meiotic DSBs. To address these possibilities, DSB frequencies were estimated in pro-oocyte and oocyte nuclei in solo and WT germaria using an antibody against γ-H2Av, a phosphorylated form of the histone variant H2Av protein that becomes enriched around DSBs shortly after their formation and that disappears when DSBs are repaired [66], [67]. γ-H2Av foci and/or short stretches were absent in region 1 germ cells from both solo and WT germaria but were present in pro-oocyte nuclei in regions 2a and 2b in both genotypes, consistent with previous reports [36], [68]. Although solo germaria exhibited fewer total foci than WT germaria, the two genotypes did not differ significantly in mean number of foci per pro-oocyte nucleus, indicating that the DSB formation is not impaired in solo mutants (Figures 6A and 6B). By contrast, unlike in WT, γ-H2Av foci were not restricted to region 2 in solo germaria. All 24 ORB-stained region 3 oocytes that were scored in solo germaria exhibited foci. The mean focus numbers did not differ significantly between region 3 and region 2 (Figures 6A, 6B and S8), suggesting a delay in DNA repair. However, foci did not persist beyond region 3; nearly all stage 2 oocytes and all stage 3 oocytes in solo mutants lacked γ-H2Av signals (Figure 6C). In this regard, solo mutants differ from DSB repair pathway mutants such as spnA, spnB and spnD, in which γ-H2Av foci persist until late pachytene [36], [68]–[71].
In principle, the delayed disappearance of γ-H2Av foci in solo mutants could reflect delayed germ cell development due to the effect of solo mutations on vasa function. In other words, if most region 3 oocytes in solo germaria are really at a developmental age typical of region 2a or 2b pro-oocytes in WT, then the persistence of foci in region 3 would have a trivial explanation. If this were the case, one would expect to see other evidence of delayed development such as failure to restrict ORB staining to a single cell. However, as described above, this was not the case. Nevertheless, to be sure that reduced vasa expression was not somehow responsible for the delayed disappearance of γ-H2Av foci, we compared the γ-H2Av phenotypes of solo; GFP::VAS and solo females. Similar to solo mutants, γ-H2Av foci persisted in region 3 oocytes but were absent in stage 2 oocytes of solo/Df; GFP::VAS/+ (Figure S9) and solo/Df; GFP::VAS/GFP::VAS (data not shown). Thus the delayed disappearance of γ-H2Av foci exhibited by solo mutant females is not due to the effect of the solo mutation on vasa function. These results indicate that solo mutations have no effect on DSB formation but cause a transient delay in DSB repair. The cause of this delay and its significance with respect to the recombination phenotype of solo mutants are unknown.
Our previous analysis of solo in Drosophila male meiosis showed it to be essential for meiotic centromere cohesion and centromere orientation. However, the idiosyncratic homolog segregation mechanism in males precluded analysis of roles of solo in homolog interactions [53], [54]. In this study we analyzed the role of solo in female meiosis and found that solo mutations disrupt a much broader range of meiotic processes in females, including centromere clustering, homologous centromere pairing, sister centromere cohesion, sister centromere mono-orientation, SC and lateral element stability, homolog exchange, and homolog bias. Moreover, SOLO protein localized to chromosome arms and along the LEs of the SCs as well as to centromeres in female meiosis. These results indicate that SOLO contributes to multiple sister chromatid and homolog interactions that underlie meiotic chromosome segregation.
solo mutations severely disrupted chromosome segregation, causing X chromosome NDJ at frequencies in excess of 50% (Table 1). The NDJ pattern, a 1∶2 ratio of sister chromatid to homolog NDJ seen also in male solo mutants and ord mutants of both sexes, is consistent with random chromatid assortment caused by loss of centromere cohesion prior to prometaphase I [48], [50], [53]. Centromere cohesion was visibly impaired by late pachytene in solo females, based on CID spot numbers that consistently exceeded eight per cell (Figure 1). Similar observations were reported for ord mutants in female and male meiosis [52], [61] and solo mutants in male meiosis [53]. Although cytological analysis of segregation in solo females has not been undertaken, FISH analysis in solo males revealed random co-segregation of chromatids at anaphase I, fully separated chromatids by mid-anaphase I and chaotic segregation at anaphase II [53] and several cytological studies of segregation in ord males and females have documented premature sister chromatid separation and disorderly segregation behavior [47]–[52], [72].
The mechanism by which solo controls centromere cohesion seems likely to involve cohesin. In male meiosis, SOLO, ORD and SMC1 are enriched on centromeres until anaphase II and all three proteins depend on the Shugoshin ortholog MEI-S332 for maintenance on centromeres after metaphase I [43], [52], [53]. In female meiosis, SOLO, ORD, SMC1 and SMC3 are all enriched on centromeres in female meiosis throughout pachytene (Figures 2 and 4) [30], [43]. When either solo or ord is mutated, no centromeric SMC cohesin foci have been detected at any stage in either sex (Figure 2) [43], [53] with the consequences summarized above. These data are consistent with the hypothesis that centromere cohesion is mediated in male and female meiosis by centromere enrichment of a cohesin complex dependent on both SOLO and ORD. However, there has been no direct demonstration that the cohesive roles of SOLO and ORD are limited entirely to regulating cohesin. There also remains no direct evidence that any of these proteins – SMC1, SMC3, ORD or SOLO – persists on centromeres after pachytene. That may be a technical detection issue of some sort but until such evidence is obtained, the possibility that cohesion is maintained during the division stages in female meiosis by some other complex cannot be ruled out.
It is worth noting that centromere cohesion persisted intact throughout early and mid-pachytene in solo mutants despite the absence of detectable SMC1 centromere foci at any stage (Figure 1). Similar observations have been reported for solo male meiosis and ord male and female meiosis [51]–[53], [60], [61], [73] and indicate the existence of centromere cohesion that is independent of both SOLO and ORD and perhaps of cohesin (although the possible presence of low levels of cohesin near centromeres in solo and ord mutants cannot be ruled out). Whether this early prophase cohesion is based on a protein complex or on chromatid entanglement remains to be determined.
Homologous centromeres are paired in nearly all germ cells and they further coalesce into 1–3 clusters at the onset of meiosis in pro-oocytes and remain paired and clustered throughout prophase I [43], [60], [61], [74]. There is considerable evidence that centromeric or heterochromatic associations between homologs underlie the robust achiasmate segregation system in Drosophila [74]–[77]. Moreover, centromere clusters serve as the first synapsis initiation sites during zygotene, accumulating the transverse filament protein C(3)G and the central element protein CONA [34], [60], [61], [78]. Both pairing and clustering (as well as synapsis initiation) was shown to depend on ord [43], [60], [61]. Here we demonstrate that solo is also required for these events. In early-mid pachytene solo pro-oocytes exhibited 6.3 foci per nucleus compared to 2.3 in WT, indicating substantial loss of pairing and clustering (Figure 1). Since SOLO and ORD are required for centromere enrichment of SMC1 as well as for centromere pairing and clustering, a logical inference is that centromere pairing is also mediated by cohesin, as previously suggested [60]. This suggestion is supported by evidence that centromere pairing is weakened in certain chromosomal backgrounds by reducing SMC1 gene copy number [79]. The mechanism by which cohesin mediates pairing and clustering is not known. Clustering may involve recruitment of SC proteins since mutations in c(3)G and cona abolished clustering [61]. However, c(3)G and cona mutations had much weaker effects on centromere pairing suggesting other mechanisms are probably involved in this process. Interestingly, yeast REC8 is also required for centromere pairing (called coupling) in early prophase I and promotes pairing by recruiting the yeast version of C(3)G, ZIP1 [80]. However, the relevance is not clear since centromere coupling in yeast is entirely promiscuous whereas Drosophila pairing is homologous [51]–[53], [73]–[76]. The mechanistic relationship between cohesion and centromere pairing remains to be elucidated. Given the association between centromere pairing and synapsis, it will also be of interest to investigate the role of solo in synapsis initiation.
What are the roles of ORD and SOLO in cohesin function? Neither protein exhibits significant homology to any of the four cohesin protein families [49], [53], yet they appear to co-localize with SMC cohesins and are required for enrichment of SMC cohesins at centromeres. We favor the idea that SOLO and ORD are subunits of a meiosis-specific cohesion complex that includes the SMC subunits. ORD and SOLO may function to replace the canonical non-SMC subunits which, with the exception of C(2)M, have yet to be identified in Drosophila meiosis. Our finding that SOLO and SMC1 reciprocally co-immunoprecipitate from ovarian protein extracts is consistent with this idea but also with alternatives such as that SOLO is a regulator rather than a subunit of cohesin. More detailed biochemical analyses will be required to resolve the composition of Drosophila meiotic cohesin and to clarify the roles of SOLO and ORD.
Cohesion between sister chromatid axes is clearly essential for maintenance of chiasmata but its role in early prophase I events such as homolog pairing, synapsis and meiotic recombination is unclear. In WT Drosophila, FISH studies indicate that sister chromatid arm sequences are tightly cohesive throughout prophase I [30], [74], but the genetic basis for arm cohesion remains to be elucidated. In c(2)M mutants, recombinant chromatids were not recovered in NDJ gametes, suggesting that chiasmata are stable and can bi-orient bivalents [45]. In ord mutants, absence of metaphase I arrest indicated an absence of chiasmata [51]. Presumably this implies that ORD also provides arm cohesion during prophase I and C(2)M does not, but direct evidence is lacking. ORD and the SMC cohesins are abundant on chromosome arms in all cells in 16-cell germ-line cysts, but ord mutants have little if any effect on intensity of SMC1/3 arm staining even in pro-oocytes and oocytes with fragmented cores [43]. Moreover, the limited FISH analysis that has been carried out thus far has not detected any disruption of arm cohesion during prophase I in ord mutants [30].
Our data show that SOLO is also expressed in all cells in 16-cell germline cysts and localizes to chromosome arms in both pro-nurse cells and pro-oocytes and oocytes. UAS-SOLO and UPS-SOLO are fully consistent in this respect (Figures 2, 4 and S3). In spread preparations co-stained with anti-SMC1 it is quite clear that the two proteins co-align very strongly even though the staining lines are thin (Figure 4D). These data suggest that SOLO may also be involved in arm cohesion. This is an important question for future research because the roles of solo in synapsis, chromosome core stability and recombination could be related to its role in arm cohesion.
Our data show that SOLO localizes to extended ribbon-like structures on the chromosome arms of pro-oocytes and oocytes, where it co-aligns with both SMC1 and C(3)G. This localization pattern is unlikely to be an artifact since it was seen with both UAS-SOLO and UPS-SOLO. However, it remains unclear whether the true pattern is the continuous staining pattern seen with UAS-SOLO or the discontinuous pattern seen with UPS-SOLO. Since UAS-SOLO appeared to be somewhat overexpressed in anterior 2a, the continuous localization could be an overexpression artifact. However, since UPS-SOLO did not fully rescue X chromosome NDJ in solo females, the discontinuous localization pattern could be an underexpression artifact. For now, we favor the continuous pattern in part because ORD localizes continuously [30], [43] and the phenotypes of ord and solo are so similar that sharply different localization patterns seem unlikely. Ascertaining the true localization pattern is an important goal.
Where exactly does SOLO localize? Overall, SOLO appeared to align slightly better with SMC1 than with C(3)G. However, this difference is not large and would not in itself suffice to assign SOLO to the LEs rather than the central regions. There are two independent reasons to favor the LEs. First, the close alignment of SOLO and SMC1 signals along unsynapsed chromosome arms of germ cells would likely persist during core assembly (Figure 4). Second, the highly correlated SMC1 and C(3)G staining phenotypes in solo mutants suggest that solo controls chromosome core stability directly rather than indirectly through effects on the central region, as even null mutations in c(3)G do not perturb chromosome core integrity (Figure 5) [45]. In other eukaryotes a distinction is often made between chromosome core proteins, which are cohesins, and non-core AE/LE proteins such as RED1, HOP1, SYCP-2, SYCP-3, etc., and there has been a spirited debate about how the two groups of proteins are organized relative to each other [6], [27]–[32], [37]–[42]. For Drosophila, the distinction would seem artificial at this point. The only proteins identified thus far that localize to the LEs – SMC1, SMC3, C(2)M, ORD, Nipped-B and SOLO – are all either cohesins or cohesion proteins with very close links to cohesins, and therefore seem likely to be components of the cores [30], [43], [45], [53], [60]–[63].
Our working model is that SOLO and ORD function as subunits of a cohesin complex that is distributed along the chromosome arms of all germ cells and likely provides cohesion between the sister chromatid axes. We do not dismiss the possibility that SOLO/ORD cohesin maintains cohesion in the chromatin loops as well but evidence has been presented that SMC cohesins are mostly confined to the axes in Drosophila germ cells [43]. In meiotic cells these arm cohesins condense along with C(2)M-cohesin (and perhaps other complexes) and assemble into continuous cores that underpin synapsis and SC formation. SOLO and ORD are unlikely to be components of different cohesin complexes since core stability was no worse when both ORD and SOLO were absent than when just SOLO was absent (Figures 5 and S7). Thus cores may consist of two cohesin complexes, one anchored by C(2)M and one anchored by ORD and SOLO. Additional cohesin complexes involving mitotic cohesins such as RAD21 might be present as well.
Our observation that pro-oocytes with fragmented, patchy or no SMC1 and C(3)G staining are abundant even at the earliest stages of pachytene in solo mutants could indicate a requirement for SOLO in assembly of cores. In addition, the progressive degeneration of cores throughout early and mid-pachytene in solo mutants might indicate a possible role in core maintenance as suggested for ord [30], [43]. A role of SOLO in core assembly seems unlikely. Full-length cores can be assembled in the absence of SOLO or of both ORD and SOLO (Figures 5 and S7) [30], [43]. However, no cores are assembled in the absence of C(2)M, suggesting that C(2)M is the motor for assembly and that SOLO and ORD play passive roles [43], [45], [60]. A maintenance function is plausible but not especially compelling since it doesn't relate in any direct way to the primary function of SOLO. In our model, SOLO is a subunit of arm cohesin complexes that become assembled into cores in meiotic cells. This would make SOLO a structural component of WT cores and its absence would be expected to compromise core structure in one of two ways. First, cores assembled with abnormal (i.e., SOLO-deficient) cohesins might be less stable than WT cores and prone to breakage or disassembly. Second, exclusion of deficient cohesins from core assembly would likely lead to monolithic cores which might lack important structural or functional properties such as flexibility or ability to complete exchanges with homologous cores.
A major strength of this hypothesis is that it does not require a fundamentally different explanation for the solo and ord phenotypes, just a difference in degree of instability of the cores. If absence of SOLO is for some reason more destabilizing than absence of ORD, then it could trigger core degeneration at earlier stages of meiosis. One way this could work is based on our proposal that SOLO and ORD are subunits of the same cohesin complex. The effect of loss of a subunit on complex stability depends on the specific role of that subunit. For example, absence of the kleisin subunit is more destabilizing for conventional cohesin than absence of the SA subunit.
In solo mutants, all of the assembled cores in early pachytene must be defective but actual fragmentation and dissolution does not begin until later in pachytene in some cells. In other cells, dissolution is already complete in region 2a. This suggests that the defect creates a fragile state and that onset of degeneration may require a stressful event of some kind to trigger it, as suggested for ord [43]. The cell-to-cell variability in phenotype could reflect stochastic variation in degree of fragility, or perhaps cell-to-cell variation in the numbers or intensity of stressors.
SOLO is required for completion of DSB repair on the normal schedule although the repair delay is brief compared with the delays caused by mutations in components of the DSB repair pathway (36,68–71). Mutations in other Drosophila chromosome core components such as c(2)M and ord have no effect at all on DSB repair (30,36,68). This is somewhat surprising in light of the often severe DSB repair defects seen in cohesin mutants in other eukaryotes (2,7–9,13,14). Additional studies will be required to determine if the transient repair delay in solo mutants contributes to its recombination phenotype.
Our data indicate that SOLO promotes homolog exchange and suppresses SCE (Tables 3–5). As SCE and homolog exchange are alternative pathways for DSB repair, suppressing SCE is likely to promote homolog exchange; direct molecular analysis of recombination intermediates in yeast confirms this [23], [24]. We conclude that a major role of SOLO in recombination is to regulate homolog bias, although this does not preclude SOLO acting in other ways to promote homolog exchange.
How might SOLO regulate homolog bias? There is a bit of a conundrum here: the primary function of SOLO is cohesion and although cohesion is very effective at promoting DSB repair, it does so by promoting SCE, presumably by reinforcing sister chromatid proximity [81]. REC8 becomes depleted around crossover sites presumably because it promotes SCE [82]. Moreover, in yeast, rec8 mutations promote homolog bias, not sister bias [24]. Therefore, simply providing extra cohesion at a recombination site is more likely to inhibit homolog exchange than to promote it. An alternative is that the chromosome cores per se are responsible for suppressing SCE. Several recent models have postulated that AE/LEs serve as “barriers to sister chromatid repair” (BSCR) [28], [29]. This mechanism seems unlikely to apply to Drosophila because c(2)M mutations completely abrogate core assembly but do not de-repress SCE at all [45].
Our proposal is that SOLO/ORD-cohesin is an unconventional cohesin that is able to flexibly regulate cohesion in the context of meiotic recombination. It becomes enriched at future DSB sites, perhaps specifically at future crossover sites, during the synapsis initiation process, where it regulates the cohesive status of chromatids involved in the recombination reaction to promote inter-homolog exchanges. For example, relaxation of cohesion between the broken chromatid and its sister may be necessary to allow a homology search and inter-homolog strand invasion [24]. We speculate that ORD/SOLO-cohesin is able to rapidly switch to a “cohesion-off” mode in response to local signaling related to DSB or recombination intermediate status. In doing so, ORD/SOLO-cohesin might be able to promote homolog exchange locally while still maintaining cohesion globally.
In conclusion, SOLO is a meiotic cohesion protein with major roles in centromere cohesion, chromosome core integrity and homolog bias. It is enriched at centromeres and chromosome cores and interacts with the SMC1 cohesin subunit. Further investigation of SOLO's meiotic functions is expected to provide insight into the roles of cohesion in inter-homolog interactions.
The solo mutants used in this paper were described previously [53]. soloZ2-0338, soloZ2-0198 and soloZ2-3534 are single-base substitutions predicted to insert stop codons in the SOLO coding sequence and truncate the proteins at amino acid positions 173, 387 and 1010 (out of 1031), respectively [53]. All three alleles are considered to be functionally null with respect to chromosome segregation. Although a closely-linked semi-lethal mutation has thus far prevented accurate measurement of NDJ in soloZ2-3534 homozygotes, both male and female sex chromosome NDJ frequencies in the other two homozygotes and in all three hemizygotes are statistically indistinguishable and consistent with random chromatid segregation [53] (Table1, unpublished data). The b vas7 stock was obtained from M. Ashburner (Cambridge University, England). The X chromosome mapping stock y pn cv m f.y+/FM7c was provided by K. McKim (The state University of New Jersey). ord5 and Df(2R)WI370 were donated by S.E. Bickel (Dartmouth College). The GFP::VAS transgenic line was provided by P. Lasko (McGill University). Other flies were from the Bloomington Drosophila Stock Center at Indiana University. Unless otherwise specified, the females being tested were crossed singly to two males in shell vials. All flies were maintained at 23°C on standard cornmeal molasses medium. Parents were removed from the vial on day 10 and progeny were counted between day 13 and day 22.
The methods for analyzing NDJ and recombination on the X and second chromosomes are explained and illustrated in Figures S1 and S2 and in Tables 1–4 and S1.
To accurately estimate the relative frequencies of sister and homolog NDJ, it is necessary to correct for the reduced viability of the sister NDJ classes which are homozygous for most or all of chromosome 2, relative to the homolog NDJ classes, which are heterozygous. The viability test was based on recoveries of the homozygote and heterozygote progeny classes from two crosses: soloZ2-0198 cn bw/b vas7 males crossed to soloZ2-0198 cn bw/Cy females and soloZ2-0198 cn bw/b vas7 males crossed to b vas7/Cy females. The viabilities of b vas7 and soloZ2-0198 cn bw homozygotes were found to be 51.76% and 63.49%, respectively, compared to their heterozygous siblings (b/cn bw). Plugging the decimal versions of those correction factors into the formula for %S NDJ gives %S = 100×(144/0.5176+106/0.6349+37)/((144/0.5176+106/0.6349+37)+(1012+36)) = 32%.
R(1)2, y1 f1/BSYy+ males were crossed to Df(2L)A267, b cn bw/CyO, cn females. The R(1)2, y1 f1/+; Df(2L)A267, b cn bw/+ F1 female progeny were crossed to y w/Y; solo, cn bw/CyO, b cn males to generate F2 R(1)2, y1/y w; Df(2L)A267, b cn bw/solo, cn bw females and sibling control R(1)2, y1/y w; +/CyO, b cn females. These F2 females were crossed to w1118/Y males and their progeny scored for the ring-X (w+) and rod-X (w). The crosses were carried out without an X chromosome balancer to enable estimation of SCE frequencies under conditions in which both homolog and sister chromatid exchanges were free to occur. The ring-X chromosome was tracked using the y/y+ marker in the F1 cross and the w/w+ marker in the F2 (test) cross. The y1 allele on the ring-X chromosome is recombinationally inseparable from the centromere, and w, which is 1.5 cM from y, does not recombine with y at appreciable rates in ring/rod heterozygotes where only double exchanges can be recovered (unpublished data). In the F1 cross, cn was used as a proxy for Df(2L)A267. solo/Df F2 females were sorted by Cy cn+ phenotype and verified (or not) on the basis of fertility and NDJ. Only regular (disjunctional) progeny were used to calculate the ring/rod recovery ratio.
pENTR-Ntag-SOLO entry vector [53] was recombined into Gateway P-element vector pPFH (Drosophila Genomics Resource Center (BDGC)), generating the germ line transformation vector P{w+mC UASp-FH::SOLO}, which contains tandem 3XFLAG and 3XHA tags at the N-terminus of SOLO fusion protein. The construct was transformed into w1118 flies (BestGene Inc.). Transgenes were mapped by standard methods and tested for ability to suppress X chromosome NDJ in solo females when expressed with the nos-GAL4::VP16 driver [83] (see Table S1, lines 4 and 5).
FH::SOLO expression was induced by nos-GAL4::VP16 in Drosophila females and 100 pairs of ovaries were collected with 1X PBS (pH 7.4). Ovaries of y w and transgenic flies were lysed using 500 µL of NP40 Cell Lysis Buffer (Invitrogen). The lysates were centrifuged at least 4 times each at 13,000 g for 10 minutes to remove tissue debris and the supernatants were used for Western blots and immunoprecipitations. Before immunoprecipitating, lysates were first pre-cleared with rabbit serum. 4 µL of rabbit serum (159 mg/ml, Sigma) were added to the 500 µL lysates and rocked for 1 hr at 4°C, then the lysates with rabbit serum were added to 100 µl of protein A agarose beads (Invitrogen) which had been washed 5 times with wash buffer (1 mM PMSF, 1 mM DTT, 1X PI (Protease Inhibitor (Roche)), 10% glycerol, 10 mM NaCl, 1X PBS, pH 7.4) rocking for 30 minutes at 4°C.
To immunoprecipitate FH::SOLO with anti-SMC1, 50 µl of pre-cleared lysates were incubated with 20 µl of anti-SMC1 rabbit antibody (1.03 mg/ml) or rabbit serum (1.06 mg/ml, diluted from original serum) and IP solution (1 mM PMSF, 1 mM DTT, 1X PI (Protease Inhibitor (Roche)), 10% glycerol, 1X PBS, pH 7.4) rocking for 4 hrs. The lysates with anti-SMC1 antibody or serum were then added to 80 µL of washed protein A agarose beads and rocked overnight in a cold room at 4°C.
To immunoprecipitate SMC1 by FH::SOLO, 50 µl of lysate pre-cleared with mouse serum and protein G agarose beads (similar procedure to rabbit serum) were incubated with 30 µL of anti-FLAG M2 (1 mg/ml, Sigma) or mouse serum (1.10 mg/ml, diluted from original serum) and the IP solution was rocked for 4 hrs. Lysates with anti-SMC1 antibody or serum were then added to 100 µl of washed protein G agarose beads and rocked overnight in a cold room (4°C).
After IP, lysates/antibody or serum/IP solutions/beads were centrifuged and beads were washed 6× times with wash buffer. 30 µL of loading buffer were added to the beads and heated to release protein binding to the beads.
The lysates from FH::SOLO and y w flies that were used in Western blot to test antibody specificity and the released solutions (from IP experiment) were run in 8% SDS-PAGE Acr/Bis electrophoresis. FH::SOLO was detected by using anti-FLAG M2 antibody (1∶1000, Sigma) and goat anti-mouse HRP-conjugated (1∶1000, Chemicon) with Supersignal West Pico (Pierce). SMC1 was detected by using anti-SMC1 (1∶200, rabbit) and goat anti-rabbit HRP-conjugated (1∶2000, JacksonImmuno) with Supersignal West Pico (Pierce).
Newly eclosed females were fattened 1–3 days in vials with yeast paste and males and then ovaries were dissected in 1X PBS (pH 7.4). Immunostaining of whole-mount ovarioles was performed according to Page and Hawley [34]. After immunostaining, ovaries were separated into individual ovarioles and transferred to slides and mounted with Prolong Antifade reagent (Invitrogen). UASp-Venus::SOLO expression was induced by nos-GAL4::VP16 and fluorescent signals were detected in the FITC channel or detected by anti-GFP antibody. Egg chambers were staged according to Matthies et al. [84]. Chromosome spreads were performed according to Webber et al. [30].
For WT germaria, pro-oocytes and oocytes in pachytene were identified by full-length C(3)G nuclear staining and enriched cytoplasmic ORB staining. For solo germaria, pro-oocytes and oocytes were identified by enriched cytoplasmic ORB staining, except in Figure S6 where C(3)G staining was used. In that figure, the “no-staining” category was not scored. For solo pro-oocytes in region 2a with abnormal C(3)G staining, the ORB-enrichment criterion ensured that zygotene cysts were not inadvertently included in the scoring. Even without ORB staining, however, zygotene nuclei could usually be distinguished from the defective pachytene nuclei by C(3)G staining. The C(3)G foci are usually smaller and more uniform in size in zygotene than the “spotty” staining in pachytene, and lengthy linear fragments are never seen in zygotene. Staging (regions 2a, 2b, and 3) was based on position of cysts in the germarium (see Figure 1) and/or shape of cysts (rounded in region 2a, flattened in region 2b). Oocytes in egg chambers were identified by ORB enrichment, C(3)G staining, nuclear size (smaller than polyploid nurse cell nuclei) and/or position in cyst (posterior). For scoring of γ-H2Av foci, pro-oocytes and oocytes were identified by enriched ORB staining. Pro-nurse cell nuclei were not scored.
Linear C(3)G staining was used to identify pachytene pro-oocytes and oocytes from the 2a region of WT and solo germaria. Nuclear boundaries were established based on margins of DAPI and C(3)G staining. Nuclei with overlapping DAPI or C(3)G staining were not used for scoring. Only non-overlapping CID spots were scored as separate spots. Size and brightness of CID spots was not considered.
Primary antibodies used : 1∶500 anti-C(3)G (mouse monoclonal and guinea pig polyclonal antibody (provided by R.S. Hawley), 1∶500 rabbit anti-GFP polyclonal antibody (Invitrogen), 1∶800 rabbit anti-CID polyclonal antibody (Active Motif), 1∶200 anti-SMC1 rabbit polyclonal antibody [53], [54], 1∶5000 rabbit anti-γ-H2Av antibody (Rockland), 1∶3000 anti-VASA antibody (P. Lasko), 1∶150 anti-ORB (6H4 and 4H8, monoclonal, Developmental Studies Hybridoma Bank (DSHB)). Secondary antibodies (IgGs) used: Alexa Fluor 488 donkey anti-rabbit, Alexa Fluor 488 goat anti-guinea pig, Alexa Fluor 555 donkey anti-mouse, Alexa Fluor 555 donkey anti-rabbit, Alexa Fluor 647 donkey anti-mouse (Invitrogen).
All images were collected using an Axioplan (ZEISS) microscope equipped with an HBO 100-W mercury lamp and high-resolution CCD camera (Roper). Image data were collected and merged using MetaMorph Software (Universal Imaging Corporation). Adobe Photoshop CS2 and Illustrator CS2 were used to process images. Each image in the immunofluorescence figures came from a sum projection of 3D deconvolved z-series stacks. All images from WT and mutants were exposed for equal periods and deconvolved and processed identically.
|
10.1371/journal.pntd.0006188 | An agent-based model of tsetse fly response to seasonal climatic drivers: Assessing the impact on sleeping sickness transmission rates | This paper presents the development of an agent-based model (ABM) to incorporate climatic drivers which affect tsetse fly (G. m. morsitans) population dynamics, and ultimately disease transmission. The model was used to gain a greater understanding of how tsetse populations fluctuate seasonally, and investigate any response observed in Trypanosoma brucei rhodesiense human African trypanosomiasis (rHAT) disease transmission, with a view to gaining a greater understanding of disease dynamics. Such an understanding is essential for the development of appropriate, well-targeted mitigation strategies in the future.
The ABM was developed to model rHAT incidence at a fine spatial scale along a 75 km transect in the Luangwa Valley, Zambia. The model incorporates climatic factors that affect pupal mortality, pupal development, birth rate, and death rate. In combination with fine scale demographic data such as ethnicity, age and gender for the human population in the region, as well as an animal census and a sample of daily routines, we create a detailed, plausible simulation model to explore tsetse population and disease transmission dynamics.
The seasonally-driven model suggests that the number of infections reported annually in the simulation is likely to be a reasonable representation of reality, taking into account the high levels of under-detection observed. Similar infection rates were observed in human (0.355 per 1000 person-years (SE = 0.013)), and cattle (0.281 per 1000 cattle-years (SE = 0.025)) populations, likely due to the sparsity of cattle close to the tsetse interface. The model suggests that immigrant tribes and school children are at greatest risk of infection, a result that derives from the bottom-up nature of the ABM and conditioning on multiple constraints. This result could not be inferred using alternative population-level modelling approaches.
In producing a model which models the tsetse population at a very fine resolution, we were able to analyse and evaluate specific elements of the output, such as pupal development and the progression of the teneral population, allowing the development of our understanding of the tsetse population as a whole. This is an important step in the production of a more accurate transmission model for rHAT which can, in turn, help us to gain a greater understanding of the transmission system as a whole.
| African trypanosomiasis is a parasitic disease which affects humans and other animals in 36 sub-Saharan African countries. The disease is transmitted by the tsetse fly, and the human form of the diseases is known as sleeping sickness. In an attempt to improve our understanding of the mechanisms which contribute to sleeping sickness transmission, a detailed, seasonally driven model of the tsetse fly has been produced, with the theory that a greater understanding of the disease vector’s life cycle will allow developments in our knowledge of disease transmission. The model incorporates previously developed spatial data for the Luangwa Valley case study, along with demographic data for its inhabitants. Tsetse and potential human and animal hosts are modelled at the individual level, allowing each contact and infection to be recorded through time. Through modelling at a fine-scale, we can incorporate detailed mechanisms for tsetse birth, feeding, reproduction and death, while considering what demographics, and which locations, have a heightened risk of disease.
| The tsetse fly (genus: Glossina) is the vector for human African trypanosomiasis (HAT) or sleeping sickness, a neglected tropical disease caused by two sub-species of the protozoan parasite Trypanosoma brucei s.l.: T. b. rhodesiense, in eastern and southern Africa and T. b. gambiense in West Africa [1]. T. b. rhodesiense HAT (rHAT) is a zoonosis, affecting a wide range of wildlife [2,3] and domestic animals, particularly cattle [4], presenting in humans as an acute disease [5]. The history of HAT in sub-Saharan Africa is characterised by long periods of endemicity where the disease self-sustains at low background levels, with periodic epidemics in regional foci [6]. As sleeping sickness is a neglected tropical disease, treatments are often out-of-date, difficult to administer, physically invasive and partially validated, with the prospect for future developments of more effective treatments being limited (e.g. [7–11]). Furthermore, where tools are available, HAT is rarely prioritised due to competing public health interests [12]. In terms of disease prevention, there is currently no immunological prophylaxis to stop infection in humans [13], made difficult to produce due to the parasite being able to evade the host's immune response by altering the antigenic character of its glycoprotein surface coat [14]. Given these difficulties with preventing and treating HAT infection in humans, it is not surprising that mitigation strategies focused on vector control have seen success (e.g. [15–18]), given that the tsetse fly is not only required for transmission, but also for several stages of parasite development [19,20]. Despite such efficacy, the control of the disease in tsetse (and, therefore, wildlife) in game reserves and other protected areas is complicated by ecological, conservationist and environmental considerations [21–23].
Gaining a greater understanding of the population dynamics in a tsetse population appears to be an attractive goal, considering that such an understanding could lead to the development of more targeted vector control strategies which have a less adverse ecological impact, while also allowing a more plausible understanding of the rHAT transmission system. For the latter, demographic growth (through the availability of food and habitat) and climate changes (affecting tsetse development and mortality rates) are two factors which could affect tsetse population dynamics, and ultimately affect the transmission system [24,25]. As a result of the significant role that a tsetse population has in determining the rate and distribution of rHAT transmission, this paper considers the tsetse sub-component of the larger rHAT transmission system in detail, with the ultimate goal being the creation of a more accurate representation of the transmission system as a whole.
Collecting comprehensive data on populations of tsetse in the field is expensive, complex and time consuming and, consequently, numerous attempts have been made to model tsetse populations as part of vector control or HAT transmission studies (e.g. [26–29]). Some models incorporate climatic drivers which create fluctuations in the tsetse population through the seasons (e.g. [30–33]). One recent example used agent-based modelling (ABM) techniques to simulate a simple fluctuation in tsetse population size through different seasons by altering the length of a predetermined lifespan for tsetse, depending on whether the tsetse emerges in the dry (2 months) or wet season (3 months) [31]. Incorporating more detail, [33] used known relationships between temperature and different life events and processes, such as mortality and the length of the pupation period, as parameters when constructing a population model for vector control.
ABMs are “a computerized simulation of a number of decision-makers or agents, and institutions, which interact through prescribed rules” [34]. ABMs have been described as a “third way” of conducting scientific research, incorporating both deductive since ABMs start with basic assumptions, and inductive approaches, as they produce simulation data to analyse [35]. However, Epstein [36] suggests that rather than inductive or deductive, ABMs should be considered as “generative” tools in that, through the initialisation of a population of autonomous agents in a relevant spatial environment, one can allow the agents to interact given a simple set of local rules, and generate, from the bottom up, the macroscopic behaviour and regularity of the population as a whole. Such an approach lends itself well to both the investigation of the HAT transmission system as a whole and the tsetse populations and their dynamics as a component. Starting with tsetse population dynamics, much is written about how varying climatic conditions have different impacts on various tsetse life events and processes e.g.: pupal period duration (e.g. [37]), probability of pupal death (e.g. [38,39]), and time between oviposition (e.g. [40,41]). Representing observations made from samples acquired both in the field and laboratory studies, these patterns provide us with a solid framework to model the larger population, for which comprehensive data are much more difficult, if not impossible, to acquire. By initialising a tsetse population as individuals, each abiding by rules set by the above behavioural patterns (and others relating to feeding, mating and age-dependent mortality), plausible population level outcomes such as fluctuations in population size should be observable as the simulation progresses.
When the HAT transmission system is incorporated into an ABM for acquiring preliminary knowledge of the disease transmission system, the constructed model becomes a representation of a complex system (e.g. [42–44]), given that the prevalence of the disease is a complicated emergent phenomenon produced by relatively simple, individual specific rules (both vector and host) concerning movement and resource acquisition. In a complex system, the causes of emergent phenomena cannot easily be decoupled and explained by specific parts of the system [45] with, in this case, the model landscape and agent behaviour creating variation in the timing, location and probability of infection as a result of their influence on variability in contact patterns between vector and host [46,47]. In this way, ABMs could be considered the most appropriate way to investigate both the HAT transmission system, and tsetse fly dynamics as a sub-component, allowing the representation of interdependent processes such as how individuals interact with each other and their environment through space and time more easily than is possible through more traditional epidemiological techniques [48].
In previous work, an ABM of rHAT transmission was produced using a spatialized approach, incorporating factors often overlooked (e.g. human behaviour and activity-based movement; density and mobility of vectors; and the contribution of additional hosts) [27]. This paper presents the first ABM which considers the effect of climatic factors on individual tsetse and their life processes in detail, while also considering the effect this has on rHAT transmission in a large study area in Eastern Province, Zambia. Through the incorporation of seasonality parameters into an existing fine spatial and temporal scale ABM of rHAT transmission in the region [27], the aim was to develop a greater understanding of tsetse population dynamics through simulation, and subsequently produce a more plausible model of rHAT transmission. The incorporation of such data is vital where transmission rates, and indeed the transmission system as a whole, are to be explored over multiple years. The existing model provided a suitable starting point for the simulation of these seasonal parameters by modelling tsetse flies at the individual level, along with different life events for which durations and probability of occurrence can be climatically constrained. Ultimately, the modified model was implemented with the aim of answering the following research questions: throughout the year, how does the tsetse fly population fluctuate both as a whole, and within different life stages (e.g. pupal, teneral, mature)? Under the caveat that a plausible model has been produced, what rates of disease transmission are observed, and how do these vary seasonally? Such a model will allow for future exploration of long-term mitigations strategies, alterations to the demographic make-up of the study area, and climate change scenarios.
Eastern Province, Zambia is situated in southern Africa, sharing borders with Malawi (to the East) and Mozambique (to the South). The Luangwa Valley is an extension of the Great Rift Valley of East Africa, traversing the Zambian Eastern, Northern and Muchinga Provinces. The valley is characterised as a flat bottomed valley bounded by steep, dissected escarpments which rise to a plateau at approximately 900–1000 m [49]. Different types of vegetation are observed at different altitudes, with valley areas consisting mainly of mopane woodland and patches of grassland, while the natural vegetation on the escarpment and plateau is miombo woodland, interspersed with munga woodland [50].
The study area spans a sparsely populated region of the Luangwa Valley. Villages are small (between 5 and 20 households) and inhabitants are predominantly subsistence farmers. The data collection area and region to be modelled consists of a 75 km transect which starts close to Mfuwe airport in the north, and runs southwards along the Lupande River and its distributaries (Fig 1). Average monthly temperature and rainfall measurements collected at the Mfuwe airport (1982–2012) weather station are reproduced in Fig 2 [51]. There are three main seasons in Zambia’s tropical climate: the rainy season spans November to April (wet and warm) with mean monthly rainfall peaking at 210 mm in January. After the rains, a cold and dry period occurs prior to August, in which May is the hottest and wettest month, with mean temperatures below 23°C and mean rainfall below 3 mm. The hot and dry season usually spans August, September and October, with mean temperatures reaching 28°C in October accompanied by 17 mm of rainfall on average, the first after four dry months in succession [49,51]. The Luangwa River and its main tributaries are perennial, and although flash flooding occurs in all rivers during the wet season, the smaller rivers which drain the valley floor dry out during the dry season and flow during the rains [52].
rHAT is endemic in the Luangwa Valley, first being reported in 1908 [53]. G. m. morsitans was not originally considered a vector of rHAT in the valley, despite 50% of domestic and game animals in the Valley having been observed to harbour trypanosomes [54]. In the early 1970s, a large rHAT outbreak occurred in Isoka (241 case in 3 years) attributed to fly encroachment from Luangwa [55]. Wildlife had been observed to reside in Isoka for several months during the rainy season, migrating away during the dry season. In 1973, early diagnosis and improved treatment methods were introduced, and case numbers fell [56].
Today, cases of rHAT continue to be reported in the Luangwa Valley. Mid-Luangwa Valley has recently experienced increased immigration of people seeking fertile land. Land pressure has resulted in human settlement in increasingly marginal, tsetse-infested areas, previously avoided for fear of disease risk to introduced livestock. Households grow cotton as a cash crop and maize and groundnuts for home consumption [49]. These anthropogenic changes have the potential to destabilise current trypanosomiasis transmission cycles, resulting in increasing prevalence of trypanosomiasis in both human and animal hosts, and the spread of rHAT into previously unaffected areas. Risk factors include human proximity to the large wildlife reservoir in the South Luangwa National Park to the north-west [2], and ever-increasing livestock and human density on the plateau. Little is known concerning tsetse-trypanosome-human interaction in the region. Therefore, the ABM has the potential to enable exploration of contact risk within communities. Furthermore, with climate changes expected to occur in the near future, such as reduced annual rainfall, increased storm events and increased temperature [57][58], it is becoming increasingly important to understand how climate factors can affect tsetse populations, particularly in areas such as this, where increases in temperature could see the tsetse habitat spreading further up the valley to more populous areas.
This paper describes a new, seasonally sensitive ABM for rHAT/animal African trypanosomiasis (AAT), based on an earlier, non-seasonal model that was constructed using data derived from a detailed rHAT, AAT, and G. m. morsitans ecological survey, undertaken in 2013, in Eastern Province, Zambia [27]. Due to the fine spatial and temporal scales used to model the system, and the number of mechanisms incorporated (e.g., tsetse reproduction, tsetse feeding, human agent movements using real-world routines and pathfinding techniques [59]), the model was complex and its data inputs were numerous. As a result, only new data and modifications to the original model are described here. A detailed description of the original, non-seasonal model framework, and the data used to construct it, can be found in [27].
The previous iteration of the model included a longer pupal duration in males than in females, as suggested in the literature (e.g. [37,60]), and so for each larva deposited during the simulation, a 35 and 30 day pupal period was included for males and females, respectively, represented as a period of inactivity. However, pupation is known to be temperature sensitive with pupal periods decreasing with increasing temperature, a relationship observed by Phelps and Burrow’s laboratory experiments at constant temperatures [37]. Hargrove [41] utilised the data to present a near perfect fit for pupal duration at temperatures between 16°C and 32°C (r2 = 0.998) (see Fig 5), represented by Eq 2:
r=k31+e(a+bt),
pupalduration=1r,
Eq 2
Where: t = temperature, for males: a = 5.3, b = -0.24 and k3 = 0.053 and for females: a = 5.5, b = -0.25 and k3 = 0.057.
Given the excellent fit to the data and the large variation in pupal periods expected within the temperature range found in the study area (19°C = ~60 days, 28°C = ~20 days), variation in pupal duration with temperature is clearly an important factor to incorporate in the model.
The previous, non-seasonal ABM provided the majority of the methods and data used in the current version of the simulation, and so readers are referred to [27] for greater detail and only a summary is provided here. Census data were used to locate and initialise the human and animal populations living in the households shown in Fig 1. A sample of resource-seeking routines sorted by gender and age was taken in the field (see supplementary information of [27]), and a set of plausible paths from each village to each resource was created using a pre-processing A* pathfinding technique [59]. For tsetse, an estimate of the total apparent population size, density and distribution was provided. Four agent types were included in the ABM, together with an areal representation of wildlife. Humans, cattle, other domestic animals and tsetse used in the ABM were constructed as four separate classes, with populations modelled 1:1 with the data collected in the census (e.g. 16,024 human agents) and the estimated tsetse population discussed previously. Each class had its own initial information and storage structures for events that occurred through the simulation. The ABM was written in Python 2.7 using an object-oriented framework, and run on the Lancaster University High End Computing (HEC) Cluster, with all spatial data being processed using Quantum GIS 1.8.0.
The subsequent sections draw attention to any modifications between the original, non-seasonal modelling framework and the new ABM model, while also describing how the climatic drivers affecting the tsetse population were incorporated into the model.
The initial iteration of the model was split into 2,400 time-step (or tick) days, as the more frequent the tick, the smaller the jumps made by agents as the simulation updates, and the less chance of missing potential interactions. However, this method was restrictive in terms of memory usage and CPU time required to run just six months of the simulation. Further tests were carried out to establish how coarse the temporal resolution could be made before the number of simulated domestic host-vector contacts was reduced, and a greater daily probability of wildlife feed was required to maintain the tsetse population levels. It was established that 600 ticks per day (2.4 minutes per tick) allowed the simulation to progress with no obvious effect on human, cattle and other domestic animal bite numbers, while requiring a very similar daily wildlife bite probability to produce a stable tsetse population (37% chance per day of a hungry tsetse taking a wildlife bite, compared with 35% in the previous version). As a result, 600 ticks per day were used to produce the results of this investigation, which required approximately 4.5 GB of RAM per simulation run on the high performance machine, and 24 hours of CPU time per simulated year.
To capture the effect of seasonality on the tsetse fly population, daily temperature was calculated every 24 hours using the interpolation method discussed previously, and set as a global variable for the simulation.
For each female, once mated, the number of days since mating was compared with the birth interval calculated using Eq 1 and the daily temperature. If and when the number of days since mating exceeded the interval calculated on a given day, a pupa was deposited. A count of the number of days since mating was replaced with a count of the number of days since last offspring, and Eq 1 was used again on a daily basis (using the alternative constants for further births), until another birth occurred. This process was repeated for the duration of a female tsetse fly’s lifespan. There was an equal chance of each tsetse offspring being male or female, and each pupa was deposited in a bush area in which the female tsetse rested during the previous night.
A rolling average of the temperature that each pupa has experienced since birth was calculated and attributed to each individual. This temperature was used to determine each individual’s pupal duration, given that if a pupa’s age exceeded the pupal duration calculated using Eq 2, the pupa would emerge as a teneral fly. It was considered important to use a rolling average of temperature here as the length of a pupal period can span months with quite different temperatures. As described previously, rather than a single probability used to decide whether a pupa would die during its entire pupal period, a variable daily probability of pupal death was included, increased in some months to account for losses observed in the rainy season. Should the probability be exceeded for a pupa, that tsetse was removed from the simulation.
Death could result from pupal mortality, starvation, or if a tsetse fly exceeded the daily mortality rate calculated by sex, age and temperature (Eq 6, Figs 7 and 8). The mortality rate was calculated individually for each teneral and mature fly, and if the probability was exceeded, the tsetse was removed from the simulation. Starvation occurred if a tsetse tried and failed to feed before a given period of time had elapsed. The starvation element was more strict for teneral flies (3 days instead of 5 days) highlighting their increased vulnerability and reduced flight strength. In the previous version of the simulation, 75 teneral files were added to the simulation for the first 35 days to account for pupae deposited prior to the start of the simulation. As this version of the simulation started in August, and the simulated climate quickly became hostile for teneral flies as temperature increased, 500 teneral tsetse were required per day for the first 45 days, which is representative of average simulated pupal maturation rates during September as the simulation progressed (see Results). In the original model, in the absence of climatic factors, a scaling factor for adult fly mortality was required to offset fly starvation within the simulation. This value was set at 55%. Although this scaling factor is still required in this iteration of the model due to the same starvation element, the incorporation of temperature dependent mortality, and more detailed mechanisms for modelling pupae, has reduced the required level of scaling to 80%
To allow the model to initialise and stabilise, the simulation was run for a year before the results for this paper were produced, allowing a ‘burn-in’ period. For example, the results presented below are representative of years 2–4 of the simulation. 100 repeat simulations were used to produce the results presented here.
At the end of the three year simulation, a relatively stable population record was observed in both the male and female tsetse populations, with both exhibiting a double peak in response to the climatic driver (Fig 9).
Each year, until peak temperature was reached in October and November, the population slowly increased, with each gender’s population size increasing by approximately 2000 flies. Such population increases during this hot and dry season could be attributable to the absence of a boosted pupal mortality which is observable during the rainy season [61], with increasing temperatures having a greater impact in reducing pupal duration and the period between births, than increasing tsetse teneral and mature tsetse mortality. During the rainy season (November-April), this population gradually fell to an annual low, a result of peaks in pupal mortality at the start of the rainy season, and high temperatures causing increased mortality in the annual peak population of teneral flies (see Fig 10) (now emerged after a high period of births discussed previously—birth numbers can be seen in Fig 11).
During this period, with a reduced number of pupae to develop, and teneral tsetse to mature and start reproducing, the higher temperatures no longer aided a growth in population as there were fewer pupal maturations and birth rates to ‘accelerate’ (Fig 11) At the end of the rainy season, the tsetse population gained a small boost due to a plateau in temperature, and the drop in population slowed through the cool and dry season (May to July), although recovery did not start during this period as temperatures were too low to aid rapid repopulation of the tsetse, and the pupal population was still recovering (Fig 10).
Fig 12 presents the different possible modes of tsetse death included in the model, and how the rates varied as the simulation progressed. Non-starvation death represented the deaths attributable to the age-temperature dependent mortality model defined by Eq 6, and was consistently responsible for the largest number of daily deaths, peaking in the period of highest temperature with approximately 350 deaths per day. Unsurprisingly, given its temperature dependency, the mortality shape closely aligned to mean monthly temperature, except for a period in February and March after the pupal population was reduced by a period of high pupal mortality during the rainy season, resulting in a reduced teneral population and, therefore, fewer adult deaths. Deaths due to starvation followed closely the general pattern of population size, with teneral starvations being particularly low–likely a result of the low daily teneral population size (ranges between 100 and 400 –Fig 10) and the teneral tsetse population having the highest age-temperature dependent mortality rate. Using the Ackley and Hargrove model [61] for pupal mortality produced peaks prior to the rainy season and, to a lesser extent, after the wettest months (Fig 12).
The ratio of pupae to mature tsetse was approximately 2:1 at any given time, with the mature to teneral population ranging between 15:1 at the peak of population size and 25:1 when population sizes were generally lower.
Across the three year simulation, the approximate incidence rate for human and cattle rHAT infections was 0.355 per 1000 person-years (SE = 0.013), and 0.281 per 1000 cattle-years (SE = 0.025). There were 11 human infections each year on average (i.e. per year, per run), and 2 cattle infections.
Fig 13 illustrates how these infections clustered spatially and by season. The aggregate number of infections across all years and each of the 100 repeats was used to produce this heat map due to the low infection numbers. There was not much spatial variability through the seasons despite the variation in tsetse population size.
However, the number of infections reduced during the second half of the rainy season with the lowest density of infections observed during the cool and dry months. Two hotspots are visible in each of the seasons, each with elongated elements suggesting that frequently used paths were sources of interaction between vector and human host. This is possibly most visible in the north as east-to-west movement here could represent movement between villages and the river, a hypothesis which is given support by observations of infection by activity (Table 1) which suggest that in each season, water collection accounted for approximately 25% of human infections, second only to school trips which accounted for 49% to 51% of infections. No human infections were acquired whilst watering or grazing cattle, while the third highest number of infections occurred when farming. There was little variation in infections by activity between the seasons.
With the observed high proportion of infections coming from school trips, it is unsurprising that 5–10 year olds and 10–18 year olds had the highest infections rates (Table 2). Infection rates were generally lower in the cool and dry season, peaking in the hot and dry season.
Table 3 shows that the highest incidence rates were observed amongst immigrant tribes, with the only indigenous tribe (the Kunda) exhibiting one of the lowest infection rate across each time period, despite making up over 70% of the population.
Infection rates observed by gender and cattle ownership were comparable across time periods, with males and cattle owning households exhibiting marginally higher infections rates in comparison to females and households without cattle (Table 4).
Infections acquired and matured within the tsetse population fluctuated as the three year simulation progressed, with a small year-on-year increase in average infections both in the midgut and salivary gland (Fig 14). On average, the peak time of salivary gland infection development was at the beginning of the rainy season, which reflects the period of highest tsetse densities plus a time-lag for development of mature infections in the fly.
The first plausible individual-based model representation of a real world tsetse population was created allowing a simulation of the system over multiple years. The model was specified using temperature-dependent parameters derived from the literature, detailed human and animal information from acquired datasets, and expert opinion, and an estimate of the initial tsetse population size and distribution. For example, the pupal population which was completely emergent from the model (as no initial pupae data were inputted) corresponded with literature findings that pupae are comparatively difficult to find in the rainy season, and that the pupal population will be greater than that of the developed flies [38], unsurprising considering that the parameters suggest that pupae are ‘safer’ than teneral flies, pupal duration is at least 3 weeks, and a constant flow of developing pupae is required to replace teneral files which are dying or maturing. In addition, the ratio of female-to-male tsetse fluctuated around 2:1, a change from the simpler, non-seasonal model [27], but more in line with estimates in the literature [67], possibly as a result of running the simulation for longer, and with the addition of climate-driven parameters. The shape of the mature population was comparable to samples of tsetse collected in the region of the South Luangwa National Park (Regional Tsetse and Trypanosomiasis Control Programme (RTTCP) data reported in [22]), Eastern Province, Zambia [68], G. pallidipes in neighbouring Zimbabwe [61], and similar, yet less detailed, ABM studies [31].
The peak adult population of around 6500 flies suggests that the relatively crude technique used to extrapolate sample data from tsetse surveys for initial model construction (see [27] for more detail) produced a reasonable estimate with 5250 flies. Furthermore, the small teneral population observed is perhaps not a surprise, given that the teneral stage is a brief transition with a gradual input of developing pupae, and high mortality rates coupled with maturation to adult fly on first feed as outputs.
The decrease in pupal population during the rainy season, combined with a consistently small teneral population highlights how one or two years with a very hot and wet rainy season could have serious consequences for a tsetse population, with a reduction in pupal development during periods of high mortality, and high temperatures killing more teneral tsetse reducing the birth rate over subsequent months. Similarly, such a relationship could occur over the coming years in response to climate change, with IPCC reports suggesting that more extreme rainfall events could occur, along with a rise in temperature over the next 50 years (e.g. [57,58]). As a result, it is not surprising that some studies have suggested that certain tsetse fly populations could face extinction within the next 50 years [69]. Future studies will consider using the present model as the basis to test future climate change scenarios and examine the response in the tsetse population to such perturbations.
The model suggested similar incidence rates for rHAT infection in humans and cattle, which is likely to be a response to both the fact that the majority of the cattle were in households at the south of the transect, away from the tsetse zone (only approximately 550 of 2925 cattle were within close proximity of the tsetse zone) [27], and that humans were modelled to be much more active than cattle in the simulation, travelling more frequently away from the home. The latter point is corroborated by similar observations of human incidence rate in both cattle owning and non-cattle owning households, particularly as no human infections occurred while tending to cattle in the field or by the river. As with observations in the previous study, collecting water and school attendance provided the highest proportion of infections by some margin, and is likely to be in response to the high frequency of both trips within the simulation and, for schools, the longer distances travelled to a sparse resource, and the time of day of the trips coinciding with tsetse activity. In support of these simulated observations, a recent study of rHAT infections in Zambia found that almost half of the observed female infections were found in school-age children [70]. The data for males suggested fewer infections in children. This perhaps reflects that school attendance in the model is overestimated for the male population, and, in reality, young men may be needed to work to provide for the family at a younger age. The high incidence rates observed in immigrant tribes gives weight to the suggestion that as populations move down the plateau and into the valley, people are increasingly occupying marginal land, and increasing their exposure to the tsetse fly. As a result, future studies using the model will look to investigate how influxes of people into the region and the associated development affects the tsetse fly population in terms of habitat availability, but also how infection patterns respond to the perturbation of the system.
Six cattle infections were used to seed the model at the beginning of the simulation (along with five goats, one dog and two pigs), to reflect the estimated prevalence of T. b. rhodesiense in the sample of animals from the study transect. The model was also implemented with 10 humans infected at the simulation start, to take into account information from medical teams in the region, who suggested that there had been two reported cases in the past year, and the known high levels of under-reporting and under-detection in the area, and further afield (e.g. [22]). For example, one study suggested that levels of under-detection of rHAT could be as high as 12 cases for every one identified [71]. Furthermore, the recent study in Zambia found that, when a period of more active surveillance was adopted, the number of diagnoses increased dramatically, suggesting high levels of under-detection in the region. In addition, the investigation found that no action was taken by approximately one quarter of people showing symptoms of rHAT infection prior to diagnosis in the study, and less than half sought medical care from a health facility on first sign of symptoms [70]. As a result, given that there is no under-detection in a simulation, two cattle and 11 human infections on average per year appears plausible, especially when considering that there is currently no removal of infection from the simulation (and no reduction in activity when infected), creating a gradually increasing reservoir of infection, and an increase in tsetse infections (Fig 14).
Despite extensive effort to incorporate seasonality accurately into the simulated system, there are some omissions which were largely unavoidable here, but which should be noted. Firstly, in reality, the spatial distribution of tsetse will change through the seasons, with tsetse concentrated in the dense woodland vegetation in the hot dry season, and more widely dispersed in the wet and cool seasons since tsetse use microhabitats to evade extremes in temperature [60,72]. Using an interpolated temperature gradient across a study area through time may allow this behaviour to be simulated, although there would be limitations as temperatures would not reflect sheltered areas utilised by tsetse. As a result, such an implementation should be used in conjunction with a variable land classification, highlighting changes in vegetation with seasons. In addition, no data were available on how human movements vary seasonally in this region at the temporal resolution being modelled, and therefore, the daily routines used are consistent through the year. Finally, it is understood that maturation rate and transmissibility of trypanosomes in tsetse varies with temperature [60], with early work in Zambia suggesting that higher trypanosome infection rates occurred in G. morsitans in the hot season than in the cold season [54]. However, very little research has been carried out on this subject and, within this study, transmission rates should be low enough for this to have little impact.
For the first time in the field of rHAT transmission research, data produced and relationships identified in different studies, focusing on different aspects of the tsetse life-cycle and tsetse-climate interactions, have been incorporated into a single detailed ABM, creating a plausible, stable model, which can ultimately produce a reasonable estimate for transmission rates. While providing an element of validation to these individual entomological studies, through the production of tsetse population curves which closely follow those produced from data collected in the nearby South Luangwa National Park, the model represents a step towards a greater understanding of disease transmission for rHAT in this case, while also being adaptable to gHAT foci in the future. As with all models, the ABM is not without its limitations, for example, variability in tsetse feeding behaviour and preference has been incorporated, but at a basic level. However, through working towards an accurate model representation of the disease landscape one can expect to achieve a greater understanding of the rHAT transmission system, which in turn can help the devising of spatially and temporally targeted mitigation strategies in the future, to help those in need with sustainable solutions, and are more appropriate for spatially marginal communities susceptible to neglected tropical diseases.
The dynamics of a tsetse population are difficult to model due to difficulties in acquiring data, and the complexity of the system, but are important to understand due to their importance in rHAT transmission. Gaining a greater understanding of tsetse population dynamics may lead to greater understanding of rHAT transmission and aid future mitigation strategies. This paper presented the first seasonally-varying rHAT transmission model, defined at a fine resolution and modelling directly individual flies, with the full tsetse life cycle as a sub-component. By incorporating numerous parameters estimated from the literature, from data and from expert opinion into such a detailed model, a range of outputs were created which can be used by scientists to analyse and evaluate our current understanding of tsetse fly dynamics and the rHAT disease transmission system, and by decision-makers to investigate alternative mitigation strategies. In its current state, including seasonally varying effects, the model lends itself to modelling future scenarios, including insecticide application and other vector control strategies, the incorporation of a changing climate, the effects of landcover change and human development adjacent to, and within, the biodiverse tsetse habitat.
|
10.1371/journal.pntd.0000429 | Australia's Dengue Risk Driven by Human Adaptation to Climate Change | The reduced rainfall in southeast Australia has placed this region's urban and rural communities on escalating water restrictions, with anthropogenic climate change forecasts suggesting that this drying trend will continue. To mitigate the stress this may place on domestic water supply, governments have encouraged the installation of large domestic water tanks in towns and cities throughout this region. These prospective stable mosquito larval sites create the possibility of the reintroduction of Ae. aegypti from Queensland, where it remains endemic, back into New South Wales and other populated centres in Australia, along with the associated emerging and re-emerging dengue risk if the virus was to be introduced.
Having collated the known distribution of Ae. aegypti in Australia, we built distributional models using a genetic algorithm to project Ae. aegypti's distribution under today's climate and under climate change scenarios for 2030 and 2050 and compared the outputs to published theoretical temperature limits. Incongruence identified between the models and theoretical temperature limits highlighted the difficulty of using point occurrence data to study a species whose distribution is mediated more by human activity than by climate. Synthesis of this data with dengue transmission climate limits in Australia derived from historical dengue epidemics suggested that a proliferation of domestic water storage tanks in Australia could result in another range expansion of Ae. aegypti which would present a risk of dengue transmission in most major cities during their warm summer months.
In the debate of the role climate change will play in the future range of dengue in Australia, we conclude that the increased risk of an Ae. aegypti range expansion in Australia would be due not directly to climate change but rather to human adaptation to the current and forecasted regional drying through the installation of large domestic water storing containers. The expansion of this efficient dengue vector presents both an emerging and re-emerging disease risk to Australia. Therefore, if the installation and maintenance of domestic water storage tanks is not tightly controlled, Ae. aegypti could expand its range again and cohabit with the majority of Australia's population, presenting a high potential dengue transmission risk during our warm summers.
| Current and projected rainfall reduction in southeast Australia has seen the installation of large numbers of government-subsidised and ad hoc domestic water storage containers that could create the possibility of the mosquito Ae. aegypti expanding out of Queensland into southern Australian's urban regions. By assessing the past and current distribution of Ae. aegypti in Australia, we construct distributional models for this dengue vector for our current climate and projected climates for 2030 and 2050. The resulting mosquito distribution maps are compared to published theoretical temperature limits for Ae. aegypti and some differences are identified. Nonetheless, synthesising our mosquito distribution maps with dengue transmission climate limits derived from historical dengue epidemics in Australia suggests that the current proliferation of domestic water storage tanks could easily result in another range expansion of Ae. aegypti along with the associated dengue risk were the virus to be introduced.
| Aedes (Stegomyia) aegypti (Linneaus) is an important vector of dengue and other arboviruses. Despite its limited flight dispersal capability [1],[2], its close association with humans and its desiccation-resistant eggs have facilitated many long distance dispersal events within and between continents, allowing it to expand its range globally from its origin in Africa. Its global emergence and resurgence can be attributed to factors including urbanisation, transportation, changes in human movement, and behaviour, resulting in dengue running second to malaria in terms of human morbidity and mortality [3],[4]. Global historical collections and laboratory experiments on this well studied vector have suggested its distribution is limited by the 10°C winter isotherm [5], while a more recent and complex stochastic population dynamics model analysis suggests the temperature's limiting value to be more towards the 15°C yearly isotherm [6].
While historical surveys in Australia have indicated that Ae. aegypti occurred over much of the continent (see Fig. 1), its range has receded from Western Australia, the Northern Territory and New South Wales (NSW) over the last 50 years. It is now only found in Queensland [7],[8], although recent incursions into the Northern Territory have required costly eradication strategies [8]. The significant reduction in vector distribution has been attributed to a combination of events including the introduction of reticulated water, which reduced the domestic water storage requirements of households that had provided stable larval sites [7],[9], as well as the removal of the railway-based water storage containers hypothesised as being responsible for the long distance dispersal events of Ae. aegypti into rural regions in NSW via steam trains [7],[10].
Today, epidemic dengue is limited to regions of Queensland where Ae. aegypti is extant, and the frequency of outbreaks has increased constantly over the past decade [11]. Historically, epidemics of dengue were recorded in northern Queensland in the late 1800s and in southeast Queensland in 1904–05 [10]. Dengue epidemics in 1926, 1942 and 1943 all extended from Queensland south into NSW, stopping only on the arrival of winter [12]. Derrick and Bicks [12] found that these dengue epidemics ceased when the outside temperature reached a wet bulb isotherm of between 14–15°C and suggested that a parameter of 14.2°C mean annual wet bulb isotherm (TW) best represented the limiting parameter for the 1926 epidemic.
The current drying of southeast Australia has placed this region's urban and rural communities on escalating water restrictions, with anthropogenic climate change forecasts suggesting that this drying trend will continue [13]. To mitigate against this regional drying effect and the stress it places on domestic water supply, state government rebate programs have been initiated to encourage the installation of large (>3000 L) domestic water tanks in towns and cities throughout this region. Data from the Australian Bureau of Statistics [14] records that in 2006, 20.6% of all Australian household dwellings had rainwater tanks.
Given the expansion of domestic rainwater tanks in southern Australia, and assuming these domestic water tanks can provide oviposition sites, we ask this question: can climate be assessed to determine the distributional limits of Ae. aegypti and dengue in Australia? We first use a genetic algorithm to develop ecological niche models for the distribution of Ae. aegypti in Australia (using data points drawn from both historical and contemporary collection sites) and evaluate the potential distributional limits of Ae. aegypti in Australia under today's climate and in future projected climate change scenarios. We map these limits in relation to published experimental and theoretical projections of Ae. aegypti's temperature limits and then compare all projections to dengue transmission climate limits obtained from epidemiological studies of historical dengue epidemics in Australian. We find that human adaptation to climate change – through the installation of large stable water storage tanks – may pose a more substantial risk to the Australian population than do the direct effects of climate change. Additionally, we find that using point occurrence data and environmental parameters of climate and elevation to map the distribution of Ae. aegypti in Australia prove deceptive and require interpretation as some Ae. aegypti collection sites exist outside our ecological niche models and both theoretical cold temperature limits. This suggests that Ae. aegypti's domestic behaviour – with a lifecycle based around human habitation that includes blood-feeding and resting indoors as well as egg laying in artificial containers around houses – plays an influencing role on distribution.
Coordinates for a total of 234 Ae. aegypti collections sites are described in Table S1. Historical collection sites were compiled [7],[9],[15],[16]. Contemporary collection sites were regarded as those collected since 1980 because most country towns had moved to reticulated water, steam powered trains had been replaced by diesel, and the common railway station water-filled fire buckets were removed [9],[17],[18]. Contemporary sites also include collections made between 1990 and 2005 from southeast Queensland (P. Mottram, unpub. data), and the Northern Territory (P. Whelan, unpub. data).
Raster ASCII grids were generated for Australia at a spatial resolution of 0.025° (approximately 2.5 km) for eight climate variables plus elevation. These included annual mean rainfall and annual mean temperature produced by BIOCLIM using the ANUCLIM software package [19] as well as mean values of maximum temperatures and minimum temperatures for the months of January and July produced by the ESOCLIM component of ANUCLIM. This procedure involved the use of monthly mean climate surface coefficients, generated by the thin plate smoothing spline technique ANUSPLIN [20] from Australian Bureau of Meteorology climate data between 1921 and 1995 [21]. The geographic coordinates of the meteorological stations were used as independent spline variables together with a 0.025° digital elevation model (DEM) for Australia generated with ANUDEM [22] which acted as a third independent variable. As atmospheric moisture is known to be an important factor in terms of the survival and longevity of adult mosquitoes, mean values of dewpoint for January and July were generated with ESOCLIM to provide this.
A further series of ASCII grids were generated from climate change scenarios using OzClim version 2 software [23],[24] at a spatial resolution of 0.25° (approximately 25 km). The scenarios used for this study were for 2030 and 2050 using CSIRO: Mk2 Climate Change Pattern with SRES Marker Scenario A1B and mid climate sensitivity. The output variables corresponded to the predicted change from the base climate for the rainfall and temperature parameters generated with ANUCLIM.
This version of OzClim outputs vapour pressure rather than dewpoint as a measure of atmospheric moisture. For the present study vapour pressure grids for the predicted change from base climate for January and July were generated and the grid cell values were converted to dewpoint by applying the inverse of Tetens' equation [dp = (241.88×ln(vp/610.78))/(17.558−ln(vp/610.78)]. This mathematical procedure was implemented with the use of ImageJ software (publicly available at http://rsbweb.info.nih.gov/ij) together with the raster operations of TNTmips (MicroImages Inc., Lincoln, Nebraska).
The environmental layers used for climate change modelling were prepared by resampling the OzClim outputs to the geographical extents and grid cell size of the ANUCLIM grids using TNTmips. The resampled outputs were then added to the corresponding ANUCLIM base climate layers to produce the environmental layers predicted for the chosen climate change scenarios.
DesktopGarp version 1_1_6 [25] was used for ecological niche modelling in a manner similar to our earlier studies [26]. Models derived from the historical climate data were generated using the record sites for Ae. aegypti as inputs together with the eight base climate layers and elevation (the ANUDEM generated DEM is described above) to model the range of Ae. aegypti. Species record sites and the climate change layers for 8 environmental parameters were derived from the climate change scenarios for 2030 and 2050 as well as the elevation layer. We utilized the medium sensitivity which corresponds to a global warming of 2.6°C for a doubling of CO2 from 280 ppm to 560 ppm [27]. The GARP procedure was implemented using half of the species record sites as a training data set for model building and the other half for model testing. Optimization parameters included 100 models for each run with 1000 iterations per model and 0.01 convergence limits. The best subsets procedure [28] was used to select 5 models which were added together using TNTmips to produce predicted range maps for each species.
Previous studies of the distributional limits of Ae. aegypti were used to develop distribution maps for Australia. Christophers [5] hypothesised a climate limit of 10°C winter isotherm based on historical global collection data and laboratory-based experiments. We also evaluated the hypothetical limit from Otero and colleagues [6], who used a complex stochastic population model that incorporates the lifecycle parameters of Ae. aegypti to suggest a 15°C annual mean isotherm. Both these values were incorporated into distributional maps of Australia using TNTmips.
Dengue transmission maps were developed using data from historical dengue outbreaks in Australia [12]. This work found that these dengue epidemics ceased when the outside temperature reached 14–15°C wet bulb isotherm and that a single parameter of 14.2°C annual mean wet bulb isotherm (TW) best approximated the limit of the 1926 epidemic – probably as a result of reducing the mosquitoes' feeding activity and the ability of the virus to replicate. This 14.2°C annual mean wet bulb isotherm value was mapped onto Australia for the current climate using TNTmips and three seasonal increments: the annual mean, the warmest quarter (December–February), and the coolest quarter (June–August).
Distribution sites for Ae. aegypti in Australia (234 sites) were collated and displayed in a single map using GPS coordinates (Table S1 and Fig. 1). Ecological niche models were built with desktop GARP to produce a best subset model that showed agreement with the full complement of Ae. aegypti collections in Australia (Fig. 2A). In this projection, much of northern, eastern and southeast Australia was projected to present a suitable niche. This model closely tracks an annual rainfall pattern of less than 300 mm. However, the excluded region around central Australia included two Ae. aegypti positive collection sites (Meekatharra in central Western Australia and Boulia in Queensland): both collection localities are small regional centres on main inland transport routes.
The projected climate change scenario for 2030 produced distributional models with small expansions of the base model envelope, mostly evident in southern Australia (Fig. 2B). Likewise the 2050 model (Fig. 2C) extended the 2030 trend, resulting in a reduced niche in north-west Australia's Pilbara region while parts of central Australia opened up as a potential niche.
The temperature limit parameters of 10°C winter isotherm [5] and 15°C annual isotherm [6] were used to build theoretical isotherm limits for Ae. aegypti in Australia (Fig. 3). Figure 3A shows a 10°C winter isotherm limit base map for the current climate and OzClim projections were then generated for 2030 and 2050 by adding the projected changes to this base map (3B and 3C respectively). The 15°C annual isotherm limits were similarly generated using a base map and adding the OzClim changes. Both the 10°C (average winter) and 15°C (average annual) limits incorporate the major state capitals cities – Brisbane, Sydney, Adelaide and Perth. When these isotherm limits were subjected to the climate change scenarios for 2030 and 2050, the projection expanded to include the other mainland state capital, Melbourne (Fig. 3B and 3C).
Several Ae. aegypti collection sites occurred well within the two theoretical cold climate limits. Table 1 details six Ae. aegypti collection sites as examples where the annual mean temperature and the mean temperature for July (calculated as (mintemp+maxtemp)/2) fall below the theoretical values and range from 12.4–15.4°C and 5.2–7.6°C respectively.
Derrick and Bicks [12] suggested that dengue transmission stopped between the 15°C and 14°C TW isotherm and suggested that a 14.2°C TW annual mean isotherm best approximated the temperature limit for transmission in the 1926 dengue epidemic. We applied this isotherm to Australia for the annual mean isotherm (Fig. 4A) as well as the warmest quarter isotherm (summer; December–February, Fig. 4B) and the coldest quarter isotherm (winter; June–August, Fig. 4C). These climate limit maps indicate that if the vector could re-establish itself throughout its former range then much of northern tropical Australia would be receptive to dengue transmission year round and transmission would be possible throughout most of Australia during the summer months.
Can the historical distribution of Ae. aegypti in Australia provide an insight into the potential distribution potential of this mosquito? Using 234 different spatial data points generated from historical and contemporary collections of Ae. aegypti in Australia, we developed ecological niche models to hypothesise the potential range expansion of this mosquito under today's climate and under future climate change scenarios for 2030 and 2050 using OzClim mid sensitivity values that correspond to a global warming of 2.6°C for a doubling of CO2 from 280 ppm to 560 ppm [27]. In Australia general warming estimates are approximately 1.0°C by 2030 and 1.2 to 2.2°C by 2050, the latter values being dependent on CO2 emissions. While rainfall (outside of far north Australia) is estimated to decrease by 2% to 5%, southern Australia is projected to encounter a 5% reduction in rainfall [13]. Our GARP model for current climate suggested that Ae. aegypti could potentially coexist with over 95% of the Australian population and this distribution did not change significantly, with regard to the Australian population distribution, under either the 2030 and 2050 climate change scenarios.
Only the highly arid central Australian region was excluded from the projection (annual rainfall less than 300 mm). The GARP model did not show southern cold climate thermal limits in Australia, probably due to the presence of several Ae. aegypti collection sites from inland New South Wales that show cool climate parameters. We then mapped two theoretical cool climate limits across Australia – the 10°C winter (July) isotherm [5] and the 15°C annual mean isotherm [6]. Of these two isotherm limits the 15°C annual mean isotherm appeared more representative of the known distribution of Ae. aegypti in Australia, although collection sites did exist outside these temperature isotherm limits.
It remains unknown if the cold climate tolerant populations were breeding in the warmer months and surviving the colder winter months as eggs [29], or were surviving as larvae. With regard to these questions, observations have been recorded of viable Ae. aegypti larvae in ice encrusted water [5],[7], while experiments have suggested that a water temperature of 1.0°C can be lethal over 24 hours, but larvae can be viable at a constant 7.0°C for over a week [5]. At the other temperature extreme, laboratory experiments show that Ae. aegypti larvae perish when the water temperature exceeds 34°C while adults start to die off as the air temperature exceeds 40°C [5]. Domestic water tanks in Australia contain thousands of litres of water that would – in combination with the mosquitoes' domestic (indoor) nature – provide a buffer to temperature extremes and assist mosquito survival in what may appear unsuitable environments. For example, Ae. aegypti exists and transmit dengue in India's Thar desert townships in north-western Rajasthan, where the mosquito utilises household pitchers and underground cement water tanks. [30].
The incongruence between the temperature limits and our ecological niche models highlights the difficulties of using what are essentially sophisticated climate pattern matching procedures to study an organism with a biology and ecology strongly influenced by human activity. Fortunately, we can directly compare our GARP model with a new mechanistic model of the same organism over the same environment [31]. This mechanistic model utilises biophysical life processes parameters such as the effects of climate on reproduction and larval development. Larval development in both rainwater tanks and smaller containers were assessed and the potential distribution of Ae. aegypti was projected across Australia. Projections using rainwater tanks larval development resembled our GARP model for Northern and central Australia, but unlike our projections, a southern cold climate thermal limit was identified which was actually lower than the published parameters displayed in Fig. 3 [5],[6]. Apart from showing the clear advantage of a bottom-up approach for modelling this mosquito, this study supports the hypothesis that domestic rainwater tanks contributed for the historical southern distribution of Ae.aegypti in Australia.
Humans not only facilitate long distance dispersal events for this mosquito, co-habitation with humans can provide thermal buffers to the outdoor climate as adults rest indoors, and domestic rainwater tanks can provide stable oviposition sites. When the theoretical distributions (GARP models and temperature limits) and actual Ae. aegypti distributions are viewed alongside the expansion of domestic water tanks underway in Australia, a trend emerges where Ae. aegypti could potentially exist year-round in today's climate throughout the southern Australian mainland. This potential distribution includes the metropolitan areas of Brisbane (pop 1.8 million), Sydney (pop 4.2 million), Adelaide (pop 1.1 million) and Perth (pop 1.5 million). Additionally the climate change temperature limit projections for the mid scenario 2050 see this range expand to include Melbourne (pop 3.6 million). The addition of a theoretical dengue virus transmission limit parameter (we used a 14.2°C wet bulb isotherm) suggests an overlapping dengue risk in many of Australia's metropolitan regions during the summer months (December–February).
The potential for dengue virus introduction to these regions through travellers from endemic regions (including north Queensland) during summer presents a transmission risk that can be inferred by the current incidence of imported and endemic cases of dengue in Australia – many of which enter Australia through national and international transport nodes. For example, for the year to June 2008 there were 250 dengue notifications for Australia, of which 113 came from Queensland (most via local transmission), 72 from NSW, 15 from NT, 12 from SA, 8 from VIC, and 28 from WA. Notifications from New South Wales, South Australia, Victoria and Western Australia exceeded the five-year mean in each jurisdiction suggesting that the frequency of dengue is increasing [32].
Understanding the relationship between climate and dengue transmission is difficult because non-linear relationships exist between the daily survival of Ae. aegypti, the extrinsic incubation period (EIP) of the virus, temperature and humidity [33]–[35]. Forecasted regional warming in Australia may lengthen and intensify the dengue transmission season by shortening the mosquitoes' EIP, although it is important to note that dengue epidemics appear to be more strongly influenced by intrinsic population dynamic (epidemiological) processes than by climate [36]. Even so, any temporal extension effect in the transmission season will follow the expansion of potential larval sites that is now underway in Australia. Thus, while the issue of regional warming is important, the expansion of large rainwater tanks throughout urban regions of Australia is at present a prevailing human adaptation with more immediate possibilities for changing vector distributions in Australia than the direct warming effects projected by anthropogenic climate change scenarios. Whether southern Australia's current drought is due to the region's natural climate variability or part of a changing climate pattern, will continue to be debated by some. Nonetheless, it is important to avoid the cycle where human changes in water storage result in an Ae. aegypti range expansion followed by dengue epidemics seeded by viremic travellers [4],[37]. Additionally, domestic water storage can sustain Ae. aegypti populations (and dengue transmission) in regions not normally suitable for its survival [38], while active government and community contributions can remove established Ae. aegypti populations (and dengue) from areas where it has been endemic [39] – and both of these are human modifications.
In Australia, ineffectively screened domestic rainwater tanks have been identified as key containers with respect to Ae. aegypti productivity [40],[41]. The introduction of reticulated water systems in towns and cities throughout Australia is believed responsible for a major range contraction of Ae. aegypti over the last 50 years. This trend may now be reversed as humans adapt to climate-change-induced drought conditions – the increased use of domestic water storage in tanks could deliver stable primary larval sites into urban neighbourhoods. In Queensland's capital city, Brisbane – which is currently Ae. aegypti free – severe water shortages resulted in escalating water restrictions with an eventual prohibition on the use of all outside reticulated water outlets (November 2007–July 2008) and 75,000 domestic water tanks being installed by late 2007. This number of tanks represents approximately 21% of households with reticulated water in the Brisbane area (F. Chandler, Brisbane City Council, pers. comm.). Additionally, ad hoc uncontrolled water tanks are now also commonly being used to store rainwater, adding to the potential surfeit of stable breeding sites around Australia that are likely to facilitate the expansion risk of Ae. aegypti into urban areas. It is unlikely that any of these water storage tanks – government approved or not – will be maintained sufficiently to prevent mosquito access in the long term.
The flight range for Ae. aegypti is understood to be generally small: mark-release recapture experiments show them to have a flight range of only hundreds of metres [42]–[44]. However, these estimates are limited in time and space, being derived from a snapshot of one or a few gonotrophic cycles which take place in the context of an abundance of ovipositing sites. Longer distance flight range dispersal may be more common, especially when ovipositing sites are rare, but this is difficult to quantify [45],[46]. Human mediated long distance dispersal events are mostly responsible for Ae. aegypti movement: their highly domestic nature and desiccation-resistant eggs facilitate successful movement via human transport routes. Surveys in Queensland in the 1990s [17] and 1990–2005 (P. Mottram, unpublished) reveal Ae. aegypti collections from over 70 townships and this number is likely an underestimate. As the numbers of individuals and populations of Ae. aegypti increase in Queensland towns, the incursion risk beyond these regions via human-induced long distance dispersal events also increases, and with the presence of new stable oviposition sites growing, the expansion of this dengue vector must now be expected.
Operations to remove Ae. aegypti incursions are resource-heavy, often requiring both government legislation and widespread community cooperation to reduce adult mosquito populations. A recent example from a 2004 incursion of Ae. aegypti into the small Northern Territory town of Tennant Creek (pop 3,200) from Queensland resulted in a two-year eradication campaign that required 11 personnel and cost approximately $1.5 million and was achieved in 2006 [8].
Determining the potential distribution of Ae. aegypti in Australia using climatic parameters can be problematic and in this case produced results that neither fully match the known distribution, nor reveal cold climate limits in Australia. Reasons for this may exist in the difficulty of relating the point occurrence data of a species' distribution that is closely tied to humans – unlike native mosquito species in Australia where GARP models appear more representative of known distributions [26],[47]. We must also consider the limited climatic parameters available through the OzClim climate scenario generator that reduced the GARP modelling to a subset of environmental parameters that may have little influence on the organism. Because the GARP models showed no cold temperature limits for Ae. aegypti in Australia, we also assessed two published theoretical cold temperature limits across Australia. These temperature limit projections also could not contain all collection sites, which may suggest that in Australia, climate - and in particular temperature - plays a less important role in determining the range of this species due to a combination of its intimate relationship with humans and our propensity to store water. This is where the use of statistical approaches and point occurrence data to evaluate species' distribution may be weak and integrating life processes parameters such as the effects of climate on reproduction and larval development may be more practical and informative.
If it is an assumption that burgeoning domestic water tanks will provide stable larval sites for Ae. aegypti, then the synthesis of our GARP modelling, the theoretical climate limits and the historical distribution of this mosquito strongly suggest that a distributional expansion is possible and could expose the majority of Australia's population to this dengue vector. Additionally, viewing this synthesis of Ae. aegypti in Australia with dengue transmission climate limits obtained from historical Australian dengue epidemics suggests a real risk of dengue transmission occurring in regions ranging well beyond north Queensland during the summer months.
We conclude that if the installation and maintenance of domestic water storage tanks is not tightly controlled today, Ae. aegypti could be spread by humans to cohabit with the majority of Australia's population, presenting a high potential dengue transmission risk during our warm summers.
|
10.1371/journal.pntd.0004519 | Sandfly-Borne Phlebovirus Isolations from Turkey: New Insight into the Sandfly fever Sicilian and Sandfly fever Naples Species | Southern Anatolia in Turkey at the border with Syria, where many refugee camps are settled, is endemic for sandfly-borne leishmaniasis. Sandfly-borne phleboviruses are also known to circulate in this region, although their relevance in terms of medical implications is virtually unknown. Therefore, the specific objectives of our study were firstly to identify isolate and characterise potentially pathogenic phleboviruses in sandflies; secondly to determine the complete genomic sequence of any viruses that we were able to isolate; and thirdly, to further our understanding of the potential medical importance and epidemiological significance of these viruses. To achieve these objectives, we organised field campaigns in 2012 and 2013. Two new phleboviruses (Toros and Zerdali viruses) were isolated and characterized by complete genome sequencing and phylogenetic analyses. Toros virus was genetically most closely related to Corfou virus within the Sandfly fever Sicilian group. Zerdali virus was most closely related to Tehran virus within the Sandfly fever Naples species. Although these new viruses belong to genetic groups that include several human pathogens, it is not yet clear if Toros and Zerdali viruses can infect humans and cause disease such as sandfly fever. Consequently, the availability of these genetically characterized infectious viruses will enable seroprevalence studies to establish their medical importance in this region and to assist the health agencies to develop appropriate and effective disease control strategies.
Many studies have presented virus sequences which suggest the existence of a variety of putative new phleboviruses transmitted by sandflies in the Old World. However, in most of these studies, only partial sequences in the polymerase or the nucleoprotein genes were characterised. Therefore to further our understand of the presence and potential medical importance of sandfly-borne phleboviruses that circulate in southern Anatolia, we initiated field campaigns in 2012 and 2013 designed to identify, isolate and characterise phleboviruses in sandflies in this region
An entomological investigation encompassing 8 villages in Adana, Mediterranean Turkey was performed in August and September 2012 and 2013. A total of 11,302 sandflies were collected and grouped into 797 pools which were tested for the presence of phleboviruses using specific primers for RT-PCR analysis and also cell culture methods for virus isolation. Seven pools were PCR positive, and viruses were isolated from three pools of sandflies, resulting in the identification of two new viruses that we named Zerdali virus and Toros virus. Phylogenetic analysis based on full-length genomic sequence showed that Zerdali virus was most closely related with Tehran virus (and belongs to the Sandfly fever Naples species), whereas Toros virus was closest to Corfou virus.
The results indicate that a variety of phleboviruses are co-circulating in this region of southern Anatolia. Based on our studies, these new viruses clearly belong to genetic groups that include several human pathogens. However, whether or not Toros and Zerdali viruses can infect humans and cause diseases such as sandfly fever remains to be investigated.
| We provide evidence that sandfly-borne phleboviruses belonging to 3 distinct genetic and phylogenetic groups (Sandfly fever Naples virus [SFNV], Sandfly fever Sicilian virus [SFSV], and Salehabad virus [SALV]) co-circulate in Adana city, in Mediterranean Turkey. While Adana virus was recently described as a new member of the SALV species, Zerdali and Toros viruses are described here as new phleboviruses genetically closely related to SFNV and SFSV, respectively. In this study, isolated and characterised these two new viruses by determining their complete genome sequence and by phylogenetic analysis. This study demonstrates that 3 distinct viruses can co-circulate in the same geographic area and based on their phylogenetic relationships and association with sandflies are likely to be transmitted by these arthropod vectors. Our molecular and phylogenetic data are important for establishing group-specific molecular detection assays in order to further understand of the possible impact of these viruses in animal and human health in this region of Turkey.
| The genus Phlebovirus (family Bunyaviridae) currently contains 9 viral species Sandfly fever Naples (SFNV), Salehabad (SALV), Rift Valley fever (RVFV), Uukuniemi (UUKV), Bujaru (BUJV), Candiru (CRUV), Chilibre (CHIV), Frijoles (FRIV) and Punta Toro (PTV) including 33 distinct serotypes, and 32 tentative serotypes as defined in the 9th Report of the International Committee on Taxonomy of Viruses (ICTV) [1]. Nevertheless, the past decade has witnessed the discovery of many new phleboviruses that remain to be classified: some are transmitted to vertebrates by sandflies (Fermo (FERV), Granada (GRAV), Punique (PUNV)) [2, 3, 4], others by ticks (Heartland (HRTV), Hunter island group (HIGV)) [5, 6], whereas some do not have recognised vectors and appear to be transmitted directly between vertebrates (Malsoor (MALV), Salanga (SGAV)) [7, 8].
In the Old World, sandfly-borne phleboviruses are transmitted between vertebrates mainly by female sandflies (genus Phlebotomus) when they take a blood meal. Some Old World sandfly-borne phleboviruses may cause self-limiting febrile illnesses (sandfly fever) or neuro-invasive infections. They are widely distributed in the Mediterranean Basin, in Africa, in the Indian subcontinent, in the Middle-East, and in far-eastern former USSR republics [9]. Annually, Toscana virus (TOSV), a serotype of SFNV is the leading cause of meningitis from May to October in central Italy [10] and one of the most prevalent human pathogenic phleboviruses in other southern European countries.
Forty years ago, seroprevalence studies showed that Sandfly fever Sicilian virus (SFSV) and SFNV were present in the Mediterranean and Aegean regions of Turkey [11, 12]. Recently, serological investigations were carried out in the Mediterranean, Aegean, and Central Anatolian regions, where outbreaks have occurred and circulation of SFSV and a SFS-like virus (Sandfly Fever Turkey virus (SFTV)) were reported [13,14,15]. The presence of TOSV was confirmed serologically and through RNA detection and sequencing [16,17,18,19]. Despite the publication of many articles, virus isolations were reported only for SFTV from a patient [13] and Adana virus (ADAV) [20] from sandflies. To further understand of the dynamics of sandfly-borne phleboviruses and sandfly fever in the Mediterranean region in the vicinity of Adana city, we organized sandfly trapping campaigns.
Sandflies were captured during August and September in 2012 and in 2013 in Adana city located in the Mediterranean region of Turkey (Fig 1) using CDC Miniature Light Traps as previously reported [21]. Live sandflies were pooled according to sex, trapping site and day of capture, with up to 30 individuals per pool and placed in 1.5mL tubes, and stored at -80°C. In order to reduce the time between capture and storage and therefore to increase the likelihood of virus isolation, morphological identification of the sandflies was not performed.
Sandfly pools were processed as previously described [22] in a final volume of 600μL, of which 200μL were used for total nucleic acid extraction using the Virus Extraction Mini Kit the BioRobot EZ1-XL Advanced (both from Qiagen). Elution was performed in 90μL of extraction buffer of which 5μL were used for RT-PCR and nested-PCR assays using primers targeting the polymerase gene and the nucleoprotein gene as previously described [20, 23, 24]. PCR products of the expected size were column-purified (Amicon Ultra Centrifugal filters, Millipore) and directly sequenced.
Two real-time RT-PCR assays (Rt-RT-PCR) were designed for specific detection of the new strains in the N gene: TORV-N-FW (AACTCTGACTCGTGTGGCTG), TORV-N-REV (GCCTTGGGTATGTCTGACCA), and TORV-N-Probe (6FAM-AGGCAATAGAAGTTGTGGAGAAC-TAMRA); ZERV-N-FW (ACTTCCTGTTACTGGAACAACAAT), ZERV-N-REV (CCATGAGCATCTGCAATAACTTC), and ZERV-N-Probe (6FAM-ATGATGCATCCTAGTTTTGCAGGA-TAMRA). Reaction conditions and cycling programs were previously described [20].
Fifty μL (derived from sandflies ground in the 600μL of EMEM as aforementioned) were inoculated onto a 12.5 cm2-flask of Vero cells, incubated at room temperature for 1 hr, and supplemented with 3mL of EMEM (5% FBS, 1% Penicillin/Streptomycin, 1% L-Glutamine 200 mM, 1% Kanamycin, and 3% Fungizone). The flasks were incubated at 37°C in 5% CO2 atmosphere and examined daily for cytopathic effects (CPE).
For detailed characterisation, Zerdali virus (ZERV) passage 5, Toros virus (TORV) strain 292, passage 3, and TORV strain 213, passage 7 were subjected to complete genome characterisation using Next Generation Sequencing (NGS). Briefly, 140μL of infectious cell culture supernatant medium was incubated with 30 U of Benzonase (Novagen 70664–3) for 7 hr at 37°C.This material was then purified using the Viral RNA Mini Kit (Qiagen). Tagged random primers for reverse transcription (RT) and tag-specific and random-primers were used for PCR (Applied Biosystems). The resulting PCR products were purified (Amicon Ultra Centrifugal filters, Millipore); 200ng of DNA were used for sequencing using the Ion PGM Sequencer (Life Technologies SAS, Saint Aubin, France). NGS reads, of 30 nucleotides minimum length, were trimmed using CLC Genomic Workbench 6.5, with a minimum of 99% quality per base and mapped to reference sequences. Parameters were set such that each accepted read had to map to the reference sequence for at least 50% of its length, with a minimum of 80% identity to the reference. From the contigs obtained, viral sequences were identified by best BLAST similarity against reference databases. Sequence gaps were completed by amplification and sequencing overlapping regions using either Sanger sequencing or NGS. The 5' and 3' extremities of each segment were sequenced using a primer including the 8-nt conserved sequence as previously described [25].
Complete genome sequencing was also performed for Corfou virus (CFUV) strain PaAr814 using the frozen cell culture supernatant medium following the methods above for comparison with the newly discovered TORV strains since they were shown to be closely related but only the complete S [26], the partial L [4] and M [27] genome sequences of the CFUV were known. Ultimately, all complete sequences obtained using NGS were verified by amplification and Sanger sequencing of overlapping regions spanning the entire genome.
Complete sequences of each of the 5 genes (L, Gn, Gc, N, Ns) were aligned without indels together with homologous sequences of selected phleboviruses retrieved from the Genbank database using CLUSTAL within the MEGA 5 program [28]. Nucleotide (nt) and amino acid (AA) distances were calculated with the p-distance method. Neighbor-joining (p-distance model) and Maximum likelihood analyses were carried out with AA sequences using MEGA version 5, with 1000 bootstrap pseudoreplications.
The Recombination Detection Program v.4.27 (RDP4) was used for recombination analysis using the nucleotide alignments. Recombination events, likely parental isolates of recombinants, and recombination break points were analyzed using RDP, GENECONV, Chimaera, MaxChi, BOOTSCAN, and SISCAN algorithms implemented in the RDP4 program with default settings [29].
To attempt identification of the sandfly species present in the TORV and ZERV positive pools, PCR was performed using 3-μL of nucleic acid extract of the pool to amplify the cytochrome c oxidase I (COI) gene using the following primers; LCO1490: GGTCAACAAATCATAAAGATATTGG and HCO2198: TAAACTTCAGGGTGACCAAAAAATCA [30]. The PCR products were processed and sequenced through NGS as described above. NGS reads were compared with available sequences in Genbank by Blastn using the CLC Genomic Workbench 6.5. For the final determination of the species the sequences were aligned with the reference sequences of regional populations of the sandfly species.
However, we would like to acknowledge that a valid protocol would be to cut off the male genitalia using a cold-stage microscope in the laboratory, so that the specimens can be identified morphologically. This would be faster and cheaper than PCR amplification followed by NGS or Sanger sequencing of the COI gene.
A total of 11,302 (4,513 females and 6,789 males) sandflies were collected in August and September 2012 and 2013 from eight villages (Fig 1) located in the surroundings of Adana city (Mediterranean Turkey). They were organized as 797 pools (494 females, 303 males) (Table 1). Two pools, #213 and #292, were positive with primers N-phlebo2S/2R and N-phlebo1S/1R [24], respectively. The 245-nt sequence obtained from pool #213 was most closely related with CFUV (Genbank no: GQ165521; 95% AA identity, 78% nt identity). The 513-nt sequence corresponding to pool #292 was also closely related with CFUV (Genbank no: GQ165521; 88% and 78% identity at the AA and nt level, respectively). These 2 pools consisted each of 20 male sandflies trapped in Damyeri village in 2012 (36S0733357 North and 4140570 East, altitude 194m).
The pool #37 (20 males trapped in Zerdali village in 2013; 36S732947 North and 4142749 East, altitude 238m) was positive using primers SFNV-S1/S2 [23]. The corresponding 390-nt sequence was most closely related to THEV (Genbank no: JF939848; 96% AA identity, 85% nt identity).
The TORV specific rt-RT-PCR confirmed that pools #213 and #292 were positive (Ct values < 26). Pool #10 (20 females collected in Damyeri in 2013) was also positive for TORV. Four-fold dilutions of the RNA were positive until the dilution 1:256 for the pools #213, #292, and #10.
The ZERV specific rt-RT-PCR confirmed pool #37 was positive (Ct value = 26.33), and detected ZERV RNA in 3 additional pools (#128–20 females, #342–20 males, and #374–29 females). Four-fold dilutions of the pool #37 RNA on the one hand and of pools #128, #342 and #374 on the other were positive until dilutions 1:4,096 and 1:1,1024, respectively. Pools#128, #342 and #374 consisted of sandflies trapped in 2012 in the respective villages of Damyeri, Camili and Koyunevi (Fig 1).
The rate of infection for TORV was 0.026%, for ZERV 0.035%, and for both TORV and ZERV 0.062% assuming that only one sandfly was infected in each pool.
Vero cells inoculated with pool #292 showed a clear CPE at day 6 post-inoculation (pi). Pool #213-inoculated Vero cells did not produce CPE during 4 serial passages. However, virus replication was demonstrated by RT-PCR (N-phlebo1 system, 24) starting from passage 3. CPE appeared at day 4 pi at passages 4 and 5 and virus replication was confirmed by RT-PCR. In a similar manner, pool #37- inoculated Vero cells provided a clear CPE at day 4 pi of passage 3 (RT-PCR was positive at passage 2). Neither virus isolation nor positive RT-PCR was obtained after 5 serial passages for pools #10, #128, #342, and #374.
Freeze-dried suspensions of ZERV-strain #37 (passage 8), TORV-strain #292 (passage 5) and TORV-strain #213 (passage 8) have been included in the collection of the European Virus Archive (www.european-virus-archive.com/) where they are publicly available for academic research at non-profit costs.
The complete genomes of both strains (#213 and #292) of the TORV consisted of 6,456 nts, 4,326 nts and 1,702 nts for the L, M and S segment, respectively (Genbank acc. no of the strain 213; KP966619, KP966620, andKP966621; Genbank acc. no of the strain 292;KP966622, KP966623, and KP966624). The polymerase gene encoded a 6,270-nt long ORF (2,090 AA), whereas the glycoprotein gene encoded a 4,077-nt long ORF (1,359AA). The small segment encoded a 738-nt and a 780-nt long ORF which when translated corresponded to the nucleocapsid protein (246 AA) and a non-structural protein (260 AA), respectively.
The complete genome of the ZERV (strain #37) consisted of 6,403 nts, 4,202 nts and 1,907nts for the L, M and S segment, respectively (Genbank acc. KP966616, KP966617, and KP966618). The polymerase gene encoded a 6,285-nt long ORF (2,095 AA), whereas the glycoprotein gene encoded a 4,002-nt long ORFs (1,334). The small segment encoded a 942-nt and a 762-nt long ORF which were translated to a nucleocapsid protein (314AA) and a non-structural protein (254AA), respectively.
The complete genome of the CFUV consisted of 6,453nts, 4,329nts, and 1,704nts for the L, M and S segment, respectively (Genbank acc. no KR106177, KR106178, and KR106179). The polymerase gene encoded a 6270-nt long ORF (2,090AA), whereas the glycoprotein gene encoded a 4,077-nt long ORFs (1,359 AA). The small segment encoded a 738-nt and a 780-nt long ORF which were translated to a nucleocapsid protein (246AA) and a non-structural protein (260AA), respectively.
Pairwise distances of the nt- and AA- sequences are presented in S1 Table. The alignment of each gene is also available in S2 Table.
AA distances between TORV and SFSV-like viruses (SFSV, SFTV, SFCV, CFUV) were ≤25.2% (N), ≤37.8% (NS), ≤43.3% (M), ≤40.7% (Gn), ≤33.7% (Gc) and ≤20.6% (L), whereas AA distances between TORV and other phleboviruses were much higher: ≥ 42.9% (N), ≥71.4% (NS), ≥58.7% (M), ≥53.3% (Gn), ≥ 47.4% (Gc) and ≥43.9% (L).
AA pairwise distances between ZERV and viruses of the SFNV species (TOSV, THEV, SFNV, PUNV, MASV, GRAV) were ≤17.3% (N), ≤58.0% (NS), ≤42.7% (M), ≤41.8% (Gn), ≤29.5% (Gc) and ≤17.4% (L), whereas AA distances between ZERV and other phleboviruses were much higher: ≥ 40.6% (N), ≥80.9% (NS), ≥66.3% (M), ≥64.8% (Gn), ≥55.0% (Gc) and ≥44.5% (L).
Gene by gene comparative analysis showed that distances observed between ZERV and viruses belonging to the SFNV species were consistently lower than the lowest distances observed between ZERV and non SFNV-phleboviruses. The same relationship was observed with distances between TORV and SFSV-like viruses on the one hand, and TORV and non-SFSV-like viruses on the other.
These findings are supportive for (i) the inclusion of TORV in the SFSV species complex (SFSV, SFTV, SFCV, CFUV), (ii) the inclusion of ZERV in the SFNV species complex (TOSV, THEV, SFNV, PUNV, MASV, GRAV).
Regardless of the gene used for phylogenetic analysis, and the tree-building programme (i.e. NJ or ML) TORV clustered with SFSV, CFUV and the other SFS-like viruses (from Turkey, Cyprus, Ethiopia) with bootstrap values ≥ 99% (Figs 2, 3, 4, 5 and 6). TORV consistently grouped together with CFUV (≥ 99%bootstrap) forming a subgroup within the SFS-like viruses that is distinct from the second subgroup including SFSV and related genotypes originating from Italy, Turkey, Cyprus, and Ethiopia. The stability of the topology and relationships between TORV and most closely related viruses suggested that the TORV genome did not contain evidence of genetic recombination or reassortment. Likewise, no recombination events were detected using any of the 6 algorithms implemented in RDP4.
ZERV consistently grouped together with THEV and Naples virus strain Yu_8–76 (subgroup I), with bootstrap values at ≥ 99 for L, Gn and Gc, and with lower values for N and Ns. Within this species, there were 3 other subgroups: (i) subgroup II: Toscana viruses; (ii) subgroup III: Naples viruses (except for Yu_8–76 included in subgroup I); (iii): subgroup IV: Granada, Massilia and Punique viruses.
Partial sequences were obtained in the polymerase gene for pool #128 (505 nt) and #10 (379 nt), and in the nucleoprotein for pools #128 (280 nt), #342 (438 nt), and #374 (245 nt). Partial polymerase sequences of the pools of #128 and #10 were identical (except 2 nts and 1 nt, respectively but 100% identical in AA) with ZERV and TORV, respectively. The partial nucleoprotein sequence from the pool #128 was also identical with the ZERV (except 1 nt but 100% identical in AA). However, the partial nucleoprotein sequences from the pool #342 (6 nt and 3 AA different) and #374 (49 nt and 2 AA different) were not identical with ZERV although originating from neighbouring localities (Fig 1); this suggests that there may be topotypes that remain to be identified and characterised.
Genotyping was performed for 7 pools. The species composition of the pools and number of reads are shown in Table 2. NGS reads were compared with available sequences in Genbank (Genbank accession numbers: KT634318, KF483675, KR349298, JQ769142, KF137560, KJ481126) by Blastn using the CLC Genomic Workbench 6.5. The species were determined when the consensus sequences had ≥98% similarity with the regional reference sequences except Sergentomyia sp. sequences which had ≥85% similarity with S. dentata from Adana (Genbank accession numbers: KU659595, KU659596, KU659597, KU659598; release date 01 July 2016). Therefore we indicated these sequences as Sergentomyia sp.
Although there are published serological data [13] and a recent report of detection of TOSV [19] in the Adana, Mediterranean region of Turkey there have been no previous reports of virus isolation. To further understand the presence of sandfly-borne phleboviruses that circulate in this endemic region for leishmaniasis [31], close to the border of Syria where many refugee camps are settled, we organised field-study campaigns in 2012 and 2013. The TORV and ZERV that were isolated during our study were most closely related to but distinct from SFS- and SFNV- like viruses, respectively. Studies conducted in 2012 led to the isolation of ADAV a novel putative member of the Salehabad species [20]. These results demonstrate that 3 phleboviruses belonging to 3 different genetic lineages co-circulate in the population of sandflies in this geographic area. Cumulative data resulting from this study and that of [20] enabled estimation of infection rates in sandflies (0.07% for sandfly-borne viruses in this region of Turkey) which is in the same order of magnitude as previously reported in France, Tunisia, Spain and Italy [2, 23, 32,33, 34].
Regardless of the gene used for analysis TORV and CFUV were grouped together in a sublineage that is clearly distinct from that including all other SFS-like viruses. CFUV was isolated from Phlebotomus major sensu lato [35] trapped in the eponymous Greek island. Interestingly, TORV has been isolated from two pools that contained P. perfiliewi sensu lato and P. tobbi, both belonging to the Larroussius subgenus as P. major sensu lato. Similarly, SFTV was detected in P. major sensu lato [21]. In contrast, other SFSV strains were isolated from P. papatasi that belongs to the Phebotomus subgenus which is distinct from the Larroussius subgenus [36]. Therefore, the two subgroups of viruses might reflect vector properties, with CFUV/ TORV associated with the Larroussius subgenus whereas other SFSV viruses are associated with P. papatasi with the exception of SFTV association with P. major sensu lato. Vector-virus association needs to be studied in a more detailed manner in order (i) to determine unambiguously the sandfly species transmitting these newly described viruses, (ii) and to verify our hypothesis that virus subgroups within viral species might be linked to specific vectors belonging to distinct taxonomic entities.
Additional field studies combined with experimental studies using sandfly colonies need to be initiated to understand the parameters driving vector capacity and competence for different strains of viruses.
ZERV was consistently grouped together with THEV and Naples virus strain YU-8-76. Interestingly, THEV and the Serbian isolate Yu 8/76 apparently do not require expression of the NSs ORF, since their replication is not impaired by the presence of either an early stop codon or a large truncation [37], whereas there is no such impairment or truncation in the ZERV genome which has a complete NSs ORF. Similar observations were also reported for a naturally attenuated RVFV strain (clone 13) that has a large in-frame deletion in the NSs coding region [37, 38].
Within the SFNV species, it is possible to discriminate 4 sublineages (I to IV); we propose to assign ZERV to sublineage I, where it was most closely related with THEV (Figs 2, 3, 4, 5 and 6). THEV was isolated from P. papatasi sandflies in Iran in 1959 [39], whereas YU-8-76 strain of SFNV was isolated from P. perfiliewi sensu lato trapped in Serbia in 1976 [37]. Subgroup III appears to be associated with P. papatasi, whereas subgroups II and IV appear to be associated with vectors belonging to the subgenus Larroussius. Subgroup I may be associated with Larroussius except THEV isolated in P. papatasi. Only P. tobbi was found to be present in all of the four ZERV positive pools. The same comment formulated above concerning the need for experimental studies to understand species-related competence and specificity of sandflies applies here.
Genetic and phylogenetic analyses support the fact that both ZERV and TORV should be considered as new strains within pre-existing SFNV species and the yet to be recognised species including SFSV and CFUV, respectively.
To date, SFSV and CFUV are listed as tentative species by the ICTV [1]. This study, based on complete genome sequences, suggests that all these viruses should be considered as members of the same species which could be further subdivided into CFUV / TORV, and SFSV / SFS-like viruses. A written proposal will be submitted to the Bunyaviridae Study Group of the ICTV.
During this two-year study, 48% of the sandflies were trapped from Damyeri village where the ecological conditions were no different from those observed in other sampling stations. However, in Damyeri, the number of domestic animals (sheep, goats, and cows) was much higher than in other localities and these animals were constantly in the close vicinity of houses producing droppings which are known to be favoured breeding sites for sandflies or we set the traps very close to the possible breeding sites by chance therefore we got higher numbers of sand flies for this village. Thus, human exposure to sandflies might be greater in Damyeri than other sampling stations.
Whether TORV and ZERV can infect humans and cause diseases such as sandfly fever is currently unknown and remains to be investigated. After the outbreak of sandfly fever occurred in Adana in 2008, specific IgM against SFSV and/or SFCV was detected in acute cases by mosaic-immunofluorescence test, although the cause of the epidemic was not formally established through virus isolation or molecular detection with sequence confirmation [13]. The region where this outbreak occurred is located less than 25 km away from our trapping stations. Despite the fact that this study does not provide results supporting that both newly discovered viruses are human or animal pathogens, both TORV and ZERV belong to genetic groups that include several human pathogens. There is no doubt that SFSV "historic strains" were causing massive outbreak of debilitation and incapacitating disease [12, 40, 41]. Moreover, SFCV was also isolated from human cases [42, 43], as well as SFTV [13]. Antibodies against CFUV / SFSV were reported in humans living in mainland Greece and on Corfu Island using the immunofluorescence assay (IFA) [44]. Viral RNA of Chios virus, closely related to CFUV, was detected in the CSF of a patient presenting with severe encephalitis (Papa and Pavlidou, 2003, Genbank no, AY293623). In contrast, specific antibodies were never described in humans for THEV [12], although SFNV a close relative of THEV and subsequently ZERV was undoubtedly the cause of explosive outbreaks in newcomers to endemic areas during summertime [9]. Importantly, although serological studies have not yet been reported we do have serological data to support the concept that the newly isolated ZERV and TORV can infect vertebrates. However, whether or not these vertebrates do play a reservoir or amplifying role is not yet clear. Interestingly, although TORV and ZERV genomic RNA was detected in female pools of sandflies, both viruses were only isolated from male pools. Based on current knowledge it is not known how male sandflies become infected. Whilst, transovarial transmission seems a likely possibility it is not yet known how significant or efficient this mechanism of transmission is in natural habitats. However, laboratory experiments have shown that the rates of infection amongst offspring are low and show a decline from the first generation to ongoing generations. Other studies suggest that venereal (horizontal) transmission from infected males to uninfected females by mating and transstadial transmission of TOSV in diapausing Phlebotomus perniciosus larvae [9] may also contribute to long-term virus survival. From what is currently known and in the absence of defined vertebrate reservoirs, maintenance and transmission of sandfly-borne phleboviruses appears to depend on the abundance and accessibility of appropriate vector species. This lack of available knowledge of virus transmission and the virus maintenance mechanisms clearly need to be investigated both in natural habitats and under experimental conditions. Future studies are planned to examine female sandfly salivary glands and heads to look for the presence of infectious virus.
Recent studies have detected sequences compatible with the existence of many putative new phleboviruses transmitted by sandflies in the Old World. In most of the cases, they were only partial sequences in the polymerase or the nucleoprotein genes. To date these limited genetic data, preclude classification by ICTV. This situation applies for viruses that may belong to (i) the SFNV species such as FERV [2], Provincia virus [48], Girne1 virus [18], and Saddaguia virus [45], (ii) the SALV species such as Adria virus [46, 47] Olbia virus [48] and Edirne virus [18], and to (iii) the SFSV / CFUV complex such as Chios virus (Papa and Pavlidou, 2003, Genbank no, AY293623), Utique virus [4], Girne2 virus [18], SFS-like viruses [4, 42, 43, 49]. Accordingly, although our knowledge of sandfly-borne phleboviruses is more extensive than it was a half-decade ago; efforts to isolate virus strains and determine their complete sequence should continue. Since virus taxonomy for the Phlebovirus genus still relies on neutralisation-based antigenic relationships, virus isolation is also essential. Nevertheless, the criteria for taxonomy appear to be evolving towards full-length genome comparative analysis.
In conclusion, the results obtained in this study together with previously published results [20] demonstrate that (i) at least 3 different phleboviruses are co-circulating in phlebotomine sandfly populationsin the Adana region of Mediterranean Turkey; (ii) these new viruses belong to 3 distinct but closely related phylogenetic groups or species (SFNV, SALV, SFSV / CFUV); (iii) all the closely related viruses are known to be sandfly-borne arboviruses; (v) we have evidence of vertebrate seroprevalence for this group of viruses. Thus whilst it is indirect, the evidence for TORV and ZERV being arboviruses is compelling. It is also important to emphasise that although >10,000 sandflies were tested in this study; TOSV was neither detected nor isolated from the field samples. This challenges recent reports of TOSV-specific antibodies in blood donors from this region of Turkey [17], and TOSV RNA / TOSV IgG in dogs from Mersin and Adana [19]. However, we recognise that the sampling points in these earlier studies do not overlap with those selected for our investigations.
|
10.1371/journal.pgen.1007135 | Histological subtypes of mouse mammary tumors reveal conserved relationships to human cancers | Human breast cancer has been characterized by extensive transcriptional heterogeneity, with dominant patterns reflected in the intrinsic subtypes. Mouse models of breast cancer also have heterogeneous transcriptomes and we noted that specific histological subtypes were associated with particular subsets. We hypothesized that unique sets of genes define each tumor histological type across mouse models of breast cancer. Using mouse models that contained both gene expression data and expert pathologist classification of tumor histology on a sample by sample basis, we predicted and validated gene expression signatures for Papillary, EMT, Microacinar and other histological subtypes. These signatures predict known histological events across murine breast cancer models and identify counterparts of mouse mammary tumor types in subtypes of human breast cancer. Importantly, the EMT, Adenomyoepithelial, and Solid signatures were predictive of clinical events in human breast cancer. In addition, a pan-cancer comparison revealed that the histological signatures were active in a variety of human cancers such as lung, oral, and esophageal squamous tumors. Finally, the differentiation status and transcriptional activity implicit within these signatures was identified. These data reveal that within tumor histology groups are unique gene expression profiles of differentiation and pathway activity that stretch well beyond the transgenic initiating events and that have clear applicability to human cancers. As a result, our work provides a predictive resource and insights into possible mechanisms that govern tumor heterogeneity.
| We developed predictive gene signatures that identify specific histological mouse mammary tumor subtypes with high fidelity to expert pathologist classifications. As a result, these signatures are a powerful tool for classification, particularly in cases of intratumor heterogeneity; where confounding results arise from differences in the tumor portions sent for pathology and separately for molecular analysis. Further, we show that despite differences in the tumor initiating oncogene, histological subtypes in mouse mammary tumor are unified in their transcriptomic profiles and activation of cell signaling pathways. We find that these transcriptomic profiles and activation of key signaling pathways are not only conserved in human breast cancer, but also other human cancer types. Further, the EMT, Adenomyoepithelial and Solid signatures were prognostic in specific human breast cancer subtypes. Indeed, this work provides a new and needed perspective on how mouse models relate to specific human breast cancer subtypes by showing that the tumor histology of the mouse mammary tumor is far more important than the initiating oncogenic event in terms of how the mouse mirrors a specific human subtype.
| One of the hallmarks of breast cancer is tumor heterogeneity at both the histological and genomic level. The histological type of the tumor refers to the morphological and cytological patterns evident within a tumor. There are a large number of special tumor histologies recognized for breast cancer [1, 2] including lobular, cribriform and several other types. The most frequently observed tumor histology is the invasive ductal carcinoma [3]. Similarly, there is a large degree of genomic heterogeneity in human breast cancer, which has been classified using gene expression analysis. Classification of breast tumors into their molecular subtypes based on unique gene expression profiles has led to tumors being described according to their “intrinsic subtype”: Basal-like, Luminal A, Luminal B, Her-2 enriched, Claudin Low and Normal-like breast group [4–6]. Importantly, these intrinsic subtypes of breast cancer provide a basis by which researchers can classify tumor heterogeneity.
Importantly, recent work has identified the gene expression relationships between intrinsic subtypes of human breast cancer and specific histological types of breast cancer [2]. Chief amongst their findings was that within intrinsic subtypes of cancer were multiple histological types of cancer. For example, both medullary and metaplastic breast cancer were categorized as claudin low. Further, individual tumors of the same tumor histological types corresponded to different intrinsic subtypes of breast cancer. For example, some medullary tumors were classified as basal and others were categorized as claudin low. These findings suggest that gene expression methods may do better job of organizing tumors into similar disease entities[2]. Collectively, these studies demonstrate that histological and genomic heterogeneity present in breast cancer establishes a complex array of distinct subtypes of tumors[2, 4].
With this complexity, modeling breast cancer in vivo requires numerous preclinical models that effectively mimic the multiple factors inherent to human breast cancer progression and parallel the molecular profiles of human breast cancer subtypes. While the use of human cell lines and patient derived xenografts offer the opportunity to study human breast cancer in vivo, they rely on immunocompromised hosts. The use of genetically engineered mouse models of cancer offer the advantage and the opportunity to study tumor progression in an immuno-competent system. As a result, a major focus has been to establish which genetically engineered mouse models have parallels in human breast cancer. [7]. Expanding upon these findings with additional tumor models and samples, numerous reports have documented mouse and human counterparts at the level of gene expression [8–12]. In addition, copy number variation at the chromosome [13]and gene level [14]has been predicted from expression data and examined similarity to human breast cancer. The gene level CNV predictions demonstrated that chromosomal alterations were associated with histological subtypes[14]. With gene expression similarities to human breast cancer, a critical need remains to address how the tumor histology of mouse mammary tumors is related to gene expression programs.
As seen in human breast cancer, a large number of histological subtypes have been observed for mouse mammary tumors [15]. This includes glandular, acinar, cribriform, papillary, solid, squamous, fibroadenoma, adenomyoepithelioma, adenosquamous, microacinar, adenocarcinoma, comedoadenocarcinoma, and medullary [8, 15–17]. Prior characterization of mouse models illustrates a number of mouse models with varied histological subtypes present across the population of tumors. For example, amongst Myc initiated tumors, epithelial to mesenchymal (EMT)-like, papillary, microacinar, solid, and squamous tumors were observed [18]. Comparison of mouse and human histological subtypes reveals key differences, for example squamous tumors are not frequently observed in human breast cancer [1, 3]. As such, it is critical to begin to understand how mouse and human tumor pathologies impact the genomic relationships between mouse models and human breast cancer.
To address the need to characterize the genomic patterns defining histological subtypes to allow a mouse / human comparison we have examined a wide spectrum of mouse model tumors. In previous work we observed that unsupervised hierarchical clustering of Myc initiated tumors resulted in subclasses that correlated with their histology [19]. Further, even in the presence of loss of the activator E2F transcription factors, clustering arranged tumors according to histology, rather than genotype[20]. This suggested that there are unique gene expression components inherent to histological subtypes apart from the initiating oncogenic events. Using gene expression data from histologically annotated mouse mammary tumors initiated by different oncogenic events, we have developed gene expression signatures that define tumors with squamous or adenosquamous, EMT-like, microacinar, solid, papillary, or adenomyoepithelial tumor histology. Applying these signatures to our published database [9] of mouse mammary tumors we scored mouse tumors for histology, tested which cell signaling pathways tightly correlate with tumor histology, and investigated signature relationships to human breast cancer. Together, this data demonstrates robust signatures that can be used to predict tumor histology and further our understanding of human breast cancer heterogeneity.
To build a gene expression signature that could identify specific histological types of tumors, we utilized publicly available gene expression data that was annotated for mammary tumor histology for each sample analyzed on array. For each tumor type that we built signatures for, histology is described in Table 1 according to descriptions from expert pathologists [15, 21]. Using significance analysis of microarrays(SAM), we identified genes uniquely and consistently differentially expressed in a specific tumor histology in a training dataset. For example, we utilized histological classifications of tumors from our previous study of the MMTV-PyMT mouse model where squamous, microacinar, and papillary tumors arise [16](Fig 1A). Using SAM, we filtered out genotype differences to identify genes consistently differentially regulated and intrinsic to the squamous identity (Fig 1B). Focusing only on the genes detected in all four comparisons, we identified 184 genes upregulated in squamous tumors. We did not detect any genes that were consistently downregulated in this comparison. We tested the performance of these genes on the training data using unsupervised hierarchical clustering. As expected, this separated adenosquamous tumors from papillary and microacinar tumors regardless of E2F status (Fig 1C).
To validate these genes, we then tested performance on a separate dataset of histologically annotated tumors (MMTV-Myc tumors). Unsupervised hierarchical clustering separated Myc-induced squamous tumors from non-squamous tumors (Fig 1D) and importantly gene set enrichment analysis (GSEA) showed that Myc induced squamous tumors were significantly enriched for upregulation of the squamous signature genes derived from the MMTV-PyMT tumor dataset (Fig 1E, Normalized Enrichment Score or NES = 1.48, nominal p-value = 0.0, FDR q-value = 0.029, fwer p-value = 0.016). This illustrated the squamous signature genes as robust and valid with the ability to properly classify squamous tumors in another gene expression dataset and in tumors initiated by a different oncogene.
Using a very similar approach, we generated gene expression signatures for EMT-like tumors (S1 Fig), microacinar tumors (S2 Fig), papillary tumors (S3 Fig), solid tumors(S4 Fig), and tumors with adenomyoepithelial(S5 Fig) content. In each case, potential signature genes were identified using SAM (q-value ≤ 5%) doing multiple comparisons between the target tumor histology and other tumor types in the dataset. Unsupervised hierarchical clustering and GSEA was used on a separate histologically annotated dataset to validate the signature.
As additional validation of our signatures, we examined individual genes for prior association with histological types in the literature. As shown in Table 2, several of the squamous signature genes have been shown to be markers for squamous tumors and keratinocytes. Similarly, many of the traditional markers (such as Zeb1, vimentin, E-Cadherin) of EMT were captured in our signatures. In addition, genes from the papillary and adenomyoepithelial signatures also had been observed as markers of these histologies. Together, the ability to detect known histological subtypes across datasets and mouse models as well as the historical use of several individual genes depicts these signatures as robust classifiers of mouse mammary tumor histology. Importantly, each of the histological signatures is provided as a supplemental file (S1 File) in GSEA “.gmt” format as a predictive resource.
To test our gene expression signatures in mouse mammary tumors, we utilized two published mouse mammary tumor model databases [9, 22]. To identify the most likely histology of each tumor in the dataset, we utilized single sample GSEA (ssGSEA) and ordered tumors according to their highest scoring signature. With this approach, we observed tumors with robust expression of signature genes for each histology (Fig 2 and S6 Fig). In addition, there were also tumors that did not show strong expression patterns for a particular histology signature, likely indicating a different histology without a predictor.
With application of these signatures, we see evidence for profound histological heterogeneity both across and within mouse models. For example, Myc, PyMT, Wnt, Large T, and p53 lines had tumors with a squamous prediction. Indeed, no histological prediction was represented by a single mouse model and most mouse models (as categorized by the driver event) showed histological heterogeneity. For example, Wnt, Met, and Myc induced tumor models presented tumors with high scores for each of the other histological subtypes, consistent with reports of histological heterogeneity in these models [18, 23, 24].
Alternatively, other models had a preponderance of a particular histological outcome. This is best represented by the Wap-Int3 and Notch induced tumors which were predominantly enriched for the papillary signature. Another model, H-Ras initiated tumors favored microacinar and solid nodular outcomes. Interestingly, models featuring inducible expression of an oncogene, showed elevation of the EMT signature in the recurrent tumors(S7 Fig); consistent with prior reports [25, 26]. Finally, predictions organized into figures for each individual mouse model are provided as additional material (S2–S28 Files).
As an additional test of the validity and capability of our signatures, we itemized tumors that had been individually annotated for a particular histology by a pathologist (see blue bars above heatmap, Fig 2). Overall, the pathologist based classification of individual tumors and the classification predicted by the expression signatures demonstrated a high degree of agreement. In addition to this, we cross-referenced the literature to determine whether any of the predicted histologies for a given mouse mammary tumor had been observed in reports for that model. As shown in Table 3, many of the predicted histological match reports for tumors from individual mouse models. Finally, MMTV-Myc tumors with mixed histology (multiple histological components within a single tumor) were noted to have strong scores for individual histology signatures. Thus, we examined matched H&E sections and find that in 89% of samples, the predicted histology was present in at least half the section and 100% concordance where the predicted histology was present in some part of the sample (S8 Fig). Thus, these signatures demonstrate the ability to resolve intra-tumor heterogeneity by identifying the dominant histological component of the tumor being transcriptomically profiled. Importantly, all scores for tumors in each dataset are provided for download (S29 File).
With our large dataset and robust performance of the histology signatures, we aimed to test for relationships between tumor histology and other features of mammary gland differentiation. To enable these comparisons, we used the histological classifications made by ssGSEA for each tumor and used standard GSEA to test for enrichment of other signatures in a comparison of predicted tumor histological subtypes (S30 File).
We noted prominent associations between histological classes of tumors and signatures for mammary cell types. As shown in Fig 3A, squamous, EMT, and tumors with high adenomyoepithelial content showed high expression signatures for mammary stem cells and mammary basal cells. Amongst these, EMT tumors displayed features most concordant with mammary stem cells. Squamous tumors showed the highest expression of the mammary basal signatures and had gene expression features (S9A and S9B Fig) that suggests these tumors are further along the differentiation spectrum than EMT tumors but not as differentiated as other histology types (S9C and S9D Fig). Papillary tumors were more luminal progenitor-like, showing moderate expression of both mammary stem cell and luminal progenitor cell signatures. Finally, the microacinar and solid tumors showed gene expression patterns consistent with those found in mature luminal cells(Fig 3A).
We also evaluated the relationships of signatures of breast cancer subtype (Fig 3B)[5]. Squamous tumors had highest expression of signatures for basal subtypes of breast cancer. As expected, EMT tumors showed high expression of a signature for claudin low subtypes and showed less luminal or basal-like features. Papillary and tumors with high predicted adenomyoepithelial content showed more moderate expression of all signatures for subtype; while microacinar and solid tumor types had high expression of signatures for luminal breast cancer. As a whole, this suggests a range of differentiation states across histological types.
We next tested for relationships between histologies and specific features with tumor progression (Fig 3C). Consistent with prior studies[27], EMT tumors showed high expression of the hallmark angiogenesis signature. In addition, microacinar and solid tumors exhibit low expression of this signature. In addition to angiogenesis, it was interesting to note differential expression of breast cancer metastasis signatures in these mouse mammary tumor types. The ‘Vantveer Breast Cancer Metastasis Up’ signature was high in microacinar tumors and low in EMT tumors, while EMT tumor showed expression of other metastasis signatures. In addition, squamous tumors showed lower expression of metastasis signatures. Together, this suggests differences in metastatic capacity and mechanism for individual tumor histologies.
In addition to phenotypic features, we also tested for key molecular aspects of each tumor histology. In many cases, the histology signatures themselves provide insight into key molecular features, as key signaling molecules were present in several of the signatures. Fig 4A, shows elevation of several pathways consistent with the relationships already detected. For example, Hedgehog and Wnt signaling in squamous tumors[28–31]. In addition, several pathways are shared between histology types. For example, EMT and squamous tumors share high expression of Kras signatures. Microacinar and solid tumors share Erbb2 signature expression, AKT1 signaling via MTOR signature expression, and very low expression of Vhl targets. Examining transcription factors (Fig 4B), a number of key relationships are predicted. Some are known markers, such as TP63 in squamous and Zeb1, Yap, and Ets transcription factors in EMT tumors are noted. However, unexpected relationships were also present such as Esr1 in microacinar tumors. Despite similarities luminal features, it is interesting to note that the E2F1 signatures distinguishes solid tumors and microacinar tumors. Signature genesets were also tested for overrepresentation in curated pathway databases, offering predictions of additional pathways of interest for each type of tumor histology (S10 Fig and S11 Fig)
Examining potential miRNAs with GSEA (Fig 4C), suggests tumor types where miRNAs may be actively expressed or lost. For example, mir-202, mir-17-3p, mir-517 targets are highly expressed in EMT tumors and lowly expressed in the more luminal tumors. Mir-486 was also interesting as its targets showed low expression almost exclusive to microacinar and EMT tumors. Similarly, mir-133A showed evidence for repression in papillary tumors. Taken together, these data suggest a number of key molecular features from pathways, transcription factors, miRNAs for each tumor histology.
Given high expression of human breast cancer signatures in certain histologies (ie- luminal signatures in microacinar),we tested whether any of the mouse tumor histology signatures were enriched in subtypes of human breast cancer using the Metabric dataset[32]. As shown in Fig 5A, a portion of the squamous signature was highly expressed in basal tumors. This suggests that mouse mammary squamous tumors are basal-like, but human basal tumors are not known to be squamous. However, human basal tumors and mouse squamous tumors shared similarly high expression of well-studied pathway ligands within the squamous signature (shared high expression of Wnt10a, Wnt6, Bmp2, Bmp7, and Jag2). Moreover, testing the 45 common highly expressed genes for overrepresentation in pathway signatures indicates possible shared activation of Hedgehog, Wnt, and Bmp pathways in mouse squamous tumors and human basal breast cancer (S12 Fig). Similarly, a subset of the genes that are highly expressed in microacinar tumors were highly expressed in luminal subtypes. Amongst these microacinar genes, many have previously been associated with luminal breast cancer and are also amongst genes that define mature luminal cells (S13 Fig). Finally, both genesets (up and down) that define EMT tumors were significantly expressed in claudin low tumors S14A and S14B Fig). This result is consistent with numerous reports that mouse EMT tumors are molecularly similar to claudin low tumors[7, 8, 12, 26, 33]. Together, these data further define appropriate mouse counterparts for study of human breast cancer.
With high expression of signature genes in certain subtypes of human breast cancer, it was important to test whether these signatures displayed predictive capacity of clinical events across human breast cancer patients. As shown by Kaplan-Meier analysis, high expression of the EMT and adenomyoepithelial signatures are associated with acceleration of tumor relapse in basal-like breast cancer (Fig 5B and 5C respectively). Adenomyoepithelial signatures were also associated with relapse and earlier onset of distant metastasis in Lum B breast cancer (Fig 5D, S14C Fig respectively), while having high expression of the solid signature was protective in luminal B (Fig 5E). Finally, high expression of the papillary signature genes were associated with accelerated progression to distant metastasis in Her-2 enriched breast cancer (S14D Fig). Together, these results suggest potential mouse tumor types for investigating these human counterparts and prognostic features.
Since some of the histology types observed in mouse mammary tumors are often found in other human cancers (ie- squamous lung tumors, papillary thyroid tumors), we sought to test whether the mouse signatures were enriched in other human cancer types. We utilized public gene expression data from the gene expression omnibus and mediated batch effects according to established protocol[9]. Using unsupervised hierarchical clustering arranged many of the tumors with squamous histology across lung, oral, melanoma, and esophageal cancer types into the same cluster with high expression of our murine squamous signature (Fig 6A, green cluster) and GSEA testing showed significant enrichment in these tumors (Fig 6B). While the mouse mammary tumor squamous signature extended to other human cancers, the murine papillary signature was not highly expressed in human papillary tumors. The other murine signature with enrichment in human cancers were the EMT signatures that showed concordant expression in a subset of melanoma and metastatic melanoma tumors (Fig 6A, blue cluster). As expected, GSEA showed significant enrichment in these tumors (Fig 6C, S15 Fig).
Given that we have itemized many similarities between gene expression profiles of mouse human tumors across cancer types, we tested for unifying features at the level of transcriptomic indicators of pathway activity and differentiation. The concise summary is shown in Fig 7 and more detailed results are available in S16–S18 Figs. Collectively, we observed that murine mammary tumors from the EMT histopathology is similar to human tumors from claudin low breast and melanoma. This includes having gene expression features similar to those found in stem cells and having Kras pathway activity. Mouse and human squamous tumors share enrichment of basal-cell genes and HRas pathway activity, and while similar pathways are active in human basal breast tumors, basal-like breast tumors were enriched for upregulation of luminal progenitor cell genes. We were unable to find human counterparts for the murine papillary tumors in our analyses. For mouse mammary microacinar and solid tumors, luminal features were observed, and like human luminal tumors, enrichment for luminal cell signatures were detected; complete with high expression of ER-target genes. Taken as a whole, these observations suggest that many features of murine tumor histologies are conserved from mouse to human and across several different cancer types.
In this study, we generated and validated signatures for specific histologies that are observed in mouse mammary tumor models. Both training and validation sets utilized prior histological annotations from expert pathologists from a number of studies. With our signature generation and validation approach (Fig 1, S1–S5 Figs), we show that features of tumor histology span oncogenic mouse models of cancer and human cancers (Figs 5–7). As shown in Fig 2 and Table 3, these signatures were predictive of known historical observations for tumor models in our dataset. Thus, we believe these signatures to be a valuable resource tool and have provided our signatures in gene set enrichment analysis format. With a robust capacity to identify tumor histology types, we used this platform to investigate and predict molecular features of each tumor histology all the way from broad features such as differentiation, to specific molecular aspects such as pathway, transcription factor, and miRNA utilization.
Based on prior studies, relationships between the mouse EMT signature and claudin low tumors [12, 33–35] were expected. In addition, recent reports have highlighted cases of melanoma that classify as claudin-low[27, 36, 37]. Unlike the relationship between human basal breast cancer and mouse squamous tumors, this relationship is likely due to similar histologies; as breast and melanoma claudin low tumors, like EMT tumors, have been reported to contain spindle-shaped cells[2]. Importantly, evidence of stem-cell like properties and Kras activation was identified in each of these cancer types. Activating mutations in Kras have been observed in mouse EMT[18, 35, 38], however in human breast cancer, the prevalence is somewhat low, as COSMIC[39] reports 80 instances that lack intrinsic subtype information (with the exception of MDA-MB-231 cells that are claudin low). In the case of melanoma, it seems likely measures for Kras activity stem from downstream activating mutations in Braf, which are common to melanomas (COSMIC reports 44% of melanomas with Braf mutations).Together, these data suggest events affiliated with the Kras pathway are important to the EMT / claudin low outcome.
We also detected relationships between squamous tumors and human basal breast cancer that seemed to stem from shared activity of multiple pathways. These shared pathways, such as Hras and hedgehog signaling, seem to come from activation events outside of mutations of those genes as both COSMIC and C-Bio-Portal illustrate a low incidence for DNA events on these genes. Although, as reported by TCGA, 32% of basal-like breast cancers harbor amplifications of Kras[40]; suggesting the Hras signature maybe measuring Kras activity in these tumors. Regardless, the shared activation of key pathways supports the use of squamous tumors as a tool for investigating human basal breast cancer at the pathway level.
Mouse microacinar tumors showed gene expression traits that define luminal breast cancers. At the pathway level, the relationships between mouse microacinar and human luminal breast cancers is still somewhat perplexing. While both the mouse and human tumors show strong expression of mature luminal cell differentiation signatures and activation of several pathways, mouse microacinar tumors also show activation of Erbb2 signaling, which is traditionally associated with the Her-2 enriched subtype of breast cancer. Furthermore, the microacinar tumors showed high expression of signatures for estrogen receptor signaling. Yet, mouse mammary tumors are notoriously ER-negative by IHC. Indeed, this does draw comparisons to the human setting where Her-2 negative tumors still classify as Her-2 enriched in intrinsic profiling despite the IHC diagnosis as Her-2 negative[41]. This might indicate that similarities to luminal breast cancer are achieved by expression of estrogen receptor target genes by a mechanism other than estrogen receptor itself.
Interestingly, several of our mouse histology gene signatures were prognostic in specific intrinsic subtypes of human breast cancer. For example, luminal B tumors with high expression of the solid tumor signature displayed prolonged times to relapse. This finding is particularly of note in light of the recent finding that HER2+ tumors with luminal B gene expression profiles benefitted significantly from trastuzumab[42]; similarly, we note elevation of a Her-2 (Erbb2) signature in murine solid tumors that also have luminal expression profiles (Fig 4A). However, our solid signature was not predictive of prognosis in Her-2 enriched tumors, suggesting the criticality of other pathways differentially regulated between luminal and Her-2 enriched tumors. High levels of EMT and adenomyopithelial signatures were associated with accelerated relapse in basal-like breast cancers (likely identifying basal-like tumors with claudin-low like properties). Indeed, relapse following chemotherapy is common in these tumors [43] and other work has shown an association of EMT phenotypes with chemo-resistance, in part due to lower rates of proliferation and apoptosis[44, 45]. Taken together, these findings are of particular significance because they may specify high risk patients where alternative therapies may be necessary. In addition, these signatures may suggest appropriate mouse models for testing new therapeutic strategies.
The fact that the same histological fates are often achieved despite differences in oncogenic events, genetic background, and promoter ultimately questions the mechanism(s) for development of a particular tumor histology. Examination of mammary cell differentiation signatures across tumors revealed unique differentiation states within each tumor histology. Indeed, it is tempting to infer that this indicates the cell of origin leading to tumor initiation and that this cell of origin ultimately drives histological outcome. Indeed, work using the PyMT model suggests that cell of origin plays a role in histological outcome [46]. Yet, more recent work counters that while cell of origin still might be a factor, the initiating oncogenic mutation plays a large role in the histological outcome[47].
In light of these findings[47, 48] and our study, one might envision that the particular combination of pathways that are activated could commit cells into a specific differentiation state. Alternatively, this could also cause selective outgrowth of specific populations of cells. Ultimately either case would result in tumors forming a particular histology. In support of this, previous work using an inducible Myc mouse model showed that after Myc withdrawal, tumors regressed, and then recurred with tumors mainly being EMT or squamous with activating mutations in Kras [25]. In part, activation of Kras was thus associated with development of these characteristic tumor pathologies. In addition, we and others have observed Kras activation in both of these histological types in other models [35, 49, 50]. Indeed, our work presented here might provide predictions as to which differentiation state of cells and which pathways drive the formation of particular histologies.
While our method provides robust classification of tumors in our large dataset, there is one important application guideline we wish to highlight. Due to gene centering techniques that are often employed with normalization of gene expression data, predicting tumor histology should be done in settings with adequate tumor heterogeneity or done using methods that adjust for skewed pathological data. In cases where heterogeneity across the dataset is low, we recommend batch adjusting[51, 52] to combine datasets of interest with large datasets such as our own[9] or others [7, 10] prior to employing gene signatures. In addition, we wish to refer to the work of Zhao et al which describes in great detail the issues surrounding gene centering and classification of homogenous cohorts while providing alternative approaches for solving such issues[53]. Indeed, the technicalities of gene centering on skewed molecular datasets highlight the necessity of conventional classification methods such as IHC and pathology of H&E stained sections to enable proper data handling.
Finally, though traditional tumor classification methods are essential, the gene-signature based classification method here offers several key advantages. First, intra-tumor heterogeneity presents challenges for accurate interpretation of the data that cannot always be addressed by conventional methods. Illustrating this, we examined a large number of tumors presenting mixed tumor histology where the portion of the tumors analyzed on microarray displayed a gene expression profile representing a major histological class. Importantly, the histology predicted by our gene expression signatures were concordant with the major component present in the associated histological section. Therefore, these signatures represent an important tool for resolving mixed cases and ensuring molecular profiles match the expected histology from H&E. Another advantage over conventional methods is the reduced variance in the clinical classification of tumors and classifying cases where histology might be misleading. This is demonstrated by Her-2 enriched Her-2-IHC negative tumors in human breast cancer [41] and ER-target gene enrichment in ER-IHC-negative microacinar tumors from mouse mammary tumor models. Finally, we demonstrate the ability of gene signatures to tie tumor cell phenotypes and functions to supporting pathways that represent therapeutic targets beyond the capacity of IHC. It is our hope that this work’s correlation of gene expression signatures to specific cell biology in the form of tumor histopathologies may provide useful inroads to understanding tumor subtype, tumor progression, and for identifying specific therapeutic strategies aimed at the biological processes upon which the tumor cells depend.
Previously published gene expression data were derived from mouse and human tumors and done in accordance to the ethics statements as reported in their respective publications.
Details for assembling the mouse mammary tumor model databases can found [9, 22]. For the squamous signature, the training data was derived from squamous and non-squamous MMTV-PyMT tumors; this data is deposited on GEO Datasets GSE104397 [54]. All animal work has been conducted according to national and institutional guidelines. These tumors were prepared by isolation of RNA samples from flash frozen tumors using the Qiagen RNeasy kit after roto-stator homogenization. RNA was submitted to the Michigan State University Genomics Core facility for gene expression analysis using Mouse 430A 2.0 Affymetrix arrays. The validation set for the squamous signature was from MMTV-Myc tumors found under GSE30805 and GSE15904[8]. The training dataset for generation of the EMT signature is published, GSE30805 and GSE15904 [8]. The validation dataset for the EMT signature can be found GSE41601[12]. Generation of the microacinar signature was done by dividing the published dataset[8] into training and validation sets with random sample selection. The training dataset for generation of the papillary signature is published, GSE30805 and GSE15904 [8]. The validation sets for the papillary signature were from Array Express E-MEXP-3663 [55] and gene expression omnibus GSE20614[56]; batch effects between datasets were mediated using combat[52]. The solid tumor signature was generated using the training dataset GSE41601[12] and validated using GSE73073[57]. Finally the signature for adenomyoepithelial content was generated using from Array Express E-MEXP-3663 [55], filtered using GSE69290 [58], and validated on GSE37223[59].
Gene expression data for human squamous and non-squamous tumors was accessed on the Gene Expression Omnibus under the following accession numbers: GSE10245, GSE10300, GSE14020, GSE17025, GSE18520, GSE2034, GSE20347, GSE21422, GSE21653, GSE2280, GSE2603, GSE27155, GSE27678, GSE29044, GSE30219, GSE30784, GSE3292, GSE33630, GSE3524, GSE35896, GSE37745, GSE39491, GSE39612, GSE43580, GSE45670, GSE4922, GSE50081, GSE51010, GSE6532, and GSE7553. These datasets were normalized using Affymetrix Expression Console. Bayesian Factor Regression Methods (BFRM) [60] was used to combine datasets and remove batch effects.(http://www.stat.duke.edu/research/software/west/bfrm/download.html).
Gene expression signatures were derived using significance analysis of microarrays [61] to detect the genes that were differentially regulated for each tumor histology as illustrated in Fig 1 and S1–S5 Figs. Venn diagrams were generated using online tool available at the following URL: http://bioinformatics.psb.ugent.be/webtools/Venn/. Unsupervised hierarchical clustering was done using Cluster 3.0 and Java Tree View. The color scheme for the heatmap and sample legends were made using Matlab. Gene set enrichment analysis [62] and single sample gene set enrichment analysis was done by converting our gene expression data and gene lists to the specified file formats and using these available modules hosted by Gene Pattern[63]. Tumors sorting for Fig 2 was by sorting tumors for the maximum single sample GSEA score for upregulated genes of any histological type. Pathway and transcription factor overrepresentation analysis was done using Innate-DB[64] and using the Broad Institute’s molecular signatures database ‘investigate gene sets’ web tool[65]. Kaplan-Meier analysis was done using the http://geneanalytics.duhs.duke.edu/Surv_sig.html tool. Samples were assigned to groups based on being above or below the median population value.
|
10.1371/journal.ppat.1000304 | Differential Ligand Binding to a Human Cytomegalovirus Chemokine Receptor Determines Cell Type–Specific Motility | While most chemokine receptors fail to cross the chemokine class boundary with respect to the ligands that they bind, the human cytomegalovirus (HCMV)-encoded chemokine receptor US28 binds multiple CC-chemokines and the CX3C-chemokine Fractalkine. US28 binding to CC-chemokines is both necessary and sufficient to induce vascular smooth muscle cell (SMC) migration in response to HCMV infection. However, the function of Fractalkine binding to US28 is unknown. In this report, we demonstrate that Fractalkine binding to US28 not only induces migration of macrophages but also acts to inhibit RANTES-mediated SMC migration. Similarly, RANTES inhibits Fractalkine-mediated US28 migration in macrophages. While US28 binding of both RANTES and Fractalkine activate FAK and ERK-1/2, RANTES signals through Gα12 and Fractalkine through Gαq. These findings represent the first example of differential chemotactic signaling via a multiple chemokine family binding receptor that results in migration of two different cell types. Additionally, the demonstration that US28-mediated chemotaxis is both ligand-specific and cell type–specific has important implications in the role of US28 in HCMV pathogenesis.
| Chemokines are small cytokines that are critical for recruiting and activating the cells of the immune system during viral infections. A number of viruses, including the large herpes virus human cytomegalovirus (HCMV), encode mechanisms to impede the effects of chemokines or have gained the ability to use these molecules to their own advantage. HCMV encodes multiple chemokine receptors including US28, which binds two different classes of chemokines namely the CC and CX3C families. In this report, we demonstrate that US28 binding to a CC chemokine elicits different responses compared to when binding to Fractalkine, the only CX3C chemokine. RANTES (CC chemokine) binding to US28 mediates smooth muscle cell migration, but Fractalkine blocks this process in a dose-dependent manner. However, Fractalkine binding to US28 can specifically mediate the migration of macrophages, another important cell type during viral pathogenesis. We explored the intracellular signaling pathways responsible for each migration event and determined that they differ in the G-proteins that are coupled to US28 following addition of ligand and that this occurs in a cell type–specific manner. These results provide a new mechanism for HCMV acceleration of vascular disease via the specific migration of macrophages and provide the first example of cell type–specific migration via multiple chemokines binding to a single receptor.
| All β and γ-herpesviruses encode molecules with the potential to modulate the host immune response, including chemokines and/or chemokine receptor homologs. The β-herpesvirus human cytomegalovirus (HCMV) encodes a CXC-chemokine (UL146), a potential CC-chemokine (UL128), and four potential chemokine receptors (US27, US28, UL33 and UL78) with the most characterized being US28 [1]–[4]. Chemokines are small, inducible cytokines that have critical roles in the induction and promotion of cellular migration and activation upon binding 7-transmembrane spanning G-protein coupled receptors (GPCRs). There are four major chemokine subfamilies that are categorized according to the spacing of the first two conserved amino-terminal cysteine residues: CC-, CXC-, CX3C- and XC-. Most chemokine receptors bind a limited subset of ligands belonging to a single subfamily. The ability to bind multiple ligands from different chemokine subfamilies is unique to a select few receptors including the Duffy antigen/receptor for chemokine (DARC-receptor) and the HHV-8-encoded chemokine receptor Orf74. These receptors have been reported to bind to both CC- and CXC-chemokines [5]–[7]. US28 also binds multiple ligands from different subfamilies. US28 contains homology to CC-chemokine receptors, with greatest homology to CCR1 [8] and binds to a broad spectrum of CC-chemokines with high affinity including: RANTES, MCP-1, MIP-1α and MIP-1β [9]. Interestingly, US28 also binds the CX3C-chemokine Fractalkine and with greater affinity than CC-chemokines. Although the N-terminal 22 amino acids of US28 have been shown to be required for binding of both chemokine classes [10], binding is not competed with saturating quantities of selected CC-chemokines [11]. Therefore, Fractalkine is predicted to bind unique regions of US28 compared to the CC-chemokines. Indeed, recent mutagenesis studies of the US28 N-terminus revealed that the phenylalanine residue at position 14 of US28 is important for binding of CC chemokines but is dispensable for Fractalkine binding, while mutation of tyrosine 16 negatively effects binding of both classes of chemokines [12].
Binding of chemokines to their respective receptors stimulates the cell type-dependent activation of a plethora of cellular signaling pathways specific to the chemokine/receptor pair. The CC-chemokines are known to be potent stimulators of cellular activation through US28. For example, in 293 cells, RANTES binding to US28 activates ERK-1/2 pathways through the G-proteins Gαi1 and Gα16 [13]. We have previously demonstrated that US28-mediated SMC migration is ligand-dependent requiring either exogenously added RANTES or endogenously expressed MCP-1 [14]. This migratory process is not blocked by treatment with pertussis toxin (PTX), a Gαi/o G-protein inhibitor, suggesting that other G-proteins are involved in this event [14]. Subsequent studies revealed that US28 couples with Gα12/13, promoting SMC migration and ligand-dependent signaling through the small G-protein RhoA [15]. US28 mediated SMC migration is also sensitive to treatment with protein tyrosine kinase (PTK) inhibitors, and the PTKs focal adhesion kinase (FAK) and Src are activated in US28 expressing cells upon RANTES binding [16]. Dominant negative inhibitory FAK molecules blocked US28 induced SMC migration suggesting that FAK activation is critical for US28 mediated SMC motility [16].
Although US28 binding to CC-chemokines leads to the activation of a multitude of cellular signaling pathways, the only activities associated with US28 binding to Fractalkine involve the modulation of constitutive signaling activity [17]–[19]. Treatment of US28 expressing cells with Fractalkine or the US28 synthetic inverse agonist VUF2274 leads to substantial decreases in the ability of US28 to promote the Gαq/11 dependent constitutive activation of phospholipase-C (PLC) and NF-κB, whereas MCP-1 and RANTES have only negligible effects on constitutive signaling levels [10],[18]. Additionally, Fractalkine treatment of US28 expressing HEK293A cells reduces constitutive US28 phosphorylation [19] and steady state levels of surface US28, but has little influence on the rapid endocytosis observed in HeLa cells [17]. The ability of US28 to efficiently bind ligands from multiple chemokine subfamilies coupled with the vastly different signaling responses elicited by divergent ligands is intriguing and suggests that US28 signaling is not only ligand and cell-type dependent, but also ligand-specific.
In the current study, we investigate the signaling potential of US28 upon stimulation with CC-chemokines compared to the CX3C-chemokine Fractalkine. We demonstrate that Fractalkine binding to US28 inhibits the ability of CC-chemokines to induce SMC migration. RANTES, MCP-1, and Fractalkine binding to US28 induced similar levels of FAK activation in fibroblasts. Overexpression studies indicate that RANTES-mediated stimulation of FAK occurs via a Gα12-dependent mechanism while Fractalkine utlilzes Gαq. In contrast to SMC, when US28 is expressed in macrophages, Fractalkine stimulation produces robust migration These results suggest that US28-signaling is ligand-specific and cell type-specific, and that RANTES and Fractalkine promote differential G-protein coupling leading to the activation of alternative signaling pathways depending on the cell-type and the complement of endogenously expressed G-proteins.
The unique ability of US28 to bind both CC- and CX3C-chemokine ligands raises the question of whether US28 signaling is not only ligand-dependent, but also ligand-specific [9],[13],[20],[21]. To determine whether US28 signaling and SMC migration are ligand-specific, we performed SMC migration and signaling assays on US28 adenovirus expressing primary rat SMC in the presence of RANTES or Fractalkine. In this assay, RANTES readily induced US28-mediated SMC migration, however, increasing concentrations of Fractalkine failed to stimulate cellular motility above Ad-tet-transactivator (Trans) infected and RANTES stimulated controls, indicating that not all US28 ligands evoke the same functional response (Figure 1A). Visual analysis of the cells prior to and following the migration assay indicated that the lack of migration was not due to overt cell death mediated by US28 expression and subsequent treatment with Fractalkine (data not shown). A competition assay was performed to determine whether Fractalkine inhibits the ability of RANTES to induce SMC migration. In these experiments, RANTES alone promoted SMC migration, as expected. However, Fractalkine, at concentrations as low as 10ng/ml, was sufficient to block RANTES-mediated SMC migration (Figure 1B) suggesting that Fractalkine is a competitive inhibitor to CC-chemokine induced SMC migration.
Since RANTES but not Fractalkine caused the migration of US28 expressing SMC and since Fractalkine blocks this migration event, we hypothesized that the difference in the ability to promote motility occurred at the level of signaling. To determine whether there exists a gross difference in the ability of these chemokine receptors/ligands to modulate intracellular signaling cascades, host transcriptional profiles were examined using DNA microarrays. Interestingly, the cellular gene expression profile of US28-expressing SMC stimulated with RANTES substantially differs from the profile obtained upon stimulation with Fractalkine. In fact, most of the genes that were up-regulated upon RANTES stimulation were down-regulated by Fractalkine. Specifically, RANTES binding to US28 induced expression of a number of cellular genes involved in cellular migration, while Fractalkine down-regulated many of these same genes (data not shown). These findings indicate that there are ligand-specific differences in US28 signaling that parallel the ability of either RANTES or Fractalkine to promote SMC migration.
To determine if the different phenotypic outcomes of RANTES or Fractalkine binding to US28 is reflected in differences at the level of signal transduction, we examined the ability each class of chemokine ligand to activate FAK through binding to US28. We have previously demonstrated that RANTES binding to US28 stimulates the activation of FAK, promoting a specific association between phosphorylated FAK and the adaptor protein Grb2. FAK is a critical mediator of focal adhesion turnover and plays important roles in cellular adhesion and motility. As such, it displays high basal activity levels in most cell types. For these experiments we developed a clean inducible signaling assay using FAK knockout mouse fibroblasts (FAK−/−) that have been reconstituted with an adenovirus vector expressing wild-type FAK concurrent with the addition of Ad-US28 [16]. To determine the ability of CC-chemokines and the CX3C-chemokine Fractalkine to promote US28 mediated activation of FAK and formation of active Grb2/FAK complexes, FAK−/− cells expressing US28 alone or in combination with FAK were stimulated with RANTES, MCP-1 or Fractalkine (40ng/ml) for 0 (unstimulated), 5, 10, 15 or 30 minutes. Grb2 was immunoprecipitated and active FAK associated with Grb2 visualized by western blotting for Phospho-Tyr [16]. RANTES, MCP-1 and Fractalkine all promoted US28-mediated FAK activation and formation of Grb2/FAK complexes with similar kinetics but slightly different magnitudes of activation (Figure 2A).
RANTES (CCL5)-induced signaling through US28 also promotes pronounced actin-cytoskeletal changes in multiple cell types [14]–[16]. Therefore, we also examined the ability of RANTES, MCP-1, or Fractalkine to promote actin cytoskeletal re-arrangements through US28 in FAK−/− fibroblasts. FAK−/− cells infected with adenoviruses expressing US28 and FAK were stimulated with RANTES, MCP-1, or Fractalkine (40ng/ml). Two hours post-ligand stimulation, fixed and permeabilized cells were incubated with antibodies directed against the Flag (US28) and HA (FAK) epitopes, and actin visualized by staining with Phalloidin. While RANTES, MCP-1, and Fractalkine failed to stimulate morphological changes in the absence of US28 (data not shown) each of the three ligands readily promoted actin cytoskeletal re-arrangements in US28 expressing cells (Figure 2B). Although RANTES, MCP-1 and Fractalkine differ with respect to their ability to promote SMC migration through US28, all are capable of promoting FAK activation and formation of active Grb2-FAK complexes, as well as re-organization of the actin-cytoskeleton in fibroblasts.
Our data indicate that although CC- and CX3C-chemokine stimulation of US28-expressing SMC produces different migratory phenotypes, both classes of ligands are capable of activating common pro-migratory signaling cascades in US28-expressing fibroblasts. We hypothesized that the disparate phenotypes seen in US28-expressing cell types is a result of differential coupling of G-proteins to US28. To identify the G-proteins involved in RANTES and Fractalkine stimulated FAK activation through US28, Grb2-FAK co-immunoprecipitation reactions were performed on lysates from reconstituted FAK−/− cells expressing US28. Cells were pre-treated with the Gαi/o inhibitor PTX or were left untreated and then stimulated with either RANTES or Fractalkine (40ng/ml) and Grb2/FAK co-immunoprecipitations were visualized by western blotting. Pre-treatment with PTX significantly enhanced both Fractalkine and RANTES mediated activation of FAK through US28, suggesting that both ligands promote coupling to G-proteins other than Gαi/o family G-proteins to induce FAK activation (Figure 3A). Interestingly, stimulation of US28 expressing cells with either RANTES or Fractalkine led to the PTX resistant activation of ERK-1/2. Unlike US28 mediated FAK activation, which was enhanced by PTX, ERK-1/2 activation was not affected by PTX pre-treatment. Therefore, US28 mediated activation of ERK-1/2 in reconstituted FAK−/− cells is independent of Gαi/o family G-proteins, differing from PTX sensitive MCP-1 and RANTES induced ERK-2 activation by US28 observed in 293 cells [13].
We have previously determined that US28-mediated SMC migration requires the Gα12/13-dependent activation of RhoA [15]. Additionally, Fractalkine stimulation of US28 has been used as an inhibitor of Gαq/11-mediated constitutive activation of phospholipase-C (PLC) and NF-κB [10]. Since RANTES and Fractalkine induced activation of FAK through US28 is independent of Gαi/o family G-proteins, and US28 is known to signal through Gα12 to promote cellular migration in SMC, we assessed the role of Gα12 in promoting RANTES and Fractalkine mediated activation of FAK. Reconstituted FAK−/− cells infected with adenoviruses expressing US28 and wild-type Gα12 were stimulated with either RANTES or Fractalkine. FAK activation was determined using Grb2-FAK co-immunoprecipitation reactions as described above. Introduction of high levels of Gα12 had little effect on the kinetics of FAK activation by RANTES, but significantly delayed and reduced FAK activation by Fractalkine (Figure 3B). In similar assays, over-expression of Gαq abrogated RANTES-mediated FAK activation while Fractalkine mediated FAK activation was unaffected by expression of this G-protein (Figure 3C). These data are consistent with the observation that Fractalkine binding to US28 specifically decreases the constitutive activation of PLC and NF-κB via a Gαq/11 dependent mechanism. This study, combined with our previous findings, shows that US28 G-protein coupling occurs in a ligand-specific manner wherein RANTES promotes US28 coupling to Gαi/o, Gα16 and Gα12/13, while Fractalkine promotes US28 coupling to Gαq [15].
Although Fractalkine binding to US28 fails to promote migration in SMC, we have demonstrated that Fractalkine stimulation causes cytoskeletal rearrangements and activates pro-migratory signaling pathways in fibroblasts via Gαq. Given that the endogenous complement of G-proteins differs between cell types, we hypothesized that Fractalkine binding to US28 may mediate migration of a second HCMV-susceptible cell type. Fractalkine (CX3CL1), is the only known CX3C chemokine and is unique among chemokines in that it has both membrane bound and soluble forms. Fractalkine is both a chemotactic signal for monocytes and sufficient for monocyte activation and adhesion under flow conditions [22]. HCMV infection of monocyte/macrophages is an important dissemination vehicle in vivo [23],[24]. We hypothesized that the capacity of US28 to bind Fractalkine with high affinity, in addition to CC-chemokine ligands, may play a role in HCMV infection of monocytes. and that, in contrast to SMC, Fractalkine stimulus may be pro-migratory in US28-expressing monocytes. We attempted these experiments in human monocytes in the context of HCMV infection. However, the presence of endogenous chemokine receptors (including RANTES-binding CCR1 and CCR5 as well as the human fractalkine receptor CX3CR1) and endogenous chemokine ligands in these cells made the experimental results difficult to interpret. To compensate for technical difficulties, US28 was expressed from an adenoviral vector in the context of a rat macrophage cell line. We reasoned that compared to ligands produced in human monocytes fewer endogenous rat chemokines would functionally interact with US28 and, similarly, fewer endogenously expressed rat chemokine receptors would signal productively in response to stimulation with recombinant human chemokines.
Using a low temperature, low volume infection protocol, rat macrophages were infected with adenovirus expressing US28 at various MOI (Figure 4A). FACS analysis was used to demonstrate US28 expression in approximately 70% of permeablized macrophages stained for the HA tag (Figure 4B) and that US28 is expressed on the cell surface of adenovirus-infected macrophages. (Figure 4C). The response of US28-expressing macrophages to treatment with recombinant human RANTES and Fractalkine was assessed using a quantitative in vitro migration assay. In these assays, Fractalkine induced robust migration of US28-expressing macrophages (Figure 4D). Statistically significant migration was seen at very low (1ng/ml) concentrations of chemokine but not in control cells expressing only Trans. In contrast, RANTES caused weak migration of macrophages presumably due to low levels of Gα12 expressed in these cells. Only the highest dose (80ng/ml) of RANTES achieved statistical significance and this response was not titratable with increasing chemokine as seen with Fractalkine stimulation (Figure 4D and 4E). These results suggest that Fractalkine is the predominant chemotactic signal in US28-expressing macrophages. We performed chemokine competition experiments similar to those performed in SMC (Figure 1B) to determine whether RANTES and Fractalkine have any synergistic effect on US28-mediated macrophage migration. Fractalkine-dependent macrophage migration was inhibited in a dose-dependent manner by increasing concentrations of RANTES as a competing ligand (Figure 4F). These results show that, in direct contrast to results seen in SMC, RANTES is a competitive inhibitor of Fractalkine mediated macrophage migration.
To demonstrate that the US28-induced macrophage migration specifically required US28-Fractalkine interaction, we expressed the US28 mutant (Y16F), which is deficient in RANTES and Fractalkine binding [12]. US28-Y16F is efficiently expressed in adenovirus-infected macrophages (Figure 5A) and is present on the cell surface (Figure 5B). Macrophages expressing Y16F mutant did not migrate in response to Fractalkine (Figure 5C). Taken together these results demonstrate that US28-expressing macrophages respond to stimulus with recombinant human chemokine in a ligand-specific manner. Furthermore, in contrast to the CC-chemokine mediated migration phenotype in SMC, Fractalkine binding to US28 produces robust migration in macrophages. These are the first data to demonstrate a specific cellular phenotype mediated by US28 binding to Fractalkine and the first example of ligand-specific chemotaxis mediated by a multiple chemokine family binding receptor.
In the current report, by examining the functional responses, signaling characteristics, and transcriptional profiles induced by US28 upon binding a diversity of ligands, we demonstrate that not only is US28-signaling ligand and cell-type dependent but also ligand and cell type-specific. While RANTES stimulation of US28 causes robust SMC migration, Fractalkine provides an anti-migratory signal in these cells. Similarly, RANTES but not Fractalkine increases transcription of genes involved in SMC migration. In contrast, Fractalkine but not RANTES provides a strong chemotactic stimulus for US28-expressing macrophages, and RANTES is able to competitively inhibit Fractalkine-mediated macrophage migration. Interestingly, while these ligands display differential signaling characteristics with respect to cellular migration, they both are capable of activating FAK and producing actin cytoskeletal rearrangements in fibroblasts. Importantly, we demonstrate that these phenotypic differences can be attributed to RANTES and Fractalkine causing differential G-protein coupling to US28. Fractalkine induced-US28 signaling occurs in a Gαq-dependent manner and is abrogated in the presence of Gα12 but not by PTX. However, RANTES induced migration and signal transduction occurrs in a Gα12 dependent manner and is blocked by overexpression of Gαq. Ultimately, our findings indicate that US28 binding to RANTES or Fractalkine results in differential G-protein coupling/activation leading to unique functional consequences.
While most chemokine receptors bind a limited subset of chemokines from a single chemokine subfamily, there are three examples of chemokine receptors that bind chemokines from multiple subfamilies: the DARC-receptor, Orf74 of HHV-8, and US28 [5]–[7],[11]. To date DARC, which binds both CC- and CXC-chemokines (CCL2, CCL5, CXCL1, and CXCL8), is the only true chemokine sink because this receptor binds and internalizes these ligands without inducing signaling events. Orf74 has also been demonstrated to bind both CC- and CXC-chemokines; however, there is a significant difference in the affinity of individual ligands for this receptor. Despite being referred to as an IL-8 receptor, Orf74 has greater affinity for GRO peptides (αβγ) than for IL-8 [6]. In competition binding assays with IL-8, Orf74 binding to the CC-chemokines MIP-1α, MIP-1β, MCP-1, and RANTES was virtually undetectable, while MCP-3 and aminooxypentane (AOP)-RANTES display affinities in the 200nm range. Interestingly, the structurally distinct, non-ELR containing CXC-chemokines IP-10 and SDF-1α can displace IL-8 binding, and function as efficient inverse agonists of Orf74 signaling at nanomolar concentrations [6]. Although Orf74 binds to chemokines from multiple chemokine subfamilies, Orf74 signaling only occurs in the presence of ELR, and pro-inflammatory/angiogenic chemokines, whereas the angiostatic non-ELR CXC-chemokines function as efficient inverse agonists. Unlike Orf74, US28 binds multiple ligands from different chemokine subfamilies with near equal affinity [9],[11], and as we demonstrate in the current report, these distinct ligands promote cellular activation upon binding US28. Therefore, to date, US28 is the only chemokine receptor capable of signaling upon binding ligands from multiple chemokine subfamilies.
We have demonstrated that both MCP-1 and RANTES promote US28-mediated SMC migration [14]. While Fractalkine is a known modulator of US28-induced constitutive signaling activity [18],[19], we have shown that Fractalkine does not promote US28-mediated SMC migration and actually inhibited RANTES mediated SMC migration. In accordance with these ligand-specific functional responses, microarray analysis of US28-expressing SMC stimulated with either RANTES or Fractalkine revealed profound differences at the level of gene induction. In the context of CMV-infection of SMC, the ability of US28 to adhere to mobilized Fractalkine, coupled with our finding that this chemokine reverses transcriptional activation required for cellular migration in SMC, suggests that Fractalkine may arrest US28-induced SMC migration and promote the subsequent adhesion of US28 expressing SMC to the vascular endothelium. The migration of HCMV infected and US28 expressing SMC from the vessel media to inflammatory sites in the vessel intima and the subsequent adhesion and accumulation of SMC in the vessel intima may have important implications in the dissemination and in vivo pathogenesis of HCMV, as well as in the exacerbation of vascular disease.
In this study, we also demonstrate that Fractalkine causes robust migration of US28-expressing macrophages, which is the first known cellular phenotype associated with Fractalkine binding to US28. This finding indicates that, in addition to being ligand-dependent and ligand-specific, the function of US28 signaling is also cell type-specific. Our finding that Fractalkine causes migration of US28-expressing macrophages suggests a further role for US28 in the development of vascular disease. US28 has been shown to be expressed in HCMV-infected peripheral blood mononuclear cells [25]. Foam cells found in atherosclerotic lesions originate as circulating monocytes and chemokines play an important role in the deposition of monocytes in lesions [26]. In particular, Fractalkine expression is known to be important for the development of atherosclerosis in mouse models of heart disease via recruitment of macrophages into atherosclerotic plaques [27],[28]. Expression of membrane-bound Fractalkine can be induced on endothelial cells by numerous cytokines including IFN-γ, TNF-α and IL-1, resulting in the recruitment of inflammatory cells and contributing to chronic inflammatory vascular diseases such as atheroscleorosis, restenosis following angioplasty and transplant vascular sclerosis [29]. Unlike other chemokines which are secreted as soluble molecules that must associate with proteoglycans and other components of the extracellular matrix to establish chemokine gradients [30], Fractalkine is generated as a membrane bound ligand with the chemokine domain presented at the top of the cell-bound mucin-like stalk [29],[31]. In many instances this ligand is more effective than other ligands in promoting leukocyte activation and migration. Our current findings suggest a secondary mechanism for US28 in CMV-mediated vascular pathology by which circulating CMV-positive monocytes infiltrate atherosclerotic plaques mediated by Fractalkine binding to US28.
We demonstrate for the first time that Fractalkine is a potent agonist capable of inducing cellular migration in macrophages and activation of signaling pathways upon binding US28. Prior to this study, Fractalkine had been employed as a modulator of US28-mediated constitutive signaling activity. Some of the signaling pathways activated by Fractalkine were similar to those activated by the CC-chemokines. For example, RANTES, MCP-1, and Fractalkine all display similar abilities to induce ERK-1/2, actin cytoskeletal rearrangements and formation FAK-Grb2 complexes in fibroblasts. Pre-treatment with PTX enhanced Fractalkine mediated FAK activation through US28, which indicated that Fractalkine promoted US28 coupling to G-proteins other than Gai/o. Expression of Gα12 delayed and reduced FAK activity via Fractalkine signaling through US28 but had no effect on RANTES/US28 activation of FAK. Importantly, overexpression of Gαq blocked RANTES signaling to FAK but had no effect on Fractalkine-mediated FAK activation. In a number of different activation scenarios FAK is a known point of signaling convergence and has been demonstrated to be phosphorylated in response to Gαq/11, Gαi/o, and Gα12/13 coupled receptors in various cell types and signaling environments [32]–[35]. In one study, lysophosphatidic acid (LPA) signaling stimulated both membrane association and autophosphorylation of FAK but these two effects were separable and mediated by different G-alpha subunits (Gαi1 and Gα12/13, respectively) presumably via signaling from two different LPA receptors [33]. Importantly, in a receptor-decoupled system of constitutively active G-alpha subunits, significant FAK phosphorylation can be observed via signaling through Gαq, Gα12 and Gα13 [35]. These results are consistent with our observations that both RANTES and Fractalkine binding to US28 can activate FAK via different signaling cascades mediated by different G-proteins. Our results suggest that overexpression of off-target G-proteins inhibit signaling from a particular ligand via competition with the G-proteins that would normally promote signaling from the ligand-bound activated receptor. Therefore, in these experiments overexpression of Gα12 may act as a dominant inhibitory molecule that prevents Gαq-receptor interactions, which would normally activate FAK following Fractalkine coupling to US28. Overexpression of Gαq prevents Gα12 coupling to the RANTES-bound activated form of US28 thereby abrogating the downstream signaling to FAK. Therefore, RANTES stimulates varying signaling pathways through different G-proteins in SMC (Gα12-dependent) and fibroblasts (Gαi/o independent). Fractalkine signals from US28 via coupling of Gαq in fibroblasts, SMC and macrophages. Together these findings demonstrate that not only is US28 signaling ligand-dependent and ligand-specific, it utilizes differential G-protein coupling to produce cell-type specific signaling and differential phenotypic responses.
In this report, we demonstrate that similar to RANTES and MCP-1, Fractalkine is a potent US28 agonist that promotes migration in macrophages, robust signaling through FAK and ERK1/2 and induces actin cytoskeletal rearrangements in fibroblasts. Unlike RANTES and MCP-1, Fractalkine fails to induce SMC migration, or increase expression of cellular genes involved in motility and signaling in SMC, thus demonstrating that US28 signaling is ligand-specific and cell type-specific. In addition, the US28 ligand-specific and cell-type dependent activation of differential signaling pathways suggest that this chemokine receptor has the capacity to couple to different G-proteins depending upon the ligand bound and the cellular G-protein environment. Therefore, US28 binds to a diversity of chemokines, which promote US28 coupling to multiple G-proteins, eliciting functional signaling through these various G-proteins. HCMV encounters and infects a multitude of distinct cell types in vivo including fibroblasts, monocyte/macrophages, endothelial cells and SMC. These cell types differ substantially with respect to the G-proteins that they express. The ability of US28 to respond to multiple signaling environments and couple to multiple G-proteins may have important implications in the persistence and pathogenesis of HCMV in these different cell-types.
The life-extended human pulmonary artery smooth muscle cell line, PAT1 [15] were maintained in Medium 199 supplemented with 20% fetal calf serum (FCS) and penicillin-streptomycin-L-glutamine (PSG; Gibco). For migration and microarray experiments, PAT1 cells were utilized between passage 5 and 30 post-telomerization. Primary F344 rat smooth muscle cells (RSMC) were maintained in Dulbecco's modified Eagle's Medium (DMEM) with 10% FCS and PSG. RSMC were used between passage 5 and 20. NR8383 rat alveolar macrophages were maintained in RPMI with 10% FCS and PSG. Mouse FAK−/− fibroblasts were maintained on gelatin coated culture dishes in DMEM supplemented with 10% FCS, PSG, non-essential amino acids (Cellgro), and G418 (Sigma; 500 µg/ml) as previously described [36],[37]. FAK−/− cells used in experiments were between passage 5 and 15.
Recombinant human RANTES, MCP-1, and Fractalkine were purchased from R&D Systems. Anti-Grb2 (C-7), anti-phosphotyrosine (PY99), anti-Gα12 (S-20), anti-Gαq (E-17) and anti-HA (F-7) antibodies were purchased from Santa Cruz Biotechnology. Phospho-specific ERK-1/2 (Thr202/Tyr204) and total ERK-1/2 antibodies were from Cell Signaling Technologies. Anti-M2-Flag antibody (F-3165) was purchased from Sigma. Secondary anti-mouse and anti-rabbit horseradish peroxidase (HRP)-conjugated antibodies (NA934V and NA931V) were purchased from Amersham.
Adenoviruses expressing Gα12 Gαq, WT-FAK, US28-Flag, and US28-HA were previously described [14]–[16]. Adenovirus vectors expressing US28-Y16F-HA were constructed by mutagenesis of the US28-HA construct in pAdTet7. This vector contains the tet-responsive enhancer within a minimal CMV promoter followed by the SV40 late poly(A) cassette, adenovirus E1A, and a single loxP site to increase recombination frequency. Complementary 30bp primers containing coding sequence for amino-acids 2–25 of US28-HA and including a phenylalanine codon in place of the tyrosine at position 16 (5′-ACGACGGAGTTTGACTTCGACGATGAAGCG-3′ and 5′-CGCTTCATCGTCGAAGTCAAACTCCGTCGT-3′) were used to PCR amplify mutated vector using Pfu Turbo DNA Polymerase (Stratagene). Non-mutated methylated parental DNA was digested using DpnI and mutated plasmid was propagated in DH5α. Recombinant adenoviruses were produced by pAdTet7 US28-Y16F-HA construct co-transfection of 293 cells expressing the Cre-recombinase with adenovirus DNA (Ad5-ψ5) that contains an E1A/E3-deleted adenovirus genome [38]. Recombinant adenoviruses were expanded on 293-Cre cells and the bulk stocks were titered on 293 cells by limiting dilution. Gene expression was driven by co-infection with Ad-Trans expressing the Tet-off transactivator as previously described [14].
To monitor surface expression of recombinant proteins and total adenovirus transduction, adenovirus-infected cells were fixed in 2% PFA for 15min, washed 2× with PBS, blocked for 15min on ice in Fc Block (PBS+20%Normal goat serum (NGS)+0.1% sodium azide). To determine the rate of adenovirus transduction, cells were permeablized with PBS containing 0.2% Saponin and 0.02%NGS for 15min on ice. For both cell surface and intracellular staining assays the cells were incubated for 30min with either mouse IgG2b isotype control or primary αHA antibody diluted 1∶200 in FACS wash buffer (PBS+1% NGS+0.01% sodium azide +/− 0.2% saponin as appropriate) on ice and washed 2× with FACS wash buffer. Primary antibody staining was detected with anti-mouse Alexa-Fluor597 antibody diluted 1∶1000 in FACS wash. After 20 min incubation with secondary antibody on ice cells were washed as above and surface expression was quantified using flow cytometry (FACS Calibur, BD Biosystems). Data analysis was performed using FlowJo software v8.8 (Treestar Inc.).
FAK−/− fibroblasts were grown in 0.1% gelatin coated 4-well chamber slides (Nalge-Nunc). US28 and/or FAK was expressed using the adenovirus vectors described above and were left untreated or were treated with MCP-1, RANTES or Fractalkine (20ng/ml) for 2 hrs. The cells were washed in PBS and fixed in phosphate buffered 2% paraformaldehyde (PFA) for 15 minutes at r.t., then permeabilized and blocked with 0.2% Saponin+0.02% BSA in PBS for 15min at r.t. Thereafter, the cells were incubated with antibodies against US28-Flag epitope or FAK-HA epitope in a 1∶200 dilution for 1 hr at room temperature. Cells were washed three times in blocking buffer and binding of the primary antibody was detected with a fluorescein isothiocyanate-tetramethyl (FITC) conjugated goat anti-mouse or rhodamine conjugated goat anti-rabbit antibody for 1 hr at room temperature. At this time the cells were also stained for actin using Phalloidin (Molecular Probes, Eugene, OR) to monitor alterations in cellular actin cytoskeleton induced by US28 and FAK. Fluorescence positive cells were visualized on an inverted Applied Precision Deltavision™ deconvolution microscope.
FAK−/− cells were plated in 10cm culture dishes and serum starved for 6 hrs upon achieving 50% confluence. The cells were co-infected with Ad-Trans and/or Ad-US28 and/or Ad-FAK WT at MOI 50. After 16 hrs the cells were stimulated with RANTES (40ng/ml), Fractalkine (40ng/ml), or MCP-1 (40ng/ml) and then harvested at times 0 (unstimulated), 5, 10, 15, and 30 minutes post ligand addition. Cells were lysed in modified RIPA buffer containing 1% Triton X-100, 1% sodium deoxycholate, and 0.1% SDS and total Grb2 was immunoprecipitated and samples analyzed by western blotting using antibodies directed against phospho-Tyr [16]. Co-precipitation of FAK-HA was demonstrated by stripping the blots in buffer containing 0.1M Tris pH 6.8, 1% SDS, and 1% 2-mercaptoethanol and staining using antibodies directed against HA. Prior to immune-complex reactions, a total of 50 µl of cellular lysate was assayed by SDS-PAGE/western blotting for the presence of input US28 and FAK using antibodies directed against the HA-epitope present on both recombinant proteins.
SMC migration assays were performed as previously described [14]. Briefly, 4×104 primary rat SMCs were added to each upper well of a transwell (12 mm diameter, 3.0 µm pore size, Costar Corning, Cambridge, MA). Cells were serum starved for 16 hrs, and then infected with Ad-Trans only or Ad-Trans and Ad-US28-HA at MOI 200. After 4 hrs, the inserts were washed and transferred to fresh 12-well plates with chemotactic stimulus. Cells migrating to the lower chamber were quantified at 48–72 hrs p.i. via fluorescence using CyQuant (Invitrogen) and read on a Molecular Devices Flexstation® II fluorescence plate reader. Migration was determined from 4–6 independent wells per assay per condition. Mean and standard deviation were calculated. Percent of control values were generated by comparing chemokine stimulated US28-expressing cells to unstimulated control cells (Trans-only) and compared using Student's t test. P values<0.05 were considered statistically significant.
NR8383 macrophages were co-infected with Ad-Trans and Ad-US28 WT or Ad-US28-Y16F at MOI 100. Macrophages were incubated with adenovirus in 200 µl total volume for 30min at room temperature, diluted into 10ml complete RPMI and incubated at 37°C. At 72 hrs post-infection, 1×105 infected macrophages were added to the top well of a chemotaxis chamber (96-well Millipore Multiscreen, 3.0 µm pore size) with Fractalkine and/or human RANTES in the bottom chamber. Chemotaxis was allowed to proceed for 1 hr at 37°C. Top chambers were discarded and migrated cells in the bottom chamber were quantified via fluorescence using CyQuant (Invitrogen) and read on a Molecular Devices Flexstation II fluorescence plate reader. Migration was determined from 4–6 independent wells per assay per condition. Mean and standard deviation were calculated. Percent of control values were generated by comparing chemokine stimulated US28-expressing cells to unstimulated control cells (Trans-only) and compared using Student's t test. P values<0.05 were considered statistically significant. Recombinant protein levels were monitored by western blotting and flow cytometry staining for total and surface expression and equalized by adjusting the adenoviral vector MOI accordingly.
|
10.1371/journal.pgen.1000075 | Demographic History of European Populations of Arabidopsis thaliana | The model plant species Arabidopsis thaliana is successful at colonizing land that has recently undergone human-mediated disturbance. To investigate the prehistoric spread of A. thaliana, we applied approximate Bayesian computation and explicit spatial modeling to 76 European accessions sequenced at 876 nuclear loci. We find evidence that a major migration wave occurred from east to west, affecting most of the sampled individuals. The longitudinal gradient appears to result from the plant having spread in Europe from the east ∼10,000 years ago, with a rate of westward spread of ∼0.9 km/year. This wave-of-advance model is consistent with a natural colonization from an eastern glacial refugium that overwhelmed ancient western lineages. However, the speed and time frame of the model also suggest that the migration of A. thaliana into Europe may have accompanied the spread of agriculture during the Neolithic transition.
| The demographic forces that have shaped the pattern of genetic variability in the plant species Arabidopsis thaliana provide an important backdrop for the use of this model organism in understanding the genetic determinants of plant natural variation. We investigated the demographic history of A. thaliana using novel population-genetic tools applied to a combination of molecular and geographic data. We infer that A. thaliana entered Europe from the east and spread westward at a rate of ∼0.9 kilometers per year, and that its population size began increasing around 10,000 years ago. The “wave-of-advance” model suggested by these results is potentially consistent with the pattern expected if the species colonized Europe as the ice retreated at the end of the most recent glaciation. Alternatively, it is also compatible with the possibility that A. thaliana—a weedy species—may have spread into Europe with the diffusion of agriculture, providing an example of the phenomenon of “ecological imperialism” described by A. Crosby. In this framework, just as weeds from Europe invaded temperate regions worldwide during European human colonization, weeds originating from the source region of farming invaded Europe as a result of the disturbance caused by the spread of agriculture.
| Arabidopsis thaliana is an important model organism for plant biology, serving as a focal species for studies of plant physiology, molecular biology, and genetics [1]–[4]. Its use as a model species is facilitated by its short generation time in the laboratory, its production of large numbers of seeds, and its reproduction primarily by self-fertilization.
Many of the same traits that contribute to the utility of A. thaliana as a model organism are important in determining the niche of the species in its natural environment. Its rapid flowering, self-fertilization, and extensive seed production are characteristic of colonizing species that grow in open or recently disturbed habitats [5],[6]. From an ecological standpoint, due to its status as a colonizing species, A. thaliana can be viewed as a weed.
A. thaliana is frequently described as native to the Eurasian landmass [6],[7], and in recent times it has been among the group of weeds from Europe that have invaded North America and Australia since the time of European colonization [8],[9]. However, relatively little is known about the prehistoric spread of the species into Europe. Because pollen from A. thaliana is very similar to that of many other species from the Brassicaceae family [10], it is often undetectable in surveys of past plant geographic distributions. Thus, investigations of patterns of present-day genetic variation have provided an important alternative method for understanding the recent history of the species.
Most European species are believed to have been restricted to southern refugia at the height of glaciation ∼18,000 BP—many in the peninsulas of Iberia, Italy, and the Balkans, and some near the Caucasus region and the Caspian Sea [11]–[13]. When the climate warmed and the ice retreated, these species expanded their ranges northwards, starting ∼16,000 BP [14]. For Arabidopsis thaliana, on the basis of population-genetic data, Sharbel et al. [15] proposed a scenario of post-glacial re-colonization of Europe from two refugia, one in the Iberian Peninsula and the other in central Asia, followed by admixture of the two ancestral populations in central and eastern Europe. However, contradicting the predictions of this model, Schmid et al. [16] found that linkage disequilibrium was more extensive in the putative source regions of Iberia and central Asia than in central Europe. Furthermore, although some population-genetic studies in A. thaliana have identified relatively unstructured patterns of genetic variation compatible with rapid range expansions from glacial refugia [17]–[20], the most recent studies of large data sets have found that genetic variation in A. thaliana shows evidence of considerable population structure [16],[21],[22]. This structure has not been extensively analyzed to determine the likely explanations for its origin, and hypotheses about the location of origin and the timing of the spread of A. thaliana have been under some debate [20],[23],[24].
In this article, we consider an alternative model for the spread of A. thaliana in Europe. Using recently developed approximate Bayesian computation and spatial modeling techniques, we re-analyzed the data of Nordborg et al. [21], one of the largest population-genetic data sets collected to date in A. thaliana. We find evidence that a migration wave from east to west is responsible for most of the genetic ancestry of European A. thaliana. We discuss this result in relation to the hypothesis of an eastern refugium, and in relation to the hypothesis that the migration of A. thaliana may have been precipitated by the spread of agriculture into Europe.
To investigate spatial population structure in European accessions of Arabidopsis thaliana, we used model-based clustering as implemented in the TESS computer program [25],[26]. Our analysis used the molecular data from 75 European accessions plus one accession from Libya (Mt-0), a total set of 876 alignments described in the study of Nordborg et al. [21] (Table S1). Using TESS, we performed an admixture analysis incorporating individual spatial coordinates and accounting for natural obstacles (see Methods and Figure S1). The program allows individuals to be distributed over Kmax clusters, estimating the most likely value for the number of clusters as a value K less than or equal to Kmax (see Methods).
The TESS runs with the smallest values of the Deviance Information Criterion, a penalized measure of how well the model underlying TESS fits the data, were obtained for Kmax greater than four (see Methods). In Figure 1, we report results for Kmax = 5 clusters. The cluster membership coefficients estimated for the central European and western European accessions suggest that clinal variation occurs along an east-west gradient separating two clusters. The western cluster grouped accessions mainly from the British Isles, France and Iberia. The eastern cluster grouped all accessions from central Europe, southern Sweden, Poland, Russia, Ukraine, and Estonia. German and Swiss accessions shared almost the same amount of membership in the western and eastern clusters. The eight northern Swedish accessions and two Finnish accessions were grouped into a separate cluster.
In previous analysis of the same data set [21], it was observed that when individual genomes were clustered by genetic similarity using the program STRUCTURE [27], European accessions sorted into K = 8 clusters, some of them corresponding to small geographic regions [21]. The TESS analysis identified a substantially lower number of actual clusters (Figure 1), consistent with more continuous allele frequency variation across geographic space and with significant isolation by distance [15],[16],[22]. Although the northern European cluster was also identified from STRUCTURE runs with K = 3 [21], some differences were found by TESS in the two continental clusters. In [21], the Iberian accessions clustered with the eastern populations, whereas TESS grouped them with the western accessions (France, British Isles). More strongly than in the STRUCTURE analysis, the TESS results suggest clinal variation of allele frequencies within central and western Europe, with Germany possibly serving as a hybrid zone separating the two clusters corresponding to these regions.
To better evaluate the direction of variation in the continental cluster, we regressed heterozygosity on geographic distance. This analysis used the approach of Ramachandran et al. [28], who showed that recurrent founder events can cause a decrease in genetic diversity in colonizing populations. Assuming a unique origin, genetic diversity is then predicted to decrease approximately linearly with geographic distance from the origin.
All accessions from the northern Sweden sample, as well as a few accessions that were poorly geographically connected to other accessions, were removed from the regression analysis. The remaining accessions were grouped into seven samples (Table S2), defined on the basis of geographic and genetic proximity. To minimize the sensitivity of the regression analysis to a particular geographic pooling of European accessions, we repeated the regression study for several combinations of seven modified samples, and the results reported can be viewed as representative of these various combinations.
For each of 300×180 points on a two-dimensional lattice covering Europe, we computed distances from each lattice point considered as a potential source for the geographic expansion of A. thaliana. The Pearson correlation coefficients of genetic diversity with distance from the source were estimated and plotted on the grid. The correlations were negative (∼ −0.5) in the east, and they were positive (∼ +0.3) in southwestern Europe. Assuming a unique site of origin, Figure 2 provides evidence that the pattern of heterozygosities is best explained by spatial expansion originating from the east.
Because this analysis is based on a relatively limited geographic sample, it is possible that it is affected by the peculiarities of this sample. Therefore, to assess the possibility of bias due to non-uniform and sparse geographic sampling, we performed spatially explicit range expansion simulations that reproduced the geographic sampling scheme of the actual data (Text S1). Assuming an origin in Anatolia (west Asia), we indeed observed a considerable shift of the position of the estimated origin to the southwest of the true origin (Figure S2). Because our data analysis identified a best-fitting origin in the Balkan region, it is thus possible that the true origin is potentially localized farther to the northeast.
Inference of demographic parameters and the choice of a best-fitting demographic model for the data were performed using an approximate Bayesian computation (ABC) analysis [29]–[31]. ABC approaches bypass the computational difficulties of using explicit likelihood functions by simulating data from a coalescent model. These methods rely on the simulation of large numbers of data sets using parameter values sampled from prior distributions. A set of summary statistics is then calculated for each simulated sample, and each set of summaries is compared with the values for the observed sample, sobs. Parameter values that have generated summary statistics close enough to those of the observed data are retained to form an approximate sample from the posterior distribution, enabling parameter estimation and model choice (see Methods).
The ABC analysis was limited to a subset of 64 individuals representing the central European and western European populations. We restricted the analysis to the non-coding part of the genomic data, using the intron and the intergenic sequences only (648 loci). Simulated data also included 648 corresponding loci, each paired to have the same length as a locus in the observed data. The loci were assumed to be in linkage equilibrium, in agreement with the median ∼100 kb distance between fragments in the genome-wide data [21] and with levels of linkage disequilibrium that decay within ∼10 kb in A. thaliana [21],[32].
Coalescent simulations were performed under four demographic scenarios (Models A–D). Model A has a constant population size, N0. Model B has an exponentially growing population size (present size, N0, ancestral size, N1, time since the onset of expansion, t0). In model C, the population size was constant in the distant past as well as in the recent past, and the growth was exponential between the two periods of constant population size (present size, N0, ancestral size, N1, time since the onset of expansion, t0, time since the end of expansion, t1). Model D is similar to model B, but it includes an ancient bottleneck before expansion. The prior distributions used in the four models are described in Table S3. Twelve summary statistics were used to capture genomic information at the 648 loci (see Methods). To make quantitative model comparison possible, we evaluated the evidence of model 1 against model 2 (where 1 and 2 are chosen among A, B, C and D) using an approximation of the Bayes factor [33]. Pritchard et al. [30] computed the Bayes factor as the ratio of the acceptance rates in Models 1 and 2. Including smooth weighting to more heavily weight the simulations that produced results that more closely matched the observed data [29], we approximated the Bayes factor aswhere Kδ is the Epanechnikov kernel and si,1 and si,2 are the ith vectors of summary statistics simulated under models 1 and 2 (see Methods).
Among all the scenarios, variants of the four models with variable mutation rates across loci were given higher statistical support, measured by the Bayes factor, than were models with fixed mutation rates - reflecting the high heterogeneity of diversity estimates among loci [21]. The best-supported model was model C with variable mutation rates, which assumed a past rapid expansion followed by a constant-size population phase (see Figure 3). The Bayes factor BA,B = 0 indicates that the model with constant population size (model A) was totally unsupported. The exponential growth model (model B) was the second best-supported model, and the evidence supporting model C against model B was moderate (BC,B = 1.9, see Figure 3). The scenario in which the population experienced a bottleneck before expansion was rejected, but less decisively than model A (model D, BD,B = 0.7).
Table 1 displays the estimates of the parameter values under the variants of model B and C with variable mutation rates. The time of onset of the expansion was dated at t0 = 10,000 BP (model B) and t0 = 12,000 BP (model C) using the Maximum A Posteriori (MAP) estimate (Figure S3). As a consequence of using broad prior bounds in the ABC analysis, similarly to [34], we observed large 95% credibility intervals. The ratio of the ancestral population size to the present population size was estimated at N1/N0 = 0.3, but the large credibility interval (0,0.6) makes it impossible to eliminate the hypothesis of a wider expansion. The MAP estimate of the mutation rate was μ = 2.0×10−8 with credibility intervals ranging from 0.9×10−8 to 12.6×10−8. The MAP estimate for the date of the end of the expansion was t1 = 5,000 BP (see Table S4 and Table S5 for posterior estimates and Bayes factors for all eight models). To investigate the relationship between the time of onset, t0, and the length of the expansion, t0−t1, the joint posterior distribution of these two quantities was computed. Figure 4 displays this joint distribution, and it indicates a positive correlation between the two values.
Because we observed considerable difference in the TESS analysis between the northernmost accessions and the main European populations (Figure 1), we performed model fitting to assess various scenarios for the split of the northern cluster. Quantifying the genetic divergence between the central European population and the northern Swedish and Finnish population by the mean number of distinct haplotypes and the mean number of private haplotypes [35], we obtained estimates of these statistics for subsamples of size two to ten. The patterns of haplotype diversity in the central European and northern European populations were typical for pairs of separated populations in which one population has larger size than the other [36]. The central European population had, on average, 3.85 distinct haplotypes for a sample of ten individuals, and the northern European population had, on average, 2.61 distinct haplotypes for a sample of ten individuals. However, in each population, about half of the haplotypes were unique to the population (Figure 5), and the genetic variation in the northern European population was not a subset of that in the central European population.
To study the split between the northern and central European populations, we used a coalescent model for the divergence between two populations at some time T in the past, with subsequent migration at rate m between these two populations (where m is the rate in each direction). We simulated the same number of fragments as in the data for both populations, and we determined the mean across 100 replicates of the sum of squared differences (SSD) between the simulated and the observed summary statistics. In a first set of simulations we increased the split time T from 0 to 135,000 years in a model with no migration (m = 0). Figure 6 shows the results for T = 0 to T = 27,000 BP superimposed on the same summary statistics computed for the observed data. For small values of T, the fit of the simulated data to the observed data was poor, with an improvement as T increased (SSD for T = 1,350 BP, distinct haplotypes: 2.56, private haplotypes: 2.12). When T was equal to 7,000 BP, the simulated data fit the observed data quite well (SSD for T = 7,000 BP, distinct haplotypes: 0.06, private haplotypes: 0.10). When T increased beyond 13,500 BP, the fit became poorer. In a second set of simulations, we used a population divergence model that incorporated migration, and we increased the values of the migration rate, m. Figure 7 shows simulated results superimposed on the observed results. The simulations fit the data relatively well for m in the range [1],[3] when T equalled 13,500 BP, and the best values were obtained for m = 3 (SSD for m = 3, distinct haplotypes: 0.04, private haplotypes: 0.08).
As m increased above the value 3, the fit of the mean number of distinct haplotypes deteriorated. We also tested values of T>13,500 together with m>3, without finding a close fit to the observed data, and the best fit was found for a model with low migration rates. A model with high migration rates was not able to replicate the observed data under the tested conditions. Thus, it is unlikely that the split occurred more recently than ∼7,000 years ago.
In the ABC analysis the scenarios that consisted solely of population size change produced patterns of DNA sequence diversity similar to those resulting from a rapid spatial range expansion [37]. To better include geographic sampling in the analysis and to estimate the rate of spread, we modeled the process of colonization of Europe in a more explicit manner [38],[39]. Range expansion was simulated under a two-dimensional wave-of-advance model [40]. We included environmental heterogeneity, borrowing topographic information from a Geographic Information System. Assuming an origin of the colonization process to the north of the Black Sea (48°N, 35°E), we divided Europe into an array consisting of 130×180 = 23,400 demes, each representing an area of ∼2,500 km2. To account for the fact that in Europe, A. thaliana grows mainly in low-altitude landscapes, carrying capacities were set to their highest values for altitudes below 200 m and were linearly decreased for altitudes higher than 1,500 m.
It has been previously recognized that the frequency spectrum may be influenced by signals of past demographic events [41],[42]. Consequently, the fit of simulated data to the pattern of polymorphism of A. thaliana was evaluated by comparing the non-coding empirical folded frequency spectrum and frequency spectra obtained from simulated individuals located at the same coordinates as the real accessions. Simulated and observed frequency spectra were compared by using the χ2 distance (see Methods).
A coarse preliminary search found that values of migration rates and growth rates corresponding to the saturation of a deme in 100–300 years and lengths of the colonization phase around 3,000–6,000 years followed by an equilibrium migration phase yielded non-significant χ2 P-values. Thus, these values provide a reasonable explanation for the observed data. They translate into a wave-of-advance of around 0.5 to 1 km/year.
In a second stage of the analysis, we investigated the time at which the range expansion began, varying this time from t0 = 5,000 BP to t0 = 20,000 BP assuming a growth rate of r = 0.6 for the oldest dates. For the most recent dates, we increased r to 0.7, 0.9 and 1.2 so that the colonization phase ended before the present day. This analysis supported the values found by the MAP estimate from the ABC analysis. Figure 8A shows that dates around 10,000–12,000 BP are consistent with the pattern of polymorphism observed today.
To better locate the origin of A. thaliana, we investigated several potential locations, and we plotted χ2 distances between simulated spectra and the empirical spectrum on an interpolated map (Figure 8B). The χ2 values ranged from 0.03 (East) to 0.3 (Spain - North Africa). Although the map does not provide an accurate localization of the onset of range expansion, it is similar to Figure 2, providing further support to the hypothesis of an eastern origin.
Figure 9A demonstrates that the empirical folded frequency spectrum computed from non-coding nucleotides deviates from neutrality through an excess of rare alleles. Figure 9B shows one simulated folded spectrum obtained from the estimated parameters (m = 0.25, r = 0.6, N1 = 5,000 and t0 = 10,000, χ2 = 0.03, P = 0.68). For this set of parameters, the estimated speed of the wave-of-advance was ∼0.9 km/year. It is clear from the search strategy used here that these parameter settings are only likely to represent a local maximum of the probability of an evolutionary scenario, and that other settings may also provide a reasonable fit to the data.
We have performed an investigation of the population structure and demographic history of European A. thaliana, using genome-wide sequence data collected in accessions from across Europe. Our main results are as follows. (1) On the basis of spatial Bayesian analysis with TESS, we observed that most European accessions were distributed over three clusters: one northern European cluster and an east-west cline of variation across continental Europe (Figure 1). (2) The level of genetic variation is greater in the east than in the west; if a single-origin model is used for modeling genetic diversity in European populations, the most likely source location is in the east and the estimated rate of westward spread is ∼0.9 km/year (Figures 2 and 8). (3) Simulations suggest that the pattern of genetic variation is explained most parsimoniously by an ancient split of the northern cluster from the central European cluster >7,000 BP. (4) Approximate Bayesian computation suggests that the European A. thaliana population began an expansion in size ∼10,000 BP, lasting 5,000 years (Figures 4 and 8).
From a biogeographic point of view, Europe is a large peninsula with an east-west orientation, delimited in the south by a strong Mediterranean barrier. During glaciation epochs, many species likely went through alternating contractions and expansions of range, involving extinctions of northern populations when the temperature decreased, and spread of the southern populations from different refugial areas after glaciation. Such colonization processes were likely characterized by recurrent bottlenecks that would have led to a loss of diversity in the northern populations.
The idea that the refugia were localized in three areas (Iberia, Italy, Balkans) is now well-established [12], although recent studies, particularly of tree species, have begun to suggest that northern and eastern refugia could have existed [43],[44]. Comparison of colonization routes has highlighted four main suture-zones where lineages from different refugia meet [11]. Two of these suture-zones correspond to the Alps and the Pyrenees, while the two others are in Germany and in Scandinavia.
We observed that genetically diverse populations of A. thaliana were localized at intermediate latitudes, as a potential consequence of the admixture of divergent lineages colonizing the continent from separate refugia. These results are potentially consistent with the pattern expected if the species colonized Europe from two separate refugia, one in the Iberian peninsula and the second in the east, as suggested by the model of Sharbel et al. [15]. Similarity with patterns of cpDNA diversity in 22 plant species that have genetically divergent populations in Mediterranean regions was also observed for the seven geographic samples considered in the regression analysis ([13] and Figure S4). Furthermore, the presence of a highly divergent accession (Mr-0) in Italy, south of the Alpine barrier, is also compatible with the view that A. thaliana was present in Mediterranean refugia during the last glaciation.
We observed that intraspecific diversity declines away from the southeast, as predicted by a model of successive founder events during colonization. We also inferred that the putative origin of most accessions in the sample is localized somewhere in a vast eastern region, encompassing refugia such as the Caucasus region and the Balkans. The direction of diffusion from the east towards the British Isles coincides with the post-glacial re-colonization of Europe for many species such as beech, alder and ash trees, or flightless grasshopers [45],[46], and it is possible that, to a large extent, this wave of expansion erased any contribution of ancient western lineages that originated in Mediterranean refugia.
The boreal regions, in which environmental conditions are often very severe, contain the northern distribution limit of many European plants. These regions are often characterized by larger fluctuations in population size, which increase the effect of drift and can lead to increased genetic differentiation [47]. Fennoscandia has recovered its flora after the last ice age, less than 10,000 years ago, via many different routes. The presence of a suture-zone in Scandinavia indicates that this area may have been colonized by A. thaliana both from the south and from the northeast. The estimated separation time of the northern European A. thaliana population and the central European population, at least 7,000 years ago, indicates that the split between the continental and northern populations took place during the early history of the re-colonization of Europe by the species.
An alternative hypothesis to the idea of a natural spatial expansion of A. thaliana is that its spread might have accompanied the spread of farming into Europe, perhaps following an earlier post-glacial wave of colonization. Between 9,000 and 5,500 BP Neolithic farming spread across Europe from the Near East, primarily northwestwards along the Danube-Rhine axis [48]–[50]. Several aspects of our results are consistent with the hypothesis that A. thaliana was part of a group of weeds that accompanied the spread of agriculture into Europe. First, the evidence for an eastern source for European A. thaliana parallels the evidence that agriculture spread into Europe from the east [48],[49]. Putative origins in the Danube basin, west of the Black Sea, received high explanatory power in our analysis, and this area was an important way-point in the route followed by the spread of agriculture. Second, the estimated time for the beginning of the A. thaliana population size expansion parallels the time for the spread of agriculture. Third, the estimated rate of westward spread of A. thaliana, ∼0.9 km/year, fits within the range 0.6–1.3 km/year estimated for the rate of agricultural expansion [48],[51]. It is believed that Neolithic agriculture advanced into Europe along two preferred routes, a Mediterranean route and a Danubian route [52],[53]; our analysis suggests that if A. thaliana followed the spread of agriculture, then it likely followed the Danubian route.
The possible prehistoric anthropogenic spread of A. thaliana in Europe is an instance of a more general pattern documented in historical times, in which land disturbances instigated through long-distance human migrations co-occur with the spread of opportunistic organisms unintentionally brought by the migrating populations from their home region. This phenomenon of “ecological imperialism” has been used to explain the current prominence of European weeds in regions of the Americas, Australia, and New Zealand that have recently been transformed by European agriculture [54]. Several lines of evidence support the view that a similar process for the spread of weeds acted during the transformation of European landscapes by the spatial advance of agriculture - that is, that a large fraction of weeds in Europe trace their geographic distributions to the Neolithic expansion of European farming. For example, based on palaeobotanical data, Pyšek et al. [55] estimated that of the presently known prehistoric alien species of central Europe, 35% arrived there during the first thousand years after the advent of agriculture. Kreuz et al. [56] detected a chronological correlation in the number of introduced weed species in central Europe and the development of the agriculturalist Bandkeramik culture. In two weedy species of Lolium, Balfourier et al. [57] found patterns of population structure explicable by the spread of agriculture, supporting the view that the A. thaliana results could be part of a general trend for prehistoric European weeds. Another source of evidence for a large-scale prehistoric agriculturalist spread of weeds into Europe is a comparison of weed species in modern plots of land in the Czech Republic. In the study of Pyšek et al. [58], introduced weeds that entered Europe in prehistoric times were comparatively more numerous in land farmed with crops dating to the origin of European agriculture (e.g. barley and wheat) than in land farmed with more recently introduced crops (e.g. maize and rapeseed), where recently introduced weeds were more numerous. Thus, the success in modern times of A. thaliana and other weedy plants brought from Europe to temperate regions worldwide may be the result of long-lasting associations with European agriculture that these plants have had since the time of the Neolithic revolution.
While our results might be explained by the simultaneous expansion of A. thaliana into Europe from multiple glacial refugia, we find that a perspective incorporating agriculture explains the data as parsimoniously as a model relying exclusively on natural dispersal. Because the sampling of accessions was denser at intermediate latitudes than it was in southern Europe, we were not able to exclude roles for Spanish or Italian refugia or for a Mediterranean route of agriculture in producing the pattern of variation in current genomes. One possibility is that A. thaliana did follow the agricultural expansion, but only after it had already arrived in Europe via a natural colonization from glacial refugia. Similarly to the diffusion of human agriculturalist genes, the continuous pattern of variation in A. thaliana would then be explained by the genetic dilution of the eastern genes that might have resulted from admixture with local populations during the agricultural expansion phase. Although the current data set has a large representation of individuals along the Danubian route of agricultural expansion, genomic analysis of a larger sample from Spain and the Balkans, as well as from the key eastern region of Asia Minor, will have greater potential to distinguish among possible models for the evolutionary history of A. thaliana in Europe.
A set of 76 individuals containing both hierarchical population samples and stock center accessions was extracted from the sample of 96 individuals studied by Nordborg et al. [21]. The subset included all accessions within an interval of latitudes of (32°N, 65°N) and within an interval of longitudes of (−10°E, 40°E), i.e. all European accessions plus one from Libya (Mt-0). For the 76 individuals, the total set of 876 reliable alignments representing 0.48 Mb of the genome was used. A thorough description of the data set can be found in the Materials and Methods of [21]. The list of accessions used in this study can be found in Table S1.
Since Arabidopsis thaliana is largely homozygous, we used a haploid setting. To enable comparisons with results obtained in [21] from the program STRUCTURE version 2.0, each fragment was treated as a multiallelic locus, so that two accessions had a different allele if they differed at any site in the fragment. To determine which clusters are generally robust to the assumption of continuous variation, we used a modified algorithm that includes spatially explicit prior distributions describing which sets of individuals are likely to have similar cluster membership [25]. In this approach, implemented in the program TESS [26], clusters correspond to spatially and genetically continuous units separated by small discontinuities that occur where genetic barriers are crossed. The incorporation of a spatial component into the clustering model has the potential to determine if clines provide a sensible description of the underlying pattern of variation.
We performed an admixture analysis using TESS version 1.1, whose individual-based spatially explicit Bayesian clustering algorithm uses a hidden Markov random field model to compute the proportion of individual genomes originating in K populations [25],[26]. The hidden Markov random field accounts for spatial connectivities by representing them as links in a network of individuals. In addition, the hidden Markov random field also incorporates decay of membership coefficient correlation with distance (computed on the network), a property similar to isolation-by-distance. The network topology merely conveys information about which pairs of individual genomes are more likely to be assigned to the same clusters, and the network was automatically generated by the TESS program using a Dirichlet tessellation obtained from the accession spatial coordinates. To better account for potential geographic barriers, we modified the network by removing several links. For our application to A. thaliana, we imposed a network topology in which the skeleton of the topographic structure of European landmasses was mimicked (Figure S1). This topology was obtained after removing the longest Dirichlet edges in the automatically generated graph.
Two values of the TESS interaction parameter were used, ψ = 0.6 and ψ = 1, which can be viewed as a moderate and a strong value. This hyperprior parameter weights the relative importance given to spatial connectivities (the value ψ = 0 recovers the model underlying STRUCTURE). Similar results were obtained from both the moderate and strong values, and only those for ψ = 0.6 are reported.
TESS and STRUCTURE proceed with the determination of the number of clusters K in a similar way. However the TESS algorithm incorporates a regularization procedure that perhaps leads to a less ambiguous decision regarding K. Indeed K can be determined by sequentially increasing the maximal number of clusters, Kmax, and by running the program until the final inferred number of clusters, K, becomes less than Kmax. We used the admixture version of TESS, and we set the admixture parameter to α = 1. The algorithm was run with a burn-in period of length 20,000 cycles, and estimation was performed using 30,000 additional cycles. We increased the maximal number of clusters from Kmax = 3 to Kmax = 8 (20 replicates for each value). Runs with Kmax = 5 led to either K = 3 or to K = 4.
For each run we computed the Deviance Information Criterion (DIC) [59], a model-complexity penalized measure of how well the model fits the data. The smallest DIC values were obtained for Kmax = 5. One accession, Mr-0 (Italy), shared nearly equal membership in each of the Kmax clusters, regardless of the value of Kmax (see the clustering tree in [60] for identification of Mr-0 as an outgroup accession). To a lesser extent, Bur-0 (Ireland) and Fei-0 (Portugal) exhibited similar patterns of shared membership. For Kmax = 5, we performed 100 additional runs (interaction parameter ψ = 0.6, admixture parameter α = 1), and we averaged the estimated admixture coefficients (Q matrix) over the ten runs with the smallest values of the DIC (DIC ∼72,000, s.d. = 30). To account for label switching and to decide which of the clusters of each run corresponded to a specific label, we used the software CLUMPP version 1.1 [61], whose greedy algorithm computed a symmetric similarity coefficient equal to 0.788 (100 random input sequences, G statistic).
Spatial interpolation of admixture coefficients was performed according to the kriging method as implemented in the R packages ‘spatial’ and ‘fields’ [62],[63]. One difficulty with fitting trend surfaces arises when the observations are not regularly spaced. To handle this issue we took the spatial correlation of the fitting errors into account by assuming that the errors had non-null covariance. Trend surfaces of degree two were adjusted using generalized least squares and exponential covariance with decay parameter h = 5.
The regression analysis of heterozygosities on geographic distances was based on 57 central European, eastern and western European accessions. The 57 individuals were grouped into seven samples as described in Table S2. The seven samples were defined on the basis of geographic and genetic proximity, and provided a balance between pooled individual accessions and actual population samples. We did not include nine individuals that were either ambiguously assigned to clusters by TESS or that were geographically isolated. The German sample was restricted to six accessions, and diversity for this sample was estimated by using a resampling procedure (mean over 100 replicates). We also ran a simulation study to evaluate the influence of the resampling strategy (Text S1).
We used an ABC approach for inferring demographic parameters under four models of population growth. In the ABC approach, we assume that there is a multidimensional parameter of interest θ, and the observed value sobs of a set of summary statistics, S, is calculated for the data. The basic rejection sampling method generates random draws (θi, si), where θi is sampled from the prior distribution, and si is measured from synthetic data, simulated from a generative model with parameter θi. Fixing the tolerance error, δ, only parameters θi such that |sobs−si|<δ are retained to form an approximate sample of size M from the posterior distribution, where | . | is the Euclidean norm. We used tolerance errors such that fractions of either 5% or 1% of the total number of simulations were retained.
The four demographic scenarios were described in the text as Models A–D. The six-dimensional parameter θ included the mutation rate per bp per generation, μ (×10−8), the population size at the onset of expansion, N1, the time since the onset of expansion, t0, the growth rate, r, the present equilibrium population size, N0, and the time elapsed since the equilibrium phase, t1. The variable mutation rate models included locus-specic rates, μj, obtained as independent realizations of an exponential prior distribution for which the hyperparameter was exponentially distributed with mean μ. Coalescent simulations were performed with the software MS [64]. Recombination within each locus was assumed, using an exponentially distributed prior of mean 0.3 for the effective recombination rate [65]. The prior distributions used in the four models are described in Table S3.
Twelve summary statistics were used to capture genomic information at the 648 loci, defined as the 25%, 50% and 75% quantiles (quartiles) of each of the distributions of the number of segregating sites, the mean number of pairwise differences between sequences, the Tajima D statistic, and the number of distinct haplotypes. The summary statistics were rescaled before comparison to the observed statistics. We divided each simulated summary statistic by the median absolute deviation – a robust estimate of the standard deviation – of the simulated statistics. Our ABC approach partially followed Beaumont et al. [29], who added regression adjustment and smooth weighting to the Bayesian rejection algorithm of Pritchard et al. [30]. We dropped the regression adjustment step because it led to a poor fit during preliminary runs (R2<0.25). The second improvement of the original method – namely, smooth weighting – was retained in our analysis. Smoothing was implemented using the Epanechnikov kernel Kδ with window size δ to weight the parameters by Kδ (|si−sobs|) [29]. The same weights were also used when estimating the mean, the quartiles and the maximum of the posterior distribution.
We computed the Bayes factor when evaluating the evidence of model 1 against model 2 (where 1 and 2 are chosen among A, B, C and D) as described in Results. The new formula can be seen as an improvement of the method that used the ratio of acceptances under the two models to approximate the Bayes factor, originally formulated aswhere Iδ is the indicator function Iδ (t) = 1 if t<δ, 0 otherwise. Note that in our case, Jeffreys' scale on degrees of belief should be interpreted more cautiously than the usual scale based on exact Bayesian computation [33]. The Bayes factors in Figure 3 and Table S5 correspond to the ratio of the weight of evidence of each model to the weight of evidence of the variant of model B with variable mutation rates. Two tolerance errors, δ0.01 and δ0.05, corresponding to the 1% and 5% quantiles of the distance between the summary statistics obtained under the variant of model B with variable mutation rates and the observed summary statistics, were used when computing the Bayes factors.
We selected 64 individuals from central Europe and western Europe and ten individuals from northern Europe (northern Sweden and Finland). From the 876 fragments, we removed indels, sites with more than 20% missing data, and monomorphic sites. A total of 795 fragments and 11,134 SNPs remained. For each site, the remaining missing data was replaced by sampling alleles from the allele frequency distribution so that the final data set did not contain any missing data.
We simulated data from model C using MS [64]. Forward in time, there is a period of constant population size followed by a period of growth and finally a period of constant population size ending in the present. We used the model parameters from the MAP estimates of model C (see Table 1), which received the most statistical support from the ABC analysis. We considered variable mutation rates per simulated fragment, taken from the same exponential distribution as used in the ABC analysis (also in agreement with [66]). The recombination rate in a simulated fragment was set to 0.3 [65]. We assumed that the population split into two subpopulations some time T in the past, scaled by NCE = 135,000, the estimated size of the central European population, and that migration occurred at rate m, scaled by NCE. The size of the northern population, NNE , was assumed to be 1/4 of the estimated size of the central European population, NCE. The growth scenario was assumed to be the same in the two populations, with only the population sizes differing.
To approximate the likelihood of the parameters, we used two haplotype diversity statistics, the mean number of distinct haplotypes and the mean number of private haplotypes. To correct the number of distinct haplotypes and the number of private haplotypes statistics for sample size differences, we used the rarefaction method [35],[36] to get estimates of these statistics for samples of size two to ten (the sample size of the northern Swedish and Finnish population was equal to ten). This was done separately for each fragment, and finally we took the average across fragments.
Simulations of a two-dimensional stepping stone model were performed using the program SPLATCHE 1.1 [40]. We modeled Europe using an array of demes that included topographic information borrowed from the online Geographic Information System GEODAS of the National Geographic Data Center. The map covered latitudes from 32°N to 65°N and longitudes in an interval of −10°E to 40°E. Topography was used to define carrying capacities for each deme. We divided Europe into an array consisting of 130×180 = 23,400 demes, each representing an area of ∼600 km2. To account for the fact that A. thaliana inhabits lower altitude landscapes, carrying capacities were set to their highest values for altitudes below 200 m (N = 5,000). They were progressively decreased to N = 100 for altitudes higher than 1500 m using a nonincreasing step function (N = 2,500 for altitudes between 200 m and 500 m, N = 1,000 for altitudes between 500 m and 1000 m, N = 500 for altitudes between 1000 m and 1500 m). At the beginning of the colonization process, a single deme was occupied. To date the onset of the spread, we based the origin at the north of the Black Sea (48°N, 35°E). We chose a logistic population growth model to describe the dynamics of population demography within each deme. The growth rate r was identical in each deme. Following [66] we set the mutation rate per base pair and per generation around u ∼10−8, and the generation time corresponded to one year. Because the memory requirements of SPLATCHE are particularly high, we modified the mutation rate and the effective size in order to accelerate the generation time from one year to ten years (this means that the model was simulated ten generations at a time).
Values of the original population size were taken equal to N1 around 10,000 (5,000–15,000). DNA sequences were simulated using the modified mutation rate v = 10−5. Rescaling the generation time to a value tR = 10 years produced a level of nucleotide diversity close to the one present in the data (Ne u = N1×v/tR ∼10−2). Note that N1 cannot be compared to the value used in the non-spatial ABC simulations unless we restore the original mutation rates and generation times. After the correction, the values used in the spatial and non-spatial scenarios were actually similar. In simulations, we assumed that the population remained constant (equal to N1) during 100 Ky before range expansion.
To compare with the data in western and central Europe, we simulated the genealogies of 66 individuals located at the same spatial coordinates as the set of 66 accessions that excluded those from northern Sweden and Finland. The fit of simulated data to the real data was assessed by evaluating the distance between the empirical folded frequency spectrum computed from the non-coding sequences, and frequency spectra obtained from individuals simulated at the same locations.
The distance used to compare folded spectra was the χ2 distance defined from four classes as follows: Class 1) minor allele frequency 1 (total 28%); Class 2) minor allele frequency 2–4 (total 26%); Class 3) minor allele frequency 5–12 (total 25%); and Class 4) minor allele frequency 13–33 (total 21%).
Five model parameters were varied: the time of the onset of spatial expansion t0, the migration rate m, the growth rate within a newly colonized deme r, the effective population size at the beginning of range expansion N1 (resized), and the location of the origin. Ideally one would use an ABC analysis to choose a subset of parameters that maximizes the posterior probability of the corresponding evolutionary scenario given prior distributions over these parameters. However performing an ABC analysis with geographically explicit simulations is prohibitively time-consuming, due to the large cost of a single simulation. In practice, we first performed a coarse search using fixed values of the starting date t0 (equal to 8,000–12,000 BP) and a random sampling design for the other parameters, exploring migration rates (m) within the range 0.1–0.8 and population size expansion rates (r) within the range 0.2–1.5, and assuming that the starting point was located at coordinates (48°N, 35°E). This preliminary search found that values of migration rates around 0.2–0.3, growth rates between 0.6 and 1, and initial sizes of 5,000–10,000 individuals yielded non-significant χ2 P-values. These ranges of parameter settings for m and r corresponded to the saturation of a deme in 100–300 years. For most of the simulations, the length of the colonization phase was around 3,000–6,000 years, which corresponded to waves of advance varying from 0.5 to 1 km/year. In a second stage, we investigated the time at which the range expansion began, varying this time from t0 = 5,000 BP to t0 = 20,000 BP using r = 0.6 for the oldest dates. For the most recent dates, we increased r to 0.7 (t0 = 10,000), 0.9 (t0 = 7,000) and 1.2 (t0 = 5,000), so that the colonization phase ended before the present day. Finally, we studied the explanatory power of twenty-four potential spatial origins throughout central and western Europe (m = 0.25, r = 0.6, Figure 8B). |
10.1371/journal.pntd.0006317 | Dietary diversity and poverty as risk factors for leprosy in Indonesia: A case-control study | Poverty has long been considered a risk factor for leprosy and is related to nutritional deficiencies. In this study, we aim to investigate the association between poverty-related diet and nutrition with leprosy.
In rural leprosy-endemic areas in Indonesia, we conducted a household-based case-control study using two controls for each case patient (100 recently diagnosed leprosy patients and 200 controls), matched for age and gender. All participants were interviewed to collect information on their demographics, socioeconomic situation, health, and diet. Body mass index, dietary diversity score, as well as anemia and iron micronutrient profiles were also obtained. By means of univariate, block-wise multivariate, and integrated logistic regression analyses, we calculated odds ratios between the variables and the occurrence of leprosy. Unstable income (odds ratio [OR], 5.67; 95% confidence interval [CI], 2.54–12.64; p = 0.000), anemia (OR, 4.01; 95% CI, 2.10–7.64; p = 0.000), and higher household food insecurity (OR, 1.13; 95% CI, 1.06–1.21; p = 0.000) are significantly associated with an increased risk of having leprosy. Meanwhile, higher education (OR, 0.34; 95% CI, 0.15–0.77; p = 0.009) and land ownership (OR, 0.39; 95% CI, 0.18–0.86; p = 0.019) have significant protective associations against leprosy. Although lower dietary diversity, lack of food stock, food shortage, low serum iron, and high ferritin were found more commonly in those with leprosy, the occurrence of leprosy was not significantly associated with iron deficiency (OR, 1.06; 95% CI, 0.10–11.37; p = 0.963).
Food poverty is an important risk factor for leprosy susceptibility, yet the mechanisms underlying this association other than nutrient deficiencies still need to be identified. With a stable incidence rate of leprosy despite the implementation of chemoprophylaxis and multidrug therapy, improving dietary diversity through food-based approaches should be initiated and directed toward high-prevalence villages. The possible underlying factors that link poverty to leprosy other than nutrient deficiencies also need to be identified.
| Despite the suggestion that nutritional deficiencies may impair host immune responses against Mycobacterium leprae, there has not been any systematic study on how various aspects of poverty interact and associate with nutrition and leprosy. In poor rural areas in Indonesia that have the highest proportion of multibacillary cases, we aimed to investigate these associations by interviewing recently leprosy diagnosed patients and measuring their anemia and iron profiles. Our findings suggested that, compared to the control population, people who are at an increased risk of contracting leprosy have lower education, lack of stable income to provide diverse types of food, and are anemic. Although low serum iron and high ferritin levels were found more commonly in those with leprosy, we did not find a significant association between iron deficiency and leprosy. Our study clarifies that food poverty is an important risk factor for leprosy susceptibility, yet the mechanisms underlying the association between diet and leprosy other than nutrient deficiencies still need to be identified. Improving dietary diversity through food-based approaches should be initiated and directed towards high-prevalence villages.
| Leprosy has long been known as a disease of poverty, yet the mechanism underlying this interaction remains unclear. Most of the affected countries are underdeveloped, in which people affected by leprosy are born and raised in poor environments and continue being pushed into poverty due to the stigma and disabilities [1]. Poverty means more than just a lack of income; it also encompasses the multiplicity of non-monetary aspects that often combine and intensify the negative effects of being poor, including lack of proper food and nutrient intakes [2]. Correspondingly, food shortage, food insecurity, and lower dietary diversity are several aspects of poverty that are more commonly found in those struggling with leprosy [3]. Previous studies have shown positive associations between food shortage and food insecurity with the occurrence of leprosy, and it was suggested that impaired host immune response against the causative bacteria as a result of insufficient nutritional intake is the possible cause of this condition [4]. However, there has been no systematic study on how various aspects of poverty interact and associate with leprosy to support the suggestion, particularly in Indonesia, which is currently the home of more than 17,000 new leprosy cases registered annually and has the highest proportion of multibacillary (MB) cases [5]. The purpose of this research, which is a part of the MicroLep Study, is to elucidate the association between poverty-related dietary intake and leprosy by determining the interaction between demographic, socioeconomic, and diet-related factors of poverty on several nutrition indicators, which encompasses people with leprosy and healthy controls in the Indonesian population.
This study was approved by the ethical review committee of the Faculty of Medicine, Universitas Indonesia, Jakarta, Indonesia (reference number: 595/UN2.F1/ETIK/2016). The Agency for National and Political Unity of Bangkalan and the District Health Office of Bangkalan, Madura, East Java Province, Indonesia have also been informed about this study and have given their approval and support prior to the beginning of the study. A signed informed consent form was obtained from each participant before starting the study.
We conducted a household-based, case-control study in rural areas of Bangkalan, Madura, East Java, Indonesia, from November to December 2016. Bangkalan has 22.38% inhabitants who are living below the poverty line, making it the second poorest region in Madura after its neighboring district, Sampang [6]. Correspondingly, this area is also endemic for leprosy; 310 new cases were diagnosed in 2015 in a total population of 1 million, yet no chemoprophylaxis therapy has been given to prevent leprosy in patient contacts [7,8].
Data on people with leprosy were gathered from the Leprosy Cohort Data Report of the Bangkalan Municipality of Health. Seventeen of 22 primary health care facilities in Bangkalan participated in this study. Selected cases between the ages of 18 and 65 who were being actively treated with the World Health Organization (WHO)-recommended multidrug therapy (MDT) regimen were chosen based on the current cohort report up to September 2016. Control subjects who lived in the village or neighborhood with common characteristics as the cases with the same sex and age range were also selected, with a ratio of 1:2 between the cases and controls. The following exclusion criteria were applied: refusal to participate, limited understanding of information, pregnant or breastfeeding, or had taken systemic antibiotics other than the WHO-MDT regimen within 30 days preceding inclusion. Additionally, control subjects who had household members with a history or newly diagnosed leprosy at the time of inclusion were also excluded.
Six trained surveyors and six trained health workers who spoke Bahasa Indonesia and Madurese collected the data using a structured questionnaire along with peripheral blood sample collection during household visits.
The questionnaire used in this research was adapted from a study that was conducted in Bangladesh [9]. The original English version of the questionnaire had been translated into Bahasa Indonesia, optimized, pre-tested, and validated prior to the study. The first section of the questionnaire focused on the demographic, socioeconomic, and health characteristics of the subjects and their households. Household size (the number of people eating in the house), occupation of the income generator and subject, land ownership, as well as the subject’s level of education, average income, income variation, food expense, and self-classification on a poverty scale were registered. The triggering cause of income variation was also included, but the difference between pre- and post-leprosy diagnosis was not asked due to the tremendous stigma in Bangkalan. As for the health characteristics, a number of questions were asked about the details of any acute and chronic diseases in the prior year, the presence of a BCG scar, history of medication, and leprosy diagnosis (for the case group). Afterward, the household food insecurity access scale (HFIAS) was administered to specify the problem concerning food insecurity during the preceding four weeks [10]. Food storage and dietary modification such as lessening the number or variety of meals was also asked in detail. For comparability purposes, food shortage was defined with the same criteria as in the Bangladesh study [1]. Finally, dietary intakes consisting of three meals a day and snacks in between were assessed by 24-hour recall, from which the individual dietary diversity score (IDDS) was calculated. The subjects described their 24-hour food consumption history in chronological order, starting from breakfast the previous day. The details of ingredients for each meal and snack were obtained, particularly for mixed dishes and processed foods. However, the food quantity was not obtained as the 24-hour recall focused on the quality of the diet composition. Subsequently, the food ingredients were categorized into nine categories and were calculated based on the Food and Nutrition Technical Assistance Project/Food and Agriculture Organization of the United Nations (FANTA/FAO) guidelines [11]. “Milk and milk products” were defined as all dairy-based products with the exception of butter, and the slight amount of milk in coffee was not counted. Moreover, garlics, shallots, and chili spices were classified as condiments due to the small amounts consumed. Considering that special feasts are usually prepared for special celebrations or specific religious holidays in Indonesia, 24-hour recall was not carried out during those particular days. Thus, we could assume that the variance among food ingredients remained stable over the period.
Following the interview, weight was assessed using a portable scale (GEA Medical, Jakarta, Indonesia) and height was measured using a measuring tape; the subjects were asked to remove their footwear and stand on a flat surface with their back against the wall. Peripheral blood from both groups was collected into EDTA and SST vacutainers (BD, Franklin Lakes, NJ, USA) by trained health workers and distributed to a laboratory in Surabaya, where blood tests were performed to measure hemoglobin and iron micronutrient profiles.
A MicroLep Study database, designed in Microsoft Excel, was established and the data were entered by well-trained data-entry personnel. Demographic, socioeconomic, and health characteristics were determined with descriptive analyses. Subsequently, four blocks consisting of several related variables were built into a framework (Fig 1). Univariate, block-wise multivariate, and integrated analyses were performed using logistic regression in SPSS version 21 with case or control as the dependent variable. Sex and age were also adjusted in order to control for confounding effects from the pair matching design [12]. Univariate and multivariate analyses within the blocks were performed first, and the variables that were relevant and significantly associated with leprosy from each block (p<0.05) were included in the integrated analysis.
A total of 276 of 419 cases were eligible for the study, of which 103 patients were randomly selected using Research Randomizer. However, only 100 cases were included in the analysis due to uncomplete data. Among 218 house visits, 200 controls were able to complete the questionnaire and were included in this study (response rate: 91.74%). In total, 300 subjects consisting of 100 cases and 200 controls were included in the analysis, with a mean age of around 35 to 36 years and approximately equal numbers of males and females. The majority of cases were MB (89%), of which 15% and 6% of patients presented with grade 1 and 2 disabilities, respectively. The demographic, socioeconomic, and health characteristic of the subjects are shown in Table 1.
Detailed information about HFIAS, food shortage, and IDDS are provided in Table 2 and Fig 2.
The HFIAS score was two times higher in people with leprosy compared to the controls (p = 0.000). In addition, food storage availability was 18% higher in the controls, which lasted on average of 6.5 weeks compared to 5 weeks in people with leprosy.
A paired t-test showed that the mean of the BMI between the cases and controls was significantly different (p = 0.002). Moreover, around 42% of the people with leprosy had anemia [13], which was 29% higher than the controls. Reflecting micronutrient deficiency, the blood iron profile showed that more people with leprosy had low iron serum levels than the controls (23% and 9%, respectively). Total iron-binding capacity (TIBC) and transferrin saturation were also lower (Table 3), while high ferritin levels were twice as common in those with leprosy than in the controls (37% and 16%, respectively).
The results of univariate and multivariate analyses per block are shown in Table 4. First, education dominated the demographic factors block (p<0.05); higher education was associated with a lower risk of leprosy.
Second, unstable income and land ownership played an important role (p<0.05) in the socioeconomic factors block; people with these factors had a greater risk of developing leprosy. In addition, both log income per capita and log food expense had a protective association against leprosy, yet the numbers were almost similar in both groups and therefore did not show significant associations with leprosy.
Third, all variables in the diet-related factors block were significantly associated with leprosy in the univariate analysis. High HFIAS and experiencing food shortage at any time in life increased the risk of having leprosy, while IDDS and food stock availability had a reverse association with leprosy. Nevertheless, only HFIAS remained significant in the multivariate analysis (p<0.05).
Fourth, all variables in the nutrition indicators block other than BMI also showed significant associations with leprosy in the univariate analysis. In this block, in addition to analyzing the original variables, we also considered the interactions between iron-TIBC and iron-ferritin [14–16], which were also statistically significant in the univariate analysis. However, only hemoglobin remained significant (p<0.05) in the multivariate analysis.
Following per block multivariate analyses, all of the significant and relevant variables were included in an integrated analysis. This final analysis aimed to reveal the connection among variables from each block. The results are presented in Table 5. Based on this analysis, variables of education, unstable income, HFIAS, and hemoglobin remained significantly associated with leprosy (p<0.05).
Our results showed that people with leprosy have less favorable socioeconomic and demographic conditions, as well as dietary consumption. Low education levels, unstable incomes, and no land ownership are some aspects of poverty that were associated with the risk of having leprosy. Moreover, although the iron profiles were not significantly associated with leprosy, the low nutritional status in people with leprosy was associated with lower IDDS and higher HFIAS.
Based on our analysis, education level had a protective association against leprosy. Thus, the more educated someone is, the lower chance they will contract leprosy. In this sense, education is regarded as a substantial factor of subjects’ self-awareness that contributes to disease elimination. This is consistent with a previous study in Brazil, where the patients were unlikely to report their symptoms to receive treatment or did not even know that they had leprosy due to lack of knowledge and awareness of the disease [17]. Additionally, a higher education level is also often associated with better economic outcomes.
Other important factors associated with leprosy were unstable income and land ownership, which are related to income inequality. Based on our analysis, people with unstable incomes were five times more likely to develop leprosy, while owning private land decreased the risk of getting leprosy by 60% (OR = 0.39 [CI 0.18–0.86], p = 0.019). While those with assets are able to provide better and more stable socioeconomic outcomes [1], freelance workers such as farmers and labors have only seasonal incomes from seasonal jobs [18,19].
In contrast to the Bangladesh study [9], our research did not find a difference in food expenditures between those with leprosy and the controls. Limited food preference, culture, and food availability in the study areas might have contributed to this value. However, heterogeneities in food consumption may vary across households even with the same food expenditures and can still influence the subjects’ nutritional intake [20]. This is consistent with our result on IDDS, which was significantly lower in the case group, who ate fewer fruits and vegetables, eggs, and legumes, nuts, and seed products (Fig 2b). Although it was not statistically significant, IDDS had a reverse association with leprosy (OR = 0.85 [CI 0.67–1.09], p = 0.213). In contrast, a higher HFIAS score was significantly associated with a higher chance of contracting leprosy (OR = 1.13 [CI 1.06–1.21], p = 0.000). Nevertheless, our score was lower than in the Bangladesh study [9]. Lower gross domestic product (GDP) per capita at purchasing power parity (PPP) in Bangladesh ($3,581) than in Indonesia ($11,612) may explain these findings [21].
In terms of food shortage, our study showed similar results with those of Feenstra et al [1] (Table 4). In the univariate analysis, around 53% of cases also experienced food shortage at some time in their lives (with a mean length of 42.84±70.49 weeks) that was significantly associated with leprosy. Although this was not statistically significant in the integrated analysis, in theory, a prolonged food shortage may result in a deficiency of essential nutrients that are needed to boost an adequate immune response against infectious agents [22], thus increasing the risk of contracting infectious diseases.
Based on our integrated analysis, those with anemia are at an increased risk of contracting leprosy (OR = 4.01 [CI 2.10–7.64], p = 0.000). There are several underlying conditions related to anemia, such as micronutrient deficiencies [23,24], infectious diseases [25], and hereditary conditions (thalassemia) [16]. Iron deficiency characterized by high TIBC or low ferritin is the most common cause of anemia. However, diagnosing iron deficiency anemia (IDA) in particular areas where infectious diseases are prevalent can be very challenging as serum ferritin levels may increase due to immune responses to the infectious agent, masking a pure iron deficiency diagnosis [16]. The iron profiles in our study population were consistent with anemia that is caused either by chronic diseases (ACD) or mixed IDA and ACD. Dietary intake has been reported to influence either hemoglobin or iron levels [26], and from the interview, we knew that those with leprosy consumed much less red meat and eggs, which are rich in iron. Hence, iron deficiency from a less diverse diet mixed with chronic infection by M. leprae might be the cause of anemia in this study. Our additional multivariable analyses demonstrated that lower dietary diversity and higher HFIAS scores escalated the risk of anemia (OR = 0.86 [CI 0.67–1.10], p = 0.227) and OR = 1.09 [CI 1.03–1.16], p = 0.003, respectively) and that dietary diversity had a reverse association with TIBC levels, which is a sensitive indicator of iron deficiency (OR = 1.37 [CI 0.60–3.11], p = 0.454). However, our final findings do not support the suggestion that iron micronutrient deficiency due to insufficient nutritional intake increases the susceptibility to leprosy. More studies are needed to identify other possible mechanisms underlying the association between poverty-related diet and leprosy. For instance, if diet-related risk factors for leprosy result from altering the skin or gut microbiota composition. Further research is currently being conducted to elucidate the role of diet-microbiota interaction in leprosy.
Although this study was carefully prepared and conducted, there were some unavoidable limitations. First, the data were collected after the diagnosis of leprosy and due to the strong stigma in the research areas, we were not allowed to specifically ask for any changes in the subjects’ income and diet after diagnosis, which made it hard to determine a causal relationship. Furthermore, the data regarding the subjects’ dietary intake were collected using a cross-sectional design. Ideally, a longitudinal study on diet and health should be conducted to compare data between those who eventually develop leprosy and those who do not. However, leprosy is a slowly developing infectious disease with a very long incubation period, so it is still difficult to determine a causal relationship using a short-term longitudinal study. In order to correct this difference, we asked all subjects for any changes in their economic status and dietary intake in general. All of the subjects had anonymously answered that they had been experiencing the same conditions and mostly consuming the same diet in the prior years. Only two subjects indicated a variation in their income due to their health, but not specifically for leprosy.
Second, the subjects were asked to reveal their dietary intake and food shortage history in the past 24 hours, past year, and in longer periods, introducing recall and response biases. In order to limit the effects on our results, the same questions were asked several times and atypical days such as parties or religious holidays were avoided. Third, in order to assess dietary intake more objectively, biomarkers for other micro- and macronutrients in blood should also be analyzed. However, we analyzed only the iron profiles considering the cost, positive correlations with dietary intake, their essential role during infection, and the limited research on the role of this micronutrient in leprosy.
In conclusion, our findings suggest that food poverty is an important risk factor for leprosy susceptibility, yet the mechanisms underlying this association other than nutrient deficiencies still need to be identified. With a relatively stable incidence rate of leprosy despite the implementation of chemoprophylaxis and multidrug therapy, improving dietary diversity through food-based approaches should be initiated and directed toward high-prevalence villages. The possible underlying factors that link poverty to leprosy other than nutrient deficiencies also need to be identified.
|
10.1371/journal.pbio.1000624 | Zyxin Links Fat Signaling to the Hippo Pathway | The Hippo signaling pathway has a conserved role in growth control and is of fundamental importance during both normal development and oncogenesis. Despite rapid progress in recent years, key steps in the pathway remain poorly understood, in part due to the incomplete identification of components. Through a genetic screen, we identified the Drosophila Zyxin family gene, Zyx102 (Zyx), as a component of the Hippo pathway. Zyx positively regulates the Hippo pathway transcriptional co-activator Yorkie, as its loss reduces Yorkie activity and organ growth. Through epistasis tests, we position the requirement for Zyx within the Fat branch of Hippo signaling, downstream of Fat and Dco, and upstream of the Yorkie kinase Warts, and we find that Zyx is required for the influence of Fat on Warts protein levels. Zyx localizes to the sub-apical membrane, with distinctive peaks of accumulation at intercellular vertices. This partially overlaps the membrane localization of the myosin Dachs, which has similar effects on Fat-Hippo signaling. Co-immunoprecipitation experiments show that Zyx can bind to Dachs and that Dachs stimulates binding of Zyx to Warts. We also extend characterization of the Ajuba LIM protein Jub and determine that although Jub and Zyx share C-terminal LIM domains, they regulate Hippo signaling in distinct ways. Our results identify a role for Zyx in the Hippo pathway and suggest a mechanism for the role of Dachs: because Fat regulates the localization of Dachs to the membrane, where it can overlap with Zyx, we propose that the regulated localization of Dachs influences downstream signaling by modulating Zyx-Warts binding. Mammalian Zyxin proteins have been implicated in linking effects of mechanical strain to cell behavior. Our identification of Zyx as a regulator of Hippo signaling thus also raises the possibility that mechanical strain could be linked to the regulation of gene expression and growth through Hippo signaling.
| Processes that control cell numbers are essential during normal development, when they are required to generate organs of the correct size, and during cancinogenesis, when they influence tumor growth. The Hippo pathway is an intercellular signaling pathway that relays information about cell-cell contact and cell polarity to a signal transduction pathway that regulates the transcription of genes controlling cell numbers. The role of Hippo signaling in controlling growth is conserved from fruit flies to humans, but many aspects of the Hippo signal transduction pathway remain poorly understood. In this article, we identify Zyx as a previously unknown component of the Hippo pathway in Drosophila, and characterize its role within the pathway. We show that Zyx plays an essential role in a branch of Hippo signaling that involves the transmembrane receptor protein Fat and its target Dachs, which is a myosin family protein. Our results suggest a model in which Fat regulates the localization of Dachs, Dachs subsequently binds Zyx, stimulating its binding with the kinase Warts/Lats, and thereby regulates downstream signaling events. Zyx is conserved in vertebrates and we suggest that vertebrate Zyx proteins might also be involved in the regulation of Hippo signaling and, thereby, organ growth.
| The Hippo pathway has emerged as an important regulator of growth during metazoan development, and its dysregulation is implicated in diverse cancers [1]–[3]. Hippo signaling is effected by transcriptional co-activator proteins, Yorkie (Yki) in Drosophila and YAP and TAZ in mammals [4]. Three interconnected, upstream branches of Hippo signaling have been characterized in Drosophila: Fat-dependent, Expanded-dependent, and Merlin-dependent [1]–[3]. These upstream branches converge on the kinase Warts (Wts), which can phosphorylate Yki. Phosphorylated Yki is retained in the cytoplasm, whereas unphosphorylated Yki can enter the nucleus and, in conjunction with DNA-binding partners, promote the transcription of downstream genes. Upstream branches of Hippo signaling regulate both the activity of Wts and its abundance. Our understanding of many steps in Hippo signaling remains fragmentary, in part due to incomplete identification of pathway components. Here, we describe the identification of Zyx102 (Zyx, FBgn0011642) as a novel component of Hippo signaling and characterize its role in the pathway.
Fat is large cadherin that acts as a transmembrane receptor for one branch of Hippo signaling [1]–[3],[5]. Fat-Hippo signaling influences the levels of Wts protein [6]. The molecular mechanism by which this is achieved is not understood, but dachs is genetically required for the influence of Fat on Wts levels, downstream gene expression, and organ growth [6]–[8]. Fat regulates the localization of Dachs to the sub-apical membrane: when fat is mutant, Dachs accumulates on the membrane around the entire circumference of the cell, and when Fat is over-expressed, Dachs is mostly cytoplasmic [7]. In imaginal discs and optic neuroepithelia, Dachs membrane localization is polarized within the plane of the tissue; this polarization reflects the graded expression of the Fat ligand Dachsous and the Fat pathway modulator Four-jointed [7],[9],[10]. The correlation of Dachs localization with Fat activity implicates Dachs regulation as a key step in Fat signaling, but how Dachs localization influences downstream events is unknown.
Zyx is a Drosophila homologue of the vertebrate Zyxin, Lipoma preferred partner (LPP), and Thyroid-receptor interacting protein 6 (TRIP6) proteins [11],[12]. These proteins have three conserved LIM domains at their C-terminus, and they have been implicated in both cytoskeletal and transcriptional regulation [13]–[15]. Gene-targeted mutations in murine Zyxin or Lpp have no significant effect on mouse development, presumably due to redundancy among family members [16],[17]. Translocations involving LPP identified it as an oncogene involved in lipomas and other cancers [13]. In cultured cell assays, Zyxin and its paralogues can affect cell motility and actin polymerization and can localize to focal adhesions and adherens junctions [13],[15],[18]. Notably, Zyxin has been implicated as playing a key role in mechanotransduction, as its localization to focal adhesions can be influenced by the application of mechanical tension to cells in culture [18].
We report here that Zyx is an essential component of the Fat-Hippo signaling pathway, required for normal Yki activity and growth in Drosophila. Using genetic epistasis tests, we position the requirement for Zyx in between fat and wts. Binding studies show that Zyx protein binds to Dachs and binds to Wts in a Dachs-regulated manner. Our observations suggest a model in which the regulated localization of Dachs to the membrane regulates Zyx-Wts binding, which then promotes Wts degradation. Dachs is a myosin protein, and its myosin motor domain contributes to interactions with Zyx and Wts, which raises the possibility that additional myosins might regulate Zyx-Wts interactions in other contexts.
In a screen for additional components of the Fat and Hippo pathways, we examined a collection of transgenic flies expressing UAS-hairpin constructs, which mediate RNAi. We focused on the X and 4th chromosomes, which are under-represented in traditional genetic screens, and looked for phenotypes when these RNAi lines were expressed in the notum under pnr-Gal4 control, and in the wing under vg-Gal4 control. To enhance the strength of RNAi, the screening was done in flies expressing Dicer2 from a UAS-dcr2 transgene [19]. One hundred and forty-eight lines exhibiting either altered tissue growth or lethality were then re-screened for possible effects on Fat-Hippo signaling by assaying the expression of downstream targets of the pathway, Wingless (Wg) and thread (th, more commonly referred to as Diap1) [20],[21], in wing discs in which RNAi lines were expressed in anterior cells under ci-Gal4 control (Table S1). The most promising candidates were then taken through four additional tests, involving confirmation of effects on additional downstream target genes, characterization of phenotypes when expressed under additional Gal4 drivers, confirmation of phenotypes with additional, independent UAS-RNAi lines, and characterization of genetic interactions with known pathway components. Based on these experiments, a single gene, Zyx102 (Zyx) [11],[12], which is located at 102F7 near the tip of the fourth chromosome, was identified as a novel component of the Fat-Hippo signaling pathway.
Reduction of Zyx in the developing wing disc, under nub-Gal4 control (Figure S1A), results in adult flies with small wings (Figure 1A–C,S). Similar phenotypes were observed using two different RNAi lines, although NIG-32018R3 (RNAi-Zyx32018), the line identified in our original screen, has slightly stronger phenotypes. Hippo signaling also regulates leg growth, and depletion of Zyx in developing legs results in shorter legs with fewer tarsal segments (Figure S1I,J). In addition to observing similar phenotypes with two independent RNAi lines, confirmation that the phenotypes observed result specifically from reduction of Zyx was provided by the observation that over-expression of Zyx from a UAS transgene rescued the RNAi phenotypes (Figure 1D,S). We also confirmed by Western blotting that that Zyx RNAi reduced Zyx protein levels (Figure S1K).
Many different genes and pathways affect organ growth. To investigate the potential connection between Zyx and the Hippo pathway, we examined the expression of downstream target genes in wing discs in which Zyx was depleted by RNAi. As downstream targets we employed reporters of expanded (ex) expression (ex-lacZ) and th expression (th-lacZ, Diap1). When Zyx was depleted from posterior cells using en-Gal4, ex-lacZ, th-lacZ, and Diap1were all reduced (Figures 2A,B, S2A). Hippo signaling regulates transcription by controlling the sub-cellular localization of Yki: activation of Hippo signaling promotes cytoplasmic localization of Yki, whereas inactivation of Hippo signaling allows nuclear localization of Yki, which corresponds to Yki activation [22],[23]. Zyx RNAi reduced nuclear Yki. This effect was subtle at late third instar, when levels of Yki in the nucleus are already low, but was evident in younger wing discs, which have higher levels of nuclear Yki (Figure 2C,D). The decreased expression of Hippo pathway target genes, together with the reduction in nuclear Yki, identifies Zyx as a regulator or component of the Hippo pathway. The Hippo pathway is generally thought of as a negative regulator of growth and gene expression, because most genes in the pathway act as tumor suppressors and negatively regulate the activity of Yki. Zyx, by contrast, is positively required for Yki activity and organ growth.
To position the genetic requirement for Zyx within the Hippo pathway, we performed a series of epistasis tests. RNAi lines targeted against several different tumor suppressor genes within the pathway (fat, ds, ex, wts, hpo, and mats), each of which phenocopy their respective mutants, were examined in combination with Zyx RNAi lines. The immediate upstream regulator of Yki is wts. Expression of a wts RNAi line under nub-Gal4 or en-Gal4 control is lethal at late third instar, but imaginal discs can be recovered and analyzed before lethality. Consistent with the expected de-repression of Yki, expression of wts RNAi resulted in upregulation of ex and Diap1 expression (Figure 3A). This upregulation of ex and Diap1 was not suppressed by Zyx RNAi (Figure 3B); hence, wts is epistatic to Zyx. Wts activity is directly regulated by a kinase, Hippo (Hpo), and a co-factor, Mats, and hpo and mats were also epistatic to Zyx (Figure S3A–D). These observations imply that Zyx acts upstream of Wts.
Upstream branches of Hippo signaling have been characterized in Drosophila as Fat-dependent, Ex-dependent, or Mer-dependent. In the developing wing, fat and ex make substantial contributions to Yki regulation, whereas Mer has a lesser role [6],[24]–[27]. Thus, we investigated the relationship between the requirement for Zyx and those for fat and ex. Expression of fat or ex RNAi throughout the wing, under nub-Gal4 control, results in overgrown wings (Figure 1E,I,S). Strikingly, the wing overgrowth phenotype associated with depletion of fat was suppressed by Zyx RNAi, resulting in adult wings of similar size to those of animals that only expressed Zyx RNAi (Figure 1B,F,S). This epistasis of Zyx to fat was also visible at the level of target gene expression (Figure 3D,E) and the subcellular localization of Yki (Figure 4G,H). Zyx is also epistatic to the Fat ligand ds (Figures 1S, S1C,D). These observations imply that Zyx acts downstream of fat.
The ex RNAi phenotype, by contrast, was only slightly affected by Zyx RNAi, as the wings of Zyx ex double RNAi animals remained overgrown (Figure 1J,S). Moreover, ex was epistatic to Zyx for effects on downstream target gene expression and Yki localization (Figure 4A–D,J,K). Together, these observations indicate that Zyx specifically affects Fat-Hippo signaling and has little effect on Ex-Hippo signaling.
To refine our placement of Zyx within Fat-Hippo signaling, we examined requirements for Zyx relative to additional pathway components. dco encodes a kinase that phosphorylates the Fat cytoplasmic domain and participates in Fat-Hippo signaling [6],[28],[29]. The requirement for Dco within Fat signaling is uncovered by expression of an antimorphic isoform, Dco3. Expression of Dco3 induces wing overgrowth (Figure 1L) [29]. This overgrowth is suppressed by Zyx RNAi, suggesting that Zyx acts downstream of dco (Figure 1P,S).
Like Zyx, dachs is required for normal wing and leg growth and acts genetically downstream of fat and dco but upstream of warts [6]–[8]. To examine the genetic relationship between Zyx and dachs, we took advantage of the observation that over-expression of Dachs can promote wing overgrowth (Figure 1Q) [7]. This overgrowth was completely suppressed by Zyx RNAi (Figures 1S, S1G), as was the influence of Dachs over-expression on ex-lacZ expression (Figure S4A,B). Thus, Zyx is required for Dachs-promoted activation of Yki. Over-expression of Zyx resulted in a mild wing overgrowth on its own (9% increase in wing area, Figure 1H,S), and synergized with Dachs over-expression, resulting in enhanced wing overgrowth (Figure 1R,S). Together, these observations suggest that the functions of Zyx and Dachs in regulating growth are closely linked. However, the observation that Zyx depletion could enhance the small wing phenotype of a putative null allele of dachs (Figures 1S, S1E,F) [7] implies that Zyx also has some Dachs-independent influence on growth.
Fat exerts a post-transcriptional influence on the levels of Wts protein [6]. The genetic placement of Zyx upstream of wts and within the Fat branch of the pathway suggested that Zyx might also affect Wts levels. Indeed, Zyx RNAi completely suppressed the reduction in Wts levels associated with fat RNAi (Figures 5A,B, S2B). Thus, Zyx is genetically required for the mechanism that links Fat activity to the regulation of Wts protein levels. The influence of fat on Warts levels also requires dachs [6]. Zyx RNAi did not detectably affect Dachs localization (Figure S4D,E), nor did Zyx RNAi affect Fat localization (Figure S5E,F). In addition to its effects on Wts, fat mutation also decreases the levels of Ex at the sub-apical membrane [30]–[33]. Zyx RNAi was not able to reverse this effect of fat on Ex levels (Figure S5G–N). Depletion of Zyx in the wing disc also did not have visible effects on F-actin (Figure S5O,P).
In addition to regulating transcription, Fat also regulates planar cell polarity (PCP) (reviewed in [1],[5]). PCP in the adult wing is manifest in the orientation of wing hairs, which point distally. The anterior, proximal wing is particularly sensitive to Fat-PCP signaling, and fat RNAi results in strong PCP phenotypes in this region, including reversals of hair polarity (Figure S1M). PCP phenotypes have also been described in this region of dachs mutant wings [34]. Zyx RNAi, by contrast, had no detectable effect on wing PCP (Figure S1N), and a PCP phenotype was also still detected in fat Zyx double RNAi wings (Figure S1O). Genes previously identified as influencing Fat-PCP signaling (i.e., fat, ds, fj, app, dachs, lft) also influence cross-vein spacing. Zyx RNAi wings sometimes have extra cross-veins, but by contrast to dachs mutants, the anterior and posterior cross-veins remain well-separated in Zyx RNAi flies (Figure 1B,C), and the influence of fat on cross-vein spacing is not suppressed by Zyx (Figure 1F). Our observations suggest that Zyx is specifically required for Fat-Hippo signaling, and not for Fat-PCP signaling, although because Zyx RNAi might not completely eliminate Zyx, we cannot exclude the possibility that low levels of Zyx are sufficient for PCP, but not for Hippo signaling.
As our anti-Zyx sera did not work for immunostaining, we made use of a V5-tagged UAS transgene that rescues the Zyx RNAi phenotype (Figure 1) to investigate the subcellular localization of Zyx in imaginal discs. We also examined a UAS-Ypet:Zyx transgene [35]. Although our localization studies are subject to the caveat that Zyx protein was over-expressed, the two different tagged Zyx proteins have similar localization profiles, and similar localization profiles were observed using different Gal4 drivers. Zyx was preferentially localized to the sub-apical membrane of disc cells (Figure 6). This sub-apical membrane staining was at the same apical-basal position as E-cadherin (E-cad), and just basal to Fat (Figure 6A–D). This is similar to the membrane localization of Dachs [7]. Indeed, when we compared Zyx and Dachs localization, using epitope-tagged constructs, we observed that the membrane staining is at the same apical-basal position and that they partially co-localize (Figure 6G,H). A distinguishing feature of Dachs localization is its polarization within the plane of the epithelium, which occurs in response to the Fj and Ds gradients (Figure 6J) [7],[9]. Zyx, by contrast, is not planar-polarized (Figure 6I); hence, Zyx and Dachs are expected to overlap on only one side of wing disc cells. A distinguishing feature of Zyx staining is that it often displays puncta of larger, more intense staining at the vertices where three cells meet (Figure 6G). Intriguingly, Ex protein also displays uneven staining, but Ex puncta are partially complementary to Zyx puncta (Figure 6E,F). These observations suggest that even though Ex and Zyx localize to a similar apical-basal position, they assemble into distinct protein complexes. Dachs localization was not visibly affected by RNAi of Zyx (Figure S4E), nor was Zyx localization affected by mutation of dachs (Figure S5B), which indicates that neither protein depends upon the other for its localization. Zyx localization was also not visibly affected by mutation or RNAi of fat, ex, or wts (Figure S5 and unpublished data).
The similar genetic requirements for Zyx and dachs in Fat-Hippo signaling, together with their partial co-localization in imaginal discs, raised the possibility that Zyx and Dachs might interact. This was investigated by expressing tagged isoforms in cultured Drosophila S2 cells and assaying for physical interactions through co-immunoprecipitation. Indeed, Zyx and Dachs could be specifically co-precipitated from S2 cells (Figure 7B). This observation suggests that Dachs and Zyx can interact directly, although it is also possible that they interact indirectly through a larger complex including endogenously expressed proteins within S2 cells.
As Dachs can also associate with Warts in co-immunoprecipitation assays [6], and both Zyx and dachs are required for the fat-dependent regulation of Wts levels, we also investigated binding between Zyx and Wts. When tagged full-length proteins were co-expressed in S2 cells, little or no Zyx-Wts co-precipitation was detected (Figure 7C,H). However, in addition to their role in Hippo signaling, functions for LATS proteins have also been identified in mitosis, and LATS1 has been localized to the mitotic apparatus [36],[37]. In the context of a study of mitotic functions of LATS1, it was reported that the C-terminus of human Zyxin, including the LIM domains, could bind to human LATS1, even though full-length Zyxin did not bind [36]. When we expressed a C-terminal polypeptide comprising the LIM domains of Zyx (Zyx-LD) in S2 cells, only very low levels of protein could be detected (Figure 7B–D). Nonetheless, this C-terminal polypeptide bound efficiently to Wts (Figure 7C). Thus, the LIM domains of Zyx can associate with Wts, but this association is normally inhibited within full-length Zyx.
The discovery of this latent ability of Zyx to bind Wts, together with our discovery of Zyx-Dachs binding, and previous identification of Dachs-Wts binding [6], indicates that Dachs, Zyx, and Wts each have the ability to bind to one another. To gain further insight into complex formation among these proteins, we mapped their interaction domains. Wts bound to the LIM domains of Zyx. Dachs, by contrast, bound most strongly to the C-terminal LIM domains but also bound to the N-terminal half of Zyx (Figure 7B). Dachs contains a large central myosin motor domain and could bind to both Zyx and Wts through this motor domain (Figure 7D,G and unpublished data). Zyx-LD bound to Wts through a region N-terminal to the Wts kinase domain (Figure 7E). Dachs bound both to this region and also to the Wts kinase domain (Figure 7F). Thus, Zyx, Dachs, and Wts interact with each other through partially overlapping domains.
To assay for potential sequential, cooperative, or competitive interactions amongst Zyx, Dachs, and Wts, we examined binding interactions when all three proteins were co-expressed together in S2 cells. A key feature of Zyx's interactions with Wts is that full-length Zyx does not bind efficiently to Wts, but the LIM domains do. However, we found that Dachs enhanced the co-precipitation of full-length Zyx with Wts (Figure 7H). Two basic models for this stimulation of Zyx-Wts association by Dachs can be envisioned: (a) Dachs might bridge Wts and Zyx within a Wts-Dachs-Zyx complex, or (b) Dachs might trigger a conformational change in Zyx that reveals the latent Wts-binding activity of the Zyx LIM domains (Figure 8A,B). By employing V5 epitope tags on both Zyx and Dachs, and assaying their co-precipitation with FLAG-tagged Wts, we could directly compare their association with Wts. A simple trimeric complex model (e.g., one subunit each of Zyx, Wts, and Dachs) would predict that Zyx and Dachs should be present within the Wts trimeric complex at equal levels. However, we found instead that Zyx could be much more abundant in Wts complexes than Dachs (Figure 7H). This suggests that rather than remaining stably associated with Zyx and Wts in a trimeric complex, Dachs is able to stimulate a conformational change in Zyx that exposes the LIM domains and enables them to bind Wts. Consistent with this model, Dachs stimulated Zyx binding to Wts but did not stimulate the binding of Zyx-LD to Wts (Figure S6A).
Zyx is a Drosophila member of a group of cytoskeletal-associated proteins with three C-terminal LIM domains [38]. These comprise two families: the Zyxin family, which in vertebrates includes Zyxin, Lipoma preferred partner (LPP), and Thyroid-receptor interacting protein 6 (TRIP6), and the Ajuba family, which in vertebrates includes Ajuba, LIM domain containing 1 (LIMD1), and Wilms tumor protein 1-interacting protein (WTIP). Drosophila have a single member of each family; Zyx is a member of the Zyxin family, and Ajuba LIM protein (Jub) is a member of the Ajuba family. Ajuba has been reported to interact with a human homologue of Warts, LATS2 [39], and Das Thakur et al. (2010) recently reported that mutation or RNAi-mediated depletion of Jub reduces growth through interactions with the Hippo pathway, and through genetic and protein interaction experiments positioned Jub as a regulator of Wts [40]. In agreement with this, we found that RNAi-mediated depletion of Jub reduces wing growth (Figure 1M,N,S), expression of Hippo pathway target genes, and nuclear Yki (Figure S7), and that wts is epistatic to Jub (Figure 3C). As for Zyx, depletion of Jub did not detectably influence wing hair PCP (Figure S1P,K).
The determination that Zyx and Jub are each genetically required for Hippo signaling suggests that they have distinct functional roles, and consistent with this, we observed that over-expression of Zyx could not rescue Jub RNAi phenotype (Figure S1H) and that Zyx Jub double RNAi induced an even greater reduction of wing size than when they were expressed individually (Figure 1O,S). Das Thakur et al. (2010) did not address the relationship of Jub to upstream regulators of Hippo signaling. Intriguingly, we found that depletion of Jub suppressed both fat and ex phenotypes. This suppression was evident upon examination of adult wings (Figure 1G,K,S), expression of downstream target genes in wing discs (Figures 3F, 4E,F), and the sub-cellular localization of Yki (Figure 4I,L). Thus, by contrast to Zyx, which functions specifically within Fat-Hippo signaling, Jub is required for both Ex-Hippo and Fat-Hippo signaling. This observation confirms that these two LIM-domain proteins have functionally distinct roles within the Hippo pathway.
The distinct genetic role of Jub in Hippo signaling is also reflected in distinct binding interactions. By contrast to the crucial role of Dachs in stimulating binding between full-length Zyx and Wts, full-length Jub binds efficiently to Wts, and full-length vertebrate homologues of Jub bind to LATS proteins [39],[40]. Moreover, Jub bound only very weakly Dachs (Figure S6B). Thus, although Zyx and Jub share the ability to associate with Wts through their LIM domains, both genetic and biochemical studies indicate that the regulation and consequences of these LIM-domain-Wts interactions are distinct.
Our characterization of Zyx identifies a role for it as a novel and integral component of the Hippo pathway, which is required for the Fat branch, but not the Ex branch, of Hippo signaling. Unlike most previously identified components, loss of Zyx reduces the activity of the key transcriptional effector of the pathway, Yki, and consequently its loss reduces organ growth. Genetic epistasis experiments position the requirement for Zyx in between fat and wts, and concordant protein binding experiments identify a Dachs-stimulated ability of Zyx to bind Wts protein. We infer that this association of Zyx with Wts then downregulates Wts, at least in part, by targeting it for degradation.
Zyx localizes to the sub-apical membrane independently of Fat or Dachs. Since Fat regulates the localization of Dachs [7], this regulated localization provides a mechanism by which Fat could modulate the interaction of Dachs with Zyx (although we note that Fat might affect the activity of Dachs in addition to affecting its localization). Since Dachs stimulates Zyx-Wts binding, this regulated localization provides a means for Fat signaling to modulate Zyx-Wts binding. We infer that Dachs effects a conformational change in Zyx, as in the absence of Dachs a Zyx LIM-domains polypeptide binds efficiently to Wts, whereas full-length Zyx binds poorly. Intriguingly, the association of vertebrate homologues of Zyx and Warts can also be post-translationally regulated, as the ability of the LIM domains of human LATS1 to bind Zyxin is masked within full-length Zyxin, but uncovered by Cdc2-mediated phosphorylation, presumably due to conformational change [36]. We hypothesize that the ability of Dachs to bind to both the N-terminus and the LIM domains of Zyx enables it to effect a conformational change in Zyx, resulting in an open configuration that can bind to Wts (Figure 8B). It is also possible that Dachs binding stimulates a post-translational modification of Zyx to induce a conformational change.
Prior studies identified two mechanisms by which Fat signaling could influence Yki activity, as fat mutation reduces both the levels of Wts protein [6] and the amount of Ex at the sub-apical membrane [31]–[33]. It has not been possible to completely uncouple these two pathways for Fat-Hippo signaling, although the observation that over-expression of Wts can efficiently suppress fat overgrowth phenotypes, but only partially suppresses ex overgrowth phenotypes [30], suggested that the influence of Fat on Wts levels might be more critical. Analysis of the influence of Zyx on Ex is complicated by its influence on ex transcription, but our observation that reduction of Zyx does not appear to suppress the influence of fat on Ex staining, even though it does suppress the influence of fat on Wts levels, also suggests that the influence of Fat on Wts levels might be more critical than its effects on Ex. Intriguingly, mutation of dachs did suppress the influence of fat on Ex levels [30]. Although it is possible that this difference between dachs and Zyx results from technical differences in the experimental paradigms (e.g., mutant clones versus RNAi), it is also possible that dachs can influence Ex levels independently from its association with Zyx.
The discovery of the Fat-specific effect on Wts levels, by contrast to the Hippo-pathway-mediated effect on Wts kinase activity, established the concept of distinct mechanisms for regulating Wts—one that affects Wts levels and another that affects Wts activity [6]. Our identification of distinct genetic requirements for Zyx and Jub provide further support for this concept. As Jub is equally required for both Fat-Hippo and Ex-Hippo signaling and acts genetically between hippo and wts [40], Jub appears to inhibit Wts activation. In our working model (Figure 8C), the epistasis of Jub to fat could be explained by an increased activity of residual Wts, which then acts catalytically to repress Yki activity. Zyx is required for the influence of fat on Wts levels. We note that when measured within a whole tissue lysate, Wts levels are only reduced to approximately half their normal levels. However, as Wts appears to function within multi-protein complexes, including some components that can localize preferentially to the sub-apical membrane [41],[42], we hypothesize that Fat signaling affects a discrete pool of Wts within a complex at the membrane that is crucial for Hippo signaling, whereas there might be additional pools of Wts within the cell that are unaffected. We also note that while we clearly see effects on Wts protein levels, our results do not exclude the possibility that Fat signaling also influences Wts activity.
Our characterization of Zyx and Jub also provides new tools for analyzing critical steps in Hippo signaling. For example, in addition to influencing Hpo and Wts kinase activity, it has been observed that Ex can bind directly to Yki and that when Ex is over-expressed it can repress Yki through a mechanism that involves direct sequestration of Yki, rather than regulation of Yki phosphorylation [43],[44]. Because this direct repression mechanism was based on over-expression experiments, the extent to which it contributes to normal Yki regulation in vivo remained uncertain. The observations that Jub acts genetically upstream of wts, yet is required for ex phenotypes, suggests that Ex regulates Yki principally through its effects on Wts activity, rather than through direct interaction with Yki.
The ability of Zyx LIM domains to interact with Wts is conserved in their human homologues [36]. Although the functional significance of this interaction in vertebrates has not yet been established, our observations raise the possibility that the oncogenic effects of human LPP mutations [13] could be due to an ability of these aberrant LPP fusion proteins to negatively regulate LATS proteins, resulting in inappropriate activation of YAP or TAZ.
One of the most intriguing aspects of Zyxin family proteins is their role in mediating effects of mechanical force on cell behavior [18]. Zyxin family proteins can localize to focal adhesions of cultured fibroblasts, and this localization is modulated by mechanical tension [15],[18],[45]. The observation that increasing tension on stress fibers stimulates Zyxin accumulation at focal adhesions is intriguing in light of our observation that Zyx tends to accumulate at higher levels at intercellular vertices in imaginal discs, as these could be points of increased tension. As the association of unconventional myosins with F-actin can also be influenced by external force [46], our discovery of binding between a myosin protein (Dachs) and Zyx raises the possibility that other myosins might also interact with Zyxin family proteins, which could potentially influence either their tension-based recruitment or their activity.
Finally, we note that theoretical models of growth control in developing tissues have proposed that growth should be controlled by mechanical tension [47],[48], and direct evidence for mechanical effects on growth has been obtained in cultured cell models [49]. However, a mechanism for how this might be achieved has been lacking. Our discovery that Zyx, a member of a family of proteins implicated in responding to and transducing the effects of mechanical tension, is also a component of the Hippo signaling pathway, a crucial regulator of growth from Drosophila to humans, raises the intriguing possibility that Zyxin family proteins might form part of a molecular link between mechanical tension and the control of growth.
RNAi screening was conducted using lines from the NIG-Fly Stock Center (http://www.shigen.nig.ac.jp/fly/nigfly/index.jsp), which were crossed to vg-Gal4 UAS-dcr2 or pnr-Gal4 UAS-dcr2. Those with growth phenotypes were then re-screened for effects on Diap1 and Wg expression in imaginal discs by crossing to ci-Gal4 UAS-dcr2 or en-Gal4 UAS-dcr2. All crosses were carried out at 28.5 C to obtain stronger phenotypes. Approximately 1,200 lines were examined in the initial screen (Table S1).
Additional RNAi lines employed include ds [vdrc 36219], fat [vdrc 9396], d [vdrc 12555], ex [vdrc 22994], Zyx [NIG-32018R3], Zyx [vdrc 21610], wts [vdrc 9928], wts [NIG-12072R1], mats [vdrc 108080], hpo [vdrc 104169], Jub [vdrc 101993], and Jub [vdrc 38442]. The effectiveness of fat and ex RNAi is illustrated in Figure S3E,F. Both Zyx RNAi lines gave similar effects on growth and gene expression in combination with multiple Gal4 lines and also behaved similarly in epistasis tests. UAS lines employed include UAS-dco3[29],[48], UAS-d:V5[9F] and UAS-d:V5[50] [7], UAS-d:citrine[28] (B.K. Staley, unpublished), UAS-Zyx:V5, and UAS-Ypet:Zyx [35]. Gal4 lines employed include Dll-Gal4, ex-lacZ en-Gal4 UAS-GFP/CyO;UAS-dcr2/TM6b, en-Gal4/CyO; th-lacZ UAS-dcr2/TM6b, ci-Gal4 UAS-dcr2[3]/TM6b, w UAS-dcr2[X]; nub-Gal4[ac-62], w; AyGal4 UAS-GFP/C yO;UAS-dcr2/TM6b, y w hs-FLP[122]; AyGal4 UAS-GFP/CyO, tub-Gal80ts/CyO,Act-GFP; tub-Gal4 UAS-dcr2/ TM6b, w; tub-Gal4/CyO-GFP. MARCM clones were made by crossing y w hs-FLP[122] tub-Gal4 UAS-GFP/FM7 ; tub-Gal80 FRT40A/CyO to fat8 FRT40A/CyO, exel FRT40A/CyO, dGC13 FRT40A/CyO or y+ FRT40A (as a control) and UAS-zyxin:V5. Flp-out clones were made by crossing y w hs-FLP[122]; AyGal4 UAS-GFP to UAS-zyxin:V5 or crossing AyGal4; UAS-d:citrine to y w hs-FLP[122]; UAS-zyxin:V5. Genetic interaction of Zyx and dachs was examined by recombining nub-Gal4 with dGC13 and crossing to dGC13; RNAi-Zyx32018.
Adult wing phenotypes were scored by crossing UAS-dcr2; nub-Gal4 females to males of RNAi lines or Oregon-R males as a control. Wings of male progeny were photographed, all at the same magnification. For quantitation, between 9 and 12 wings per genotype were traced using NIH Image J, and wing areas were normalized to the average area in control males. Standard error of the mean (s.e.m.) and t tests were calculated using Graphpad Prism software.
For analysis of gene expression in imaginal discs, ex-LacZ en-Gal4 UAS-GFP; UAS-dcr2 females were crossed to RNAi line males, and larvae were kept at 28.5 C until dissection. For analysis of Zyx:V5 or Ypet:Zyx localization, expression was driven by en-Gal4, AyGal4, or tub-Gal4. Discs were fixed in 4% paraformaldehyde and stained using as primary antibodies: goat anti-ß-galactosidase (1∶1,000, Biogenesis), mouse anti-Diap1 (1∶200, B. Hay), rat anti-E-cad (1∶200, DSHB), guinea pig anti-Ex (1∶2000, R. Fehon), rat anti-Fat (1∶400) [29], mouse anti-V5 (1∶400, Invitrogen), mouse anti-Wg (1∶400, DSHB), and rabbit anti-Yki (1∶400) [22]. F-actin was stained using Alexa Fluor 546 phalloidin (1∶100, Invitrogen), and DNA was stained using Hoechst (Invitrogen).
Details of plasmid construction are in Text S1.
Co-immunoprecipitation assays were performed as described previously [6]. Cell lysates were cleared using protein G beads (Sigma). Anti-V5 or anti-FLAG M2 beads (Sigma) were incubated with cell lysates overnight at 4°C, then washed six times with RIPA buffer and boiled in SDS-PAGE loading buffer. Primary antibodies used for blotting include rabbit anti-V5 (1;10,000, Bethyl), mouse anti-V5 (1∶10,000, Invitrogen), and mouse anti-FLAG M2 (1∶10,000, Sigma), and were detected using anti-mouse IRdye680 and goat anti-rabbit IRdye800 (1∶10,000, LiCor) and scanning on a LiCor Odyssey.
For analysis of Wts protein levels, tub–Gal4 UASdcr2/ TM6b females were crossed to white (control), RNAi-fat, RNAi-Zyx, RNAi-fat; RNAi-Zyx, or UAS-Zyx:V5 males, and wing discs were dissected from third instar larval progeny and lysed in RIPA buffer. Amounts loaded were adjusted to try to load equivalent amounts of total protein in each lane. Wts was detected using a published Wts anti-sera [6] at 1∶4,000. Protein bands were detected using anti-mouse IRdye680 and goat anti-rabbit IRdye800 (1∶10,000, LiCor) and scanning on a LiCor Odyssey. Bands were quantified using LiCor Odyssey software. Relative Wts levels were determined by comparison to bands detected by anti-Actin antibodies (mouse anti-Actin at 1∶5,000, Calbiochem). To enable the relative levels of Wts to be averaged across different blots, we normalized the ratios on each blot to that detected for the control lane, which was set as 1.
For confirmation of the influence of Zyx RNAi on Zyx protein levels, tub–Gal4 UASdcr2/TM6b females were crossed to white (control), or RNAi-Zyx32018, and cultured at 29 C, and wing discs were dissected from third instar larval progeny and lysed in RIPA buffer. A rabbit anti-Zyx sera was used at a 1∶2,000 dilution, and subsequently the blot was re-probed with rabbit anti-actin (1∶10,000, Sigma). Fluorescent detection was performed as described above. Anti-Zyx sera was obtained by immunization of rabbits with a KLH conjugated peptide (KRRLDIPPKPPIKY), performed by Open Biosystems.
|
10.1371/journal.ppat.1003757 | A Small Molecule Glycosaminoglycan Mimetic Blocks Plasmodium Invasion of the Mosquito Midgut | Malaria transmission-blocking (T-B) interventions are essential for malaria elimination. Small molecules that inhibit the Plasmodium ookinete-to-oocyst transition in the midgut of Anopheles mosquitoes, thereby blocking sporogony, represent one approach to achieving this goal. Chondroitin sulfate glycosaminoglycans (CS-GAGs) on the Anopheles gambiae midgut surface are putative ligands for Plasmodium falciparum ookinetes. We hypothesized that our synthetic polysulfonated polymer, VS1, acting as a decoy molecular mimetic of midgut CS-GAGs confers malaria T-B activity. In our study, VS1 repeatedly reduced midgut oocyst development by as much as 99% (P<0.0001) in mosquitoes fed with P. falciparum and Plasmodium berghei. Through direct-binding assays, we observed that VS1 bound to two critical ookinete micronemal proteins, each containing at least one von Willebrand factor A (vWA) domain: (i) circumsporozoite protein and thrombospondin-related anonymous protein-related protein (CTRP) and (ii) vWA domain-related protein (WARP). By immunofluorescence microscopy, we observed that VS1 stains permeabilized P. falciparum and P. berghei ookinetes but does not stain P. berghei CTRP knockouts or transgenic parasites lacking the vWA domains of CTRP while retaining the thrombospondin repeat region. We produced structural homology models of the first vWA domain of CTRP and identified, as expected, putative GAG-binding sites on CTRP that align closely with those predicted for the human vWA A1 domain and the Toxoplasma gondii MIC2 adhesin. Importantly, the models also identified patches of electropositive residues that may extend CTRP's GAG-binding motif and thus potentiate VS1 binding. Our molecule binds to a critical, conserved ookinete protein, CTRP, and exhibits potent malaria T-B activity. This study lays the framework for a high-throughput screen of existing libraries of safe compounds to identify those with potent T-B activity. We envision that such compounds when used as partner drugs with current antimalarial regimens and with RTS,S vaccine delivery could prevent the transmission of drug-resistant and vaccine-breakthrough strains.
| To achieve malaria elimination, the consensus expert opinion is that new approaches to drug and vaccine design are desperately needed. We have undertaken a novel, comprehensive approach towards the development of a malaria transmission-blocking drug based on the strategy of inhibiting Plasmodium development in the mosquito by interfering with obligate cellular interactions between the parasite and the mosquito-midgut epithelium. We have successfully designed a potent transmission-blocking small molecule (VS1) that mimics the structure of molecules on the mosquito-midgut surface called glycosaminoglycans (GAG), which are thought to serve as ligands for parasite attachment prior to cell invasion. Using assays in which mosquitoes were fed with infectious blood, we tested the effect of VS1 on Plasmodium development in the mosquito and found that the GAG mimic dramatically reduced the intensity of infection in the midgut. Binding experiments and immunofluorescence microscopy indicate that VS1 binds to the circumsporozoite- and TRAP-related protein (CTRP), a micronemal protein expressed by ookinetes essential for midgut invasion. This interaction profoundly inhibits a key step of parasite development, thereby abrogating downstream events necessary for mosquito-to-human transmission. The work described lays the framework for bringing a truly novel transmission-blocking drug to fruition.
| Each year more than half a million people die from malaria, a disease caused by protozoan parasites in the genus Plasmodium. The life cycle of Plasmodium parasites includes asexual development in the human host and obligatory sporogonic development in the Anopheles mosquito vector with transmission from person to person only made possible through the bite of an infected anopheline. Despite substantial investment in malaria research, it is widely accepted that current interventions are insufficient to achieve the ultimate goal of eradication and that a combination of anti-malaria strategies including those that target parasite transmission give eradication efforts the best chance to succeed [1]. Moreover, the evolutionary capacity of vectors and parasites to overcome chemical- and drug-based interventions emphasizes the need for new weapons in the anti-malaria arsenal. It is in this context that malaria transmission-blocking (T-B) interventions (vaccines and drugs) have received significant attention [2]. In fact, recent progress has shown that probing the basic biology underlying mosquito-Plasmodium interactions can identify novel intervention targets not only in the parasite, but in the mosquito as well [3]–[6]. Importantly, seminal work by Delves, et al. [7] have brought increased attention to the potential T-B activity of drugs that failed to demonstrate efficacy against Plasmodium asexual stages but have been resurrected as novel T-B candidate compounds. These efforts highlight the need for T-B molecules and open new avenues for the development and/or repurposing of compounds that have direct activity against the parasite soon after ingestion into the mosquito midgut during blood feeding.
In the mosquito blood meal, Plasmodium gametocytes differentiate into gametes and fuse to form zygotes, which then develop into motile ookinetes. For parasite development to continue, ookinetes must find and adhere to membrane-associated ligands on the midgut epithelial surface, a pre-requisite for cell invasion. Experimental evidence from Plasmodium berghei and Plasmodium falciparum suggests that ookinete attachment and invasion is mediated by micronemal proteins, including the circumsporozoite protein and thrombospondin-related anonymous protein-related protein (CTRP) [8]–[13] and von Willebrand factor A domain-related protein (WARP) [10], [14]. The function of WARP is unclear, while CTRP has a demonstrated role in ookinete motility [9], [15]. However, both are essential for midgut epithelial cell invasion by Plasmodium ookinetes. Once inside the cell, ookinetes make their way to the midgut basal lamina where they differentiate into oocysts, each giving rise to thousands of sporozoites that are released into the hemocoel upon maturation and rupture. Sporozoites are then swept into the circulating hemolymph and carried to the salivary glands. Following successful salivary gland invasion, sporozoites remain in the lumen of the salivary duct until host delivery during blood feeding. Clearly, negotiating the midgut tissue barrier in the vector is crucial for successful establishment of the parasite in the mosquito and hence, subsequent transmission to human hosts.
In this study we exploit knowledge of crucial molecular interactions between Plasmodium ookinetes and the apical microvillar surface of the mosquito midgut to design proof-of-concept small molecules that interfere with ookinete attachment. Previous work demonstrated that sulfated glycosaminoglycans (GAGs) are present on both the apical and basal surfaces of the midgut epithelium, with chondroitin sulfate (CS) predominant on the apical side (i.e., facing the midgut lumen) and heparan sulfate (HS) predominant on the basal side [5], [16] (Figure S1A). RNAi-mediated knockdown of the Anopheles gambiae peptide-O-xylosyltransferase, an enzyme that catalyzes the first step in CS and HS biosynthesis, resulted in mosquitoes with CS-depleted midgut apical surfaces [5]. When infected with P. falciparum and P. berghei in feeding assays, these mosquitoes demonstrated significantly lower oocyst infection intensities relative to controls. The study also showed binding affinity for two types of CS (CS-A and CS-E) in mature ookinetes, consistent with a previous report demonstrating that the ookinete micronemal proteins PfWARP and PfCTRP bind to sulfated GAGs in vitro [10]. Inspired by the idea that these molecular interactions could be disrupted by small molecules that mimic the charged structural elements involved in ligand binding, two short-chain, water-soluble compounds were synthesized for in vitro and in vivo T-B studies based on their potential to interfere with parasite protein-GAG interactions (Figure 1A).
Here we describe efforts to test this strategy with the underlying hypothesis that when a mosquito ingests these small molecules in an infectious blood meal, the compounds will interfere with ookinete-GAG interactions, therefore preventing midgut invasion and abrogating subsequent developmental steps in sporogony. Based on key studies in the literature [9]–[15], we further hypothesize that the molecular basis of ookinete-GAG interactions, and hence those between ookinetes and our GAG-mimetics, involve the Plasmodium micronemal proteins CTRP and/or WARP. Our findings showed that this novel strategy dramatically reduced infection intensity in the mosquito midgut, and multiple lines of evidence suggest that the mechanism underlying the T-B effect involved binding of our GAG-mimetic decoy to one or more von Willebrand factor A (vWA) domains found in the protein CTRP.
Synthesis of polysulfonated polymers (VS1 and VS2-PVP) proceeded with the addition of 100 mg of potassium persulfate to 5 g (0.38 mmol) of vinyl-sulfonic acid (VS1) sodium salt water solution and adjusted to basic pH with sodium hydroxide. The final solution was warmed for 20 hr at 80°C, then cooled to room temperature (RT), diluted with water and ultra-filtered through a membrane with a nominal cut-off of 10,000 Da. The fraction retained was freeze-dried and the product obtained was a white powder. Size exclusion chromatography was used following different reaction times to obtain oligomers of different length and molecular mass. These compounds were then purified by ultra-filtration through different cut-off membranes (500 Da, 1000 Da, 5000 Da), and average molecular weights were measured by size exclusion chromatography and MALDI spectrometry.
Transmission-blocking assays for vinyl-sulfonic acid compounds (including preliminary VS1 and VS2-PVP experiments) were tested using both in vivo and in vitro systems. In vivo studies were performed using the murine malaria parasite P. berghei (ANKA 2.34) following IACUC approved protocols. For each experiment, two to three naïve, donor mice (Swiss Webster, 20–24 g) were inoculated (i.v.) with blood stage P. berghei and then checked for parasitemia by blood smear five to six days later. Once parasitemia reached ≥10%, donors were sacrificed via cardiac puncture and parasitemic blood was used to inoculate (i.v.) eight to ten experimental mice per test compound. Two to three days post-inoculation, experimental mice were smeared and checked for exflagellating gametocytes. Mice demonstrating an average of at least 1 and fewer than 6 exflagellations per 40× field were assigned to a treatment group, weighed and anesthetized. For each mouse, a pre-injection 500 ml cup of Anopheles stephensi mosquitoes (n = 50) were allowed to feed for 15 to 20 min. The mouse was then removed from the mosquito cup, injected with either a vinyl sulfonic acid compound (250 µg/ml or 500 µg per 24 g body weight), polyvinylpyrrolidone (PVP, same dose as VS1or VS2-PVP), or the carrier only (PBS) via tail vein injection (iv) and then allowed to recover for 10 to 15 min. Following recovery, a post-injection cup of mosquitoes (n = 50) was allowed to feed as before. Unfed mosquitoes were then removed from both pre- and post-injection cups via mouth aspiration. For each control and test compound, three to five pre- and post-injection sets of mosquitoes were maintained on sucrose and water for 10 days at 19°C, 80% relative humidity. On day 10, midguts were dissected from all surviving mosquitoes and stained with 0.1% mercurochrome for 20 mins. Oocyst number for each midgut was determined by microscopy and at least three independent experiments were performed for each compound.
In vitro studies were performed using the human malaria parasite P. falciparum and the old-world malaria vectors Anopheles gambiae and An. stephensi as described [4]–[5]. With the exception of the parasite load experiment, each set of studies consisted of independent experiments in which the age of the gametocyte culture (16–17 days), the age of mosquitoes (4–6 days), and the blood-meal gametocytemia (0.3%) and hematocrit (45%) were kept consistent. In the parasite-load experiment, all else was the same except for the blood-meal gametocytemia which varied as described in the Results. For each experimental treatment, VS1 or the control compound (PVP) were prepared in PBS and diluted 1∶10 to the final experimental dose with infected blood.
Full-length PvWARP excluding the signal peptide (nt 88–867) and a fragment of PvCTRP containing the first vWA domain (nt 79–921) were PCR-amplified from genomic DNA of the Salvador I strain with a C-terminal 6×His tag appended to the reverse primer. Fragments were cloned into the EcoRV sites of the vector pEU-E01-MCS (Cell Free Sciences, Matsuyama, Japan). The PvCTRP-vWA1 and PvWARP were expressed in the wheat germ cell-free expression system (Cell Free Sciences, Matsuyama, Japan) as described [17] and purified using Ni-affinity chromatography.
Biotinylated VS1-NH2 in a volume of 100 µl (10 µg/ml) was applied to each well of a streptavidin-coated (2 µg/ml) 96-well microtiter plate that had been blocked with PBS, 1% BSA (Thermo Pierce) and incubated for two hours at RT. During VS1 incubation, 6×His-tagged recombinant PvWARP and PvCTRP (5 µg/ml) were each mixed separately with heparin and CSA (100 µg/ml) in blocking buffer and incubated for 2 hr at RT. The microtiter plate was subsequently washed three times with PBS to remove excess VS1, and then 100 µl of 6×His-tagged recombinant PvWARP or PvCTRP alone (10 or 5 µg/ml) or of the recombinant protein-GAG mixture was added to each well, with the exception of the no-protein and irrelevant-protein controls. The latter received a 6×His-tagged recombinant glycosyltransferase from An. gambiae. Binding and/or competition with VS1 was allowed to proceed for 2 hr at RT. Following three washes with PBS, anti-His MAb (Sigma) was added to each well and incubated for 1 hr at RT. After three washes with PBS + Tween-20 (0.05%), anti-mouse secondary antibodies conjugated to HRP were added and incubated for 1 hr at RT. Following another wash step, TMB ELISA substrate (Pierce) was used for detection. VS1 binding was quantified by measuring the OD at 450 nm with a SPECTRA MAX PLUS microplate reader (Molecular Devices).
Ookinete samples were fixed with 4% paraformaldehyde and prepared for immunofluorescence microscopy by washing three times with PBS containing 0.1 M glycine (rinsing buffer). To permeabilize samples the parasites were incubated with rinsing buffer containing 0.2% Triton X-100 for 10 min and then washed as before. After the washes, samples were incubated in rinsing buffer for 30 min and blocked with PBS containing 0.05 mM glycine, 0.2% fish skin gelatin and 0.05% sodium azide for 2 hr. The samples were then incubated with biotinylated-VS1 and anti-Pbs21 (P. berghei) or anti-Pfs28 (P. falciparum) for 1 hr at RT or overnight at 4°C. Cells were washed as before and incubated with Streptavidin, DyLight 488 conjugated (Thermo), and Goat Anti-Rabbit IgG (H+L), DyLight 594 conjugated, for 1 hr at RT. Following incubation, the cells were washed three times with rinsing buffer, resuspended in PBS, spotted on slides and allowed to air dry. ProLong Gold antifade reagent with DAPI (Invitrogen) was added prior to the coverslip and slides were incubated for 24 hr at RT protected from light. Samples were examined with SPOT software using a Nikon Upright E800 microscope.
Homology modeling of the CTRP vWA domain was performed using SWISS-MODEL in two different modes of operation [18]–[21]. In the full-automated mode, the Toxoplasma gondii micronemal protein 2 I domain (2XGG, chain B, residues 75–212) was selected as the optimal template to calculate the CTRP model (residues 1–148) and is based on 22% sequence identity. In the template identification mode and using the InterPro Domain Scan method [22], the von Willebrand factor A1 domain (1AUQ, chain A, residues 1276–1463) was selected as the optimal target template (residues 1–193). The model quality was assessed using the QMEAN server [23] and the Z-score.
To determine significance between treatment and control groups in the feeding assays, the nonparametric Mann-Whitney U test was used due to the non-normal distributions typical of oocyst counts. For the in vitro assays, the test was performed comparing the distribution of oocyst counts per midgut for each treatment group to that of the PVP control, followed by a Bonferroni correction of z-scores to adjust for multiple tests. For the in vivo assays, the test was performed comparing oocyst counts per midgut between mosquitoes fed on P. berghei-infected mice pre- and post-injection with VS1, PVP, or PBS alone.
All experimental studies using vertebrate animals (mice) were performed in accordance with Johns Hopkins University (JHU) ACUC (Animal Welfare Assurance #A3272-1) regulations. The Animal Protocol (#MO12H232) used for these studies was reviewed and approved by the JHU ACUC and are in compliance with the United States Animal Welfare Act regulations and Public Health Service (PHS) Policy. No human subject research was performed during this study.
Our primary goal was to design synthetic polymers that can mimic sulfated CS-GAGs that have been shown to bind ookinetes (Figure 1A, Figure S1A). However, the likelihood of CS-GAGs, with C4S and C6S sulfation on the midgut microvillar surface was challenged [16]. Using capillary electrophoresis with laser-induced fluorescence detection, we confirmed the presence of both C4S and C6S chondroitin GAGs on An. gambiae midgut brush border microvilli vesicles (Figure S1A, Text S1). Based on these data, we hypothesized that polymers with high sulfation densities would be appropriate for our study. As such, two polysulfonated polymers were generated by polymerization of vinyl-sulfonic acid (VS1) and copolymerization of vinyl-sulfonic acid with 1-vinyl-2-pyrrolidone (VS2-PVP). The sulfate groups on the polysulfonated polymers are anionic at physiologic pH and would presumably bind to ookinete proteins that would naturally bind to GAGs on the midgut surface. Note that it was previously shown that the blocking phenomenon is predicted to occur at the apical midgut surface, and that basal lamina GAGs do not influence the ultimate read-out of these T-B studies, which is oocyst prevalence and intensity measurements at 8 or 10 days post-blood feeding, for P. falciparum and P. berghei, respectively [5].
Initial data from malaria T-B studies indicated that VS1 was non-toxic and well tolerated by both mice and mosquitoes. Although VS2-PVP was tolerated by mice, mosquitoes that ingested the compound had poor survivorship in multiple experiments within 24 hrs following blood feeding. This low survivorship prevented a proper comparison of infection intensity between midguts dissected from pre- and post-injection mosquitoes; and consequently, this compound was not pursued further. In preliminary in vitro standard membrane feeding assays (SMFAs), VS1 demonstrated 98.5% and 92.3% inhibition of P. falciparum oocyst development in An. gambiae and An. stephensi, respectively. In vivo direct feeding assays (DFA) with mice infected with P. berghei demonstrated a somewhat lower effect of the compounds on parasite development in An. stephensi, with VS1 reducing oocyst intensity relative to controls by 77%. These preliminary data from both in vitro and in vivo malaria models demonstrated the potential for VS1 to act as a potent T-B compound. Moreover, to exclude the possibility that micro- and macrogamete fertilization events could be adversely affected by VS1, we tested the effect of the compound on male microgamete exflagellation and noted that the number of exflagellation centers were unaffected (data not shown). We therefore pushed this small molecule forward as the lead compound for further testing, which included (i) assays to determine if T-B activity varies according to polymer length, (ii) dose-ranging assays to estimate the IC50 of VS1, (iii) ELISA-based binding and competition assays using a candidate gene approach, specifically recombinant versions of the ookinete micronemal proteins CTRP and WARP, and (iv) immunofluorescence microscopy to confirm binding of VS1 to wild-type ookinetes from both P. falciparum and P. berghei, as well as ookinetes from knockout lines of P. berghei.
To assess the influence of polymer length, the product following VS1 synthesis was fractionated into three molecular-weight categories by size exclusion chromatography, VS1-10,000; VS1-3,000; and VS1-1,000 (Figure 1A). Based on preliminary dose ranging experiments, each new compound was then tested at a concentration of 250 µg/ml using both in vitro (SMFA) and in vivo (DFA) malaria models. Polyvinylpyrrolidone, a non-sulfated, neutrally charged control (PVP), which represents the unsulfated VS1 backbone, was used as a control. In SMFAs the three VS1 compounds were tested in parallel against P. falciparum in two replicate experiments using two different vector species, An. gambiae and An. stephensi. In all four experiments, each of the VS1 compounds significantly reduced oocyst intensity (Figure 1B–E). VS1-3,000 consistently performed the best; inhibiting oocyst development by 86.0%–99.0% in An. gambiae (Figure 1 B, C) and 88.0%–93.5% in An. stephensi (Figure 1 D, E). Experiments with the in vivo system yielded similar results, as all three VS1 compounds strongly inhibited P. berghei oocyst development in An. stephensi (Table 1, Figure 2, Table S1). Results from two replicate experiments per compound showed that all six VS1 treatment groups experienced a highly significant reduction in median oocyst intensity when comparing mosquitoes fed on pre-injection mice with those fed on mice injected with VS1 (Table 1, Table S1). In five of these treatment groups, oocyst development was inhibited by >90%. VS1-3,000 had the strongest effect, demonstrating ≥98.0% inhibition on average in both experiments (Table 1, Table S1B). Conversely, pre- and post-injection comparisons of the oocyst burden in mosquitoes fed on mice from PBS and PVP groups demonstrated no consistent effect of either the VS1 carrier or the unsulfated polymer (Table 1, Table S1). To confirm that VS1 was available to mosquitoes in blood meals, and hence the likely cause of T-B activity, the presence of VS1 in the mouse bloodstream following injection was confirmed by HPLC (Figure S2, Text S1).
Due to its consistent T-B activity, VS1-3,000 was selected for use in P. falciparum NF54 SMFA experiments to assess the compound's effectiveness across variations in parasite load in the blood meal (gametocytemia) and to estimate the IC50 of VS1 in two different anopheline vectors. For the parasite-load experiments, we chose to test VS1 potency at levels of gametocytemia that captured values routinely observed during the conduct of membrane feeding assays in the field [24]–[25]. With the concentration of VS1-3,000 set at 250 µg/ml, a SMFA was performed in which four gametocyte concentrations were tested in the presence and absence of VS1. In this experiment, a day 17 gametocyte culture at 3.0% gametocytemia was pelleted and the packed red blood cells (RBCs) were diluted with uninfected blood to 0.3% and 0.1% gametocytemia. Each of these dilutions was in turn diluted 1∶10 with uninfected blood yielding 4 concentrations that ranged from 0.01%–0.3% gametocytemia (∼800–24,000 stage V gametocytes per µl of blood). The level of gametocytemia commonly used in laboratory-based SMFAs (0.3%) is typically much higher than that found in the field to ensure consistent and robust infections in mosquitoes, allowing more rigorous tests of T-B activity [25]. Though a widely accepted approach, a criticism of the SMFA is that the assay better tests the effects of compounds (or antibodies) on oocyst intensity than prevalence of infection among mosquitoes. Since the ultimate goal is to reduce the latter to zero, we wanted to perform the SMFA over a range of gametocytemias once we established that VS1 consistently reduces the oocyst burden at the usual gametocytemia of 0.3%. In this set of experiments, VS1-3,000 once again demonstrated a potent reduction in oocyst intensity at 0.3% gametocytemia, reducing the median oocyst number per midgut from 92.0 to 8.0 (Figure 3A). However, the prevalence of infection was unchanged between carrier-only and VS1 treatments at this level of gametocytemia. Interestingly, as the gametocytemia was reduced from 0.3%, the effect of VS1 on infection prevalence increased while the reduction in oocyst intensity remained high (Figure 3A, B). In fact, at the two levels of gametocytemia most relevant to the field (i.e., 0.03% and 0.01%), the median oocyst number was reduced to 0 in the VS1 treatments while prevalence was reduced from 89% and 67% in carrier-only treatments to 21% and 13% in VS1 treatments, respectively (Figure 3B, C). In other words, at levels of gametocytemia where untreated mosquitoes averaged fewer than 5 oocysts per midgut and where most mosquitoes were infected, VS1 treatment reduced infection prevalence 4–5 fold and infection intensity by 10 fold. Under these conditions, most VS1-treated mosquitoes were uninfected, while the few that were tended to have a single oocyst.
In a set of two dose-ranging experiments with An. gambiae and two with An. stephensi, VS1-3,000 was fed to mosquitoes in infectious blood meals using serially diluted concentrations from 400 µg/ml to 12.5 µg/ml. The four experiments revealed a consistent pattern of percent inhibition characterized by a linear increase from little to no inhibition at 12.5 µg/ml to approximately 80% at 100 µg/ml and then a plateau >90% at concentrations >200 µg/ml (Figure 3D). From these data the IC50 of VS1-3,000 was approximated to be 25 µg/ml.
Since VS1 is a hypothesized structural mimetic of midgut-microvillar sulfated GAGs, we sought to determine whether VS1 can directly bind to Plasmodium ookinetes. To this end, we used biotinylated VS1-NH2 (Figure 1A) to probe non-permeabilized and permeabilized P. berghei ookinetes generated in vitro, as well as ex vivo blood-meal derived P. falciparum ookinetes isolated from dissected mosquito midguts. Only permeabilized Plasmodium ookinetes showed strong binding affinity to VS1 by immunofluorescence microscopy. The VS1 staining pattern suggested that it is not associated entirely with the ookinete surface since it did not consistently bind to non-permeabilized ookinetes nor did it localize with the abundant ookinete surface marker P28 (also called Pbs21 in P. berghei) in either P. falciparum (WTPf) or P. berghei (WTPb) (Figure 4A, 4B). VS1 binding appears to be centrally and apically localized in the cytoplasm, suggestive of interaction with micronemal proteins since these proteins are not constitutively secreted to the ookinete surface and are stored in the micronemes [10], [26]. In addition to ookinetes, VS1 also bound to P. falciparum retorts (i.e., developing ookinetes) (data not shown), and a portion of permeabilized P. falciparum and P. berghei round cells. These cells are likely to be zygotes or unfertilized macrogametes (Figure S3E–L) since they are stained with P28, a surface marker well described in the literature, which is expressed from macrogametes to early oocyst [27]–[29] ruling out the possibility that these cells are P. berghei or P. falciparum gametocytes. Furthermore, P. falciparum stage IV and V gametocytes are not round but have a distinctive elongated morphology, and these cells did not stain withVS1 in subsequent immunofluorescence experiments targeting gametocytes (Figure S4A–D, Text S1). We cannot rule out an effect of VS1 in macrogametes and/or zygotes because those stages were not the focus of the current research, but any effect in these stages will potentiate the effect of VS1 against transmission of Plasmodium. We evaluated the effect of VS1 in the ability of microgametes to exflagellate in P. berghei and found that VS1 had no inhibitory effect on exflagellation between pre- and post-injection mice (Table S2). Furthermore, dissection of mosquitoes 24 hours after feeding with either PBS or VS1 showed presence of ookinetes in the blood meal (data not shown). Nevertheless, more studies are needed to more thoroughly assess the effect of VS1 on macrogametogenesis, fertilization, and ookinete development. Because we are working with a T-B compound, targeting more than one stage of the parasite in the mosquito gut would strengthen the outcome of our final goal, completely blocking transmission of malaria.
We used a candidate gene approach to identify potential targets or binding ligand(s) of VS1 among the repertoire of Plasmodium ookinete micronemal proteins [30], focusing particularly on those with established roles in midgut attachment or invasion. The literature indicates that two such proteins, CTRP and WARP, bind to sulfated GAGs [10]; and although both are in the apicomplexan TRAP/MIC2 family of proteins, their domain architectures are quite different [30]. WARP is an approximately 40 KDa protein with a signal peptide and a single vWA domain, and we hypothesize that based on the published data WARP is secreted from the ookinete microneme and can thus work as an extracellular adaptor protein, potentially bridging the parasite surface (or surface molecules) with midgut apical membrane ligands. The much larger CTRP is approximately 230 KDa and contains a signal peptide followed by six contiguous vWA domains, seven contiguous thrombospondin (TS) domains, a transmembrane domain, and a short acidic cytoplasmic domain at the C-terminus that interacts with the motility actomyosin machinery [30]. The first four vWA domains of CTRP are more similar to one another than to vWA domains 5 and 6 when comparing six species of Plasmodium. Interestingly, a phylogenetic analysis of the vWA domains from TRAP, CTRP, and WARP among these species shows that WARP and CTRP form a single clade and that the vWA domain of WARP most recently shared a common ancestor with the fifth vWA domain of CTRP, suggesting that WARP evolved from CTRP [30]. We emphasize that the domain architectures and amino acid sequences of these two proteins are highly conserved across Plasmodia [10], [30] and argue that VS1's potency against both rodent and human malaria, as well as its ookinete staining pattern as reported here, suggests that VS1's binding partner(s) is likewise highly conserved across Plasmodia. Thus given the above and the aim of identifying the mechanism of action for VS1, we sought to investigate the binding activity of Plasmodium vivax CTRP and WARP to VS1, following the argument that VS1 should bind to these two molecules. We produced soluble recombinant WARP and the first vWA domain of CTRP from P. vivax using a cell-free wheat germ system (Figure S5) and evaluated binding affinity via ELISA. As expected, both recombinant PvCTRP and PvWARP bound to VS1 in a dose-dependent manner (Figure 4C). To better delineate binding specificity, competition assays with heparin and chondroitin sulfate A (CSA) were performed. If VS1 binds primarily to the putative GAG-binding sites on the vWA domains of CTRP and WARP, then we would expect that heparin, and perhaps to a lesser extent CSA, at a concentration of 100 µg/ml should completely inhibit binding. However, we observed that both heparin and CSA only partially inhibited VS1 binding to PvCTRP and PvWARP (Figure 4C), suggesting that VS1 binds to additional sites not used by heparin on either recombinant protein or that it binds to them with greater affinity.
The ELISA results demonstrated that VS1 binds to recombinant WARP and the first vWA domain of CTRP in vitro, so to test binding in vivo we obtained three lines of P. berghei in which CTRP had been either completely [9], [15] or partially knocked out [15] and compared VS1 staining patterns by immunofluorescence microscopy. One of the partial knockouts, a line known as ΔA6, expresses CTRP that is missing all six of the vWA domains but contains all seven of the thrombospondin domains. Conversely, CTRP expressed by the other partial knockout line, ΔTS7, includes the six vWA domains but lacks any of the thrombospondin domains. Since the recombinant CTRP protein used in the ELISA consisted only of the first vWA domain, we predicted a priori that the VS1 binding pattern to ΔTS7 ookinetes would be similar to wild type ookinetes, while the VS1 signal in both the CTRP knockout (CTRPKO) and ΔA6 lines would be diminished. However, if VS1 also binds to any of the TS domains in vivo, the VS1 signal would be much lower in the CTRPKO than in either of the partial knockouts. A caveat to this approach is that if VS1 also binds to WARP in vivo, we would expect some portion of the VS1 signal observed in wild type ookinetes to be shared among all of the knockout lines.
Immunofluorescence microscopy images from CTRPKO (Figure 4D) indicate that VS1 binds to CTRP in permeabilized ookinetes. Furthermore, comparisons of staining patterns from the partial knockouts strongly suggest that VS1 binding involves vWA domains but not the TS domains (Figure 4E, F). Furthermore, the apparent loss of VS1 signal in both the CTRPKO and ΔA6 lines also suggests that VS1 localizes to the micronemes (as suggested by Figure 4B) and that CTRP is the primary micronemal target of VS1and not WARP (Figure 4D, F). These data further reconcile the observed cytoplasmic staining of putative zygotes/macrogametes (Figure S3) with the previously reported staining of round forms with CTRP antisera [11]. Without a WARP knockout line we cannot rule out that binding to WARP may occur in vivo or that some of the T-B activity we observed is due to such an interaction. It should be noted that WARP expression/secretion is not well understood and may be temporally regulated or even midgut contact-dependent for different Plasmodium species. Thus, in vitro generated P. berghei ookinetes or P. falciparum ookinetes isolated from the blood-meal bolus may not express detectable levels of WARP. However, the CTRPKO and ΔA6 microscopy data are persuasive and we note that VS1 binding by ELISA is consistently stronger for the first vWA domain of CTRP than for WARP. Moreover, the phylogenetic relationship among CTRP and WARP vWA domains [30] in combination with the ELISA and immunofluorescence microscopy data reported here, suggest that VS1 primarily targets the first four vWA domains of CTRP. The divergence of vWA domains 5 and 6 and their evolutionary relationships with WARP suggest that these domains bind VS1 secondarily or not at all.
In the absence of a CTRP crystal structure, we used homology modeling to predict heparin-binding sites on CTRP (Figure 5A–E). The quality of the models was assessed with QMEAN; and models 1 (PDB: 1AUQ, Figure 5B, C) and 2 (PDB: 2XGG, Figure 5 D, E) had Qmean scores of 0.572 (Z-score = −2.892) and 0.584 (Z-score = −2.314), respectively, indicating comparable model reliability. Model 1 is based on the human von Willebrand factor A1 domain, while Model 2 is based on the vWA-integrin like domain from the Toxoplasma gondii MIC2 protein. The structure of the former has been extensively studied due to its essential role in platelet adhesion [31]–[33], while the latter is a well-described adhesin involved in host-cell invasion [34]. Despite the selection of markedly different templates, both with low sequence identity to CTRP (<25%), both models had the same overall α/β Rossmann fold. The positively charged residues adopted similar patterns between the two models (Figure 5 B–E), which were also in general agreement with the predicted heparin-binding domains on vWFA1 (Figure 5A). A superposition of Model 1 with vWFA1 (1AUQ) (Figure S6A–B) suggests that the two models predominantly differed in their overall length and the conformation of the loops connecting the beta sheet core and flanking alpha helices. A superposition of Models 1 and 2 also demonstrated a difference in overall length as well as the presence of a large alpha helix in Model 1 that is absent in Model 2 (Figure S6C–D). Furthermore, a number of basic residues fall well outside the predicted vWFA1 heparin-binding regions in CTRP (Figures 5A–E), which may represent an extended electropositive surface and additional binding sites for sulfated polymers such as heparin and VS1.
Host-cell GAGs have been shown to be important mediators of Plasmodium development in its two hosts, including merozoite invasion of RBCs [35]–[36], infected RBC sequestration to placenta [37], ookinete invasion of the midgut [5], and sporozoite invasion of mosquito salivary glands [38] and vertebrate hepatocytes [39]. Here we tested a strategy that exploits this feature of Plasmodium biology and demonstrated that VS1, a putative GAG-mimetic, reduced midgut oocyst development by as much as 99% in mosquitoes fed with P. falciparum or Plasmodium berghei. Through direct-binding assays, we observed that VS1 bound to two ookinete micronemal proteins necessary for midgut invasion, each containing at least one vWA domain: (i) CTRP and (ii) WARP. By immunofluorescence microscopy, we observed that VS1 stains permeabilized P. falciparum and P. berghei ookinetes but does not stain P. berghei CTRP knockouts or transgenic parasites lacking the vWA domains of CTRP while retaining the thrombospondin repeat region. Finally, we used structural homology models of the first vWA domain of CTRP to identify residues likely involved in binding GAGs, as well as the VS1 compound. Based on these data, our working model for the mechanism underlying VS1's T-B activity is that it binds to CTRP once the protein is secreted from the micronemes of ookinetes prior to midgut attachment and invasion. CTRP is essential for gliding motility [9], [15] and contains six vWA domains, which commonly play roles in cell adhesion to GAGs [8], [10], [30]–[34]. Our data suggest that VS1 either interrupts the gliding process on the midgut apical microvillar surface or coats the surface of the ookinete through its interaction with CTRP, thus preventing attachment to ligands (e.g., chondroitin sulfate [5]) on the apical surface of the midgut epithelium. Nevertheless, due to observed binding of VS1 to permeabilized round cells, we cannot rule out that VS1 may potentially have an added benefit and affect additional parasite stages found in the blood meal, particularly macrogametes and/or zygotes. Further studies into these beneficial side effects are necessary.
When designing the GAG-mimetic strategy, data from the literature suggested that sulfation density per disaccharide unit and the manner of presentation (i.e., how the underlying structure of the sugar scaffold influences the 3D projection of sulfated moieties) are critical factors in inhibiting pathogen-GAG interactions. Boyle et al. [35], for example, found that heparin and the E. coli-derived K5 polysaccharide inhibits merozoite entry into RBCs and that variations in the average number of sulfate groups/saccharide unit for K5, which consists of glucoronate as opposed to iduronate, exhibited different inhibitory effects against merozoites, with sulfate densities >3/disaccharide producing the most potent IC50 estimates. Therefore, we sought to determine the minimal T-B polymer length of VS1 with the hope of minimizing the likelihood of diverse structural conformations that can occur with longer polymers, which could in turn affect the presentation/projection of anionic moieties. Although we were able to demonstrate that VS1-3,000 was the most effective polymer, we cannot, however, predict its structure. VS1-3,000 is unlikely to remain linear in solution or in the midgut after blood feeding. With this caveat in mind, we suspect that binding to the recombinant or native CTRP and WARP molecules may engender a specific VS1 conformation. Regardless, we expect that VS1 binding is largely due to the predicted GAG-binding motifs on the vWA domain(s) of CTRP and WARP [10]. However, the concentration of heparin used in our studies, which would otherwise result in the near complete inhibition of high affinity protein-GAG interactions [40]–[42] only reduced VS1 binding by ∼25%. It should be noted, however, that cases exist in the literature where soluble heparin cannot completely outcompete vWA domain-GAG interactions. For example, heparin-BSA binding to the vWA domain of PfTRAP can be competed between 45–66% using 50 µg/ml of soluble heparin and that a 10-fold increase in heparin concentration reduced binding to 9–27% of control [43]. Even more striking is a report that neither a 50-, 100-, or 1,000-fold molar excess of soluble heparin could completely inhibit binding between the PfTRAP vWA domain and the surface of HepG2 cells, which was thought to be GAG mediated [44]. In this set of experiments, each concentration reduced binding by approximately 15%, 55%, and 70%, respectively, suggesting that the PfTRAP vWA domain utilizes both GAG and a non-GAG binding sites. In combination, these data suggest that each recombinant protein in our study has either stronger binding affinity for VS1 than for either heparin or CSA, that the predicted GAG-binding regions do not completely explain the interactions of VS1 with PvCTRP or PvWARP, or more likely, a combination of these two scenarios. Tertiary structures of mammalian heparin-binding proteins have also been shown to enhance affinity and specificity [41]. We cannot rule out the possibility of cryptic GAG-binding sites on CTRP and WARP that provide cooperative binding to VS1, as suggested at least in part by potential basic residue patches identified on two homology models of the first vWA domain of CTRP, which appears to be a primary ligand of VS1. In terms of sulfation density and propensity to form various non-linear conformations, VS1 is clearly different from Heparin and natural GAGs. In this context, cooperative binding may be conferred by VS1 “wrapping around” CTRP and interacting with basic residues along different faces of the protein. The presence of potential additional binding sites suggests that CTRP (as opposed to other GAG binding proteins) can be specifically targeted by the next generation of VS1-based chemical mimetics. Clearly, a crystal structure for CTRP is needed to clarify the hypotheses generated by our two models.
To date, the antimalarial pipeline is filled with compounds that act on related biochemical pathways (e.g., folate biosynthesis), which also increase the likelihood of the development of parasite cross-resistance to these “new” compounds. The need to discover drugs that act on unpredicted or uncharacterized biochemical pathways that are completely different from those associated with current antimalarials is paramount [45]. Our approach fits this mold, as it represents a completely novel mechanism of action compared to those associated with the existing, new, and now “rediscovered” list of antimalarials and T-B compounds [7].
Among the various T-B strategies, drugs offer a distinct advantage over vaccines since the efficacy of the compound is dose dependent and human immune-system independent, the latter being a potentially significant issue given that individuals in malaria endemic regions may suffer from malnourishment and concomitant infections by immune-modulating pathogens such as HIV and helminths. Although we have shown that VS1 is a potent T-B molecule, we emphasize that it cannot be used as a drug in its current form. However, we intend to use the data reported here to establish a high-throughput approach for identifying a next-generation “druggable” malaria T-B compound that would inhibit ookinete invasion of the midgut beyond that observed for VS1 (i.e., achieve zero infection prevalence among treated mosquitoes). We recognize, however, that if the next generation compound only replicates the T-B activity reported for VS1 and were used alone in the field, it would unlikely reduce infection prevalence in the mosquito population below the threshold necessary for sustained transmission . Nevertheless, it is widely believed that no anti-malarial intervention on its own will lead to regional elimination and eventual eradication [1], [2]. We envision that in this context, such compounds may be valuable in a range of epidemiologic settings. Potential applications include (i) general use in conjunction with existing artemisinin combination therapies, which we emphasize do not kill stage V gametocytes, to prevent recurrent transmission from the treatment-seeking segment of the population, (ii) use in regions with unstable malaria (e.g., highlands) to curb transmission during epidemics, (iii) use in combination with a T-B vaccine targeting sexual stage parasites to act as a safety net to “mop up” break-through parasites, and (iv) at the end game of the malaria eradication effort, as mass distribution of T-B compounds may offer a cost-effective approach to preventing asymptomatic, gametocytemic individuals, who would not otherwise seek treatment, from infecting anopheline mosquitoes, thus preventing resurrection of epidemic malaria transmission.
|
10.1371/journal.pcbi.1004844 | A Knowledge-Based System for Display and Prediction of O-Glycosylation Network Behaviour in Response to Enzyme Knockouts | O-linked glycosylation is an important post-translational modification of mucin-type protein, changes to which are important biomarkers of cancer. For this study of the enzymes of O-glycosylation, we developed a shorthand notation for representing GalNAc-linked oligosaccharides, a method for their graphical interpretation, and a pattern-matching algorithm that generates networks of enzyme-catalysed reactions. Software for generating glycans from the enzyme activities is presented, and is also available online. The degree distributions of the resulting enzyme-reaction networks were found to be Poisson in nature. Simple graph-theoretic measures were used to characterise the resulting reaction networks. From a study of in-silico single-enzyme knockouts of each of 25 enzymes known to be involved in mucin O-glycan biosynthesis, six of them, β-1,4-galactosyltransferase (β4Gal-T4), four glycosyltransferases and one sulfotransferase, play the dominant role in determining O-glycan heterogeneity. In the absence of β4Gal-T4, all Lewis X, sialyl-Lewis X, Lewis Y and Sda/Cad glycoforms were eliminated, in contrast to knockouts of the N-acetylglucosaminyltransferases, which did not affect the relative abundances of O-glycans expressing these epitopes. A set of 244 experimentally determined mucin-type O-glycans obtained from the literature was used to validate the method, which was able to predict up to 98% of the most common structures obtained from human and engineered CHO cell glycoforms.
| Our objective being to model the enzymes of mucin-type O-linked glycosylation, we first developed a model language to represent O-glycan structures succinctly in linear string form, to which a set of pattern-matching rules was then applied to simulate the activities of a set of 25 glycosyltransferase and sulfotransferase enzymes. The modelling language (a formal language), together with the set of transformation rules representing the enzymes of the model. comprise the deductive apparatus of a formal system. The system, implemented in software, was able to predict a highly heterogeneous set of structures when all enzymes were allowed to act, including many clinically important epitopes such as sialyl-Lewis X. We studied the effects of single-enzyme knockouts on the properties of the resulting enzyme-catalysed reaction networks and determined the enzymes most likely to be responsible for heterogeneity.
| Glycosylation is a major post-translational modification of proteins, in which a carbohydrate moiety, called a glycan, is covalently attached to an amino acid of the polypeptide, to form a glycoprotein [1]. N-linked glycans are attached to an asparagine (N) residue appearing in the consensus sequence Asn-X-Ser/Thr, where X is not Pro, while O-linked glycans are attached to the hydroxyl group of a serine or threonine residue. A study of potential glycosylation sites indicated that three quarters of proteins may be glycosylated, with about 10% of these O-glycosylated [2]. Glycans are formed by the sequential addition of monosaccharides from nucleotide-sugar donors to the glycoprotein acceptor, a process that is catalysed by glycosyltransferase enzymes, which are located in the endoplasmic reticulum and Golgi apparatus.
Mucins are a class of large glycoproteins that contain a large number of Ser/Thr in close proximity, which can be heavily O-glycosylated. The initial step of mucin-type glycosylation is the attachment of a GalNAc (N-acetyl-d-galactosamine) to an unoccupied Ser/Thr on the protein acceptor. Modification of mucin O-glycosylation is an important biomarker in cancer detection [3–8]. In the innate immune response, cell-cell recognition is dependent on the expression of a number of different carbohydrate epitopes on carrier proteins, which include both sulfated and non-sulfated versions of Lewis X (Lex), Lewis A (Lea), Lewis B (Leb) [9] and, more rarely, Lewis Y (Ley) antigens [10].
Of the several theoretical treatments of glycosylation which have now appeared, most have considered N-glycosylation rather than O-glycosylation [11]. The method of Kawano et al. [12] for predicting glycan structures from gene expression data was able to predict the appearance of a variety of glycosylated structures, including O-linked. The model by Gerken and co-workers focused on the initiation of O-glycosylation [13]. Liu et al. [14] described an object-oriented method of construction of networks of O-glycan biosynthesis that was used to predict levels of sialyl-Lewis X (SLex), an important antigenic determinant, and more recently a computational approach based on MATLAB has been used to predict pathways of N- and O-linked glycosylation [15, 16]. In the present work, we have taken an alternative, bottom-up, approach to modelling the de novo biosynthesis of mucin O-glycans. In order to facilitate computational analysis, we introduce a formal language (see [17]) for identifying individual glycan structures, a method for representing glycans graphically, based on these identifiers, and describe a method for generating networks of reactions based on the activities of enzymes involved in mucin protein O-glycosylation. A mathematical model of N-linked glycosylation has been developed, [18] whose structure identifiers are based on Linear Code; Spahn et al. have developed a Markov-chain model based on this system. [19]. As it seeks to uncover the nature of the reaction networks of O-glycosylation, this work both validates and extends the approach used by these earlier studies.
With a rapidly increasing number of studies employing nuclease-based genome-editing technologies, such as zinc-finger nuclease (ZFN) [20] and CRISPR/Cas9 [21], for biotechnological applications, it is important to consider the possible phenotypic effects that may result from knock-ins or knockouts of the glycosyltransferase genes, and the corresponding changes to the glycome. We anticipate that the methods we describe here will be of use in predicting such changes within the context of O-glycosylation networks.
A study of the GalNAc-linked oligosaccharides within the online repository of the Consortium for Functional Glycomics [22] revealed the five most commonly occurring monosaccharides to be d-galactose (Gal), N-acetylgalactosamine (GalNAc), N-acetylglucosamine (GlcNAc), l-fucose (Fuc) and N-acetylneuraminic acid (Neu5Ac). The five most commonly encountered sugars were: Gal (32.3%), GalNAc (22.7%), GlcNAc (20.7%), Fuc (11.2%) and Neu5Ac (9.6%). Four residues, which included N-glycolylneuraminic acid (Neu5Gc) and 2-keto-3-deoxy-d-glycero-d-galacto-nononic acid (Kdn), made up the remaining 4% of the total monosaccharide composition. Methylated and sulfated variants were included in the analysis.
At the time of writing, 1654 transferases are listed in the IUBMB Enzyme Nomenclature, of which 280 involve the transfer of a monosaccharide from a nucleotide-sugar donor to an acceptor. An examination of the latter subset of reactions reveals that the class of monosaccharides employed is quite small, with over 90% of the glycosyltransferase reactions involving only 8 distinct sugar species, Fuc, Gal, GlcA, GalNAc, Glc, GlcNAc, Neu5Ac and Xyl. Combined with the result of the analysis of the CFG database, this suggested that the language of O-glycosylation has a limited alphabet, though with a potentially rich vocabulary. A formal language was developed that uses a single-letter code for the five most commonly encountered monosaccharides, with uppercase letters for d-sugars and lowercase for the less common l isomers. The symbols of the language and their meanings are summarised in Table 1.
The strings generated by the language, which we refer to as structure identifiers, are a further contraction of the short-form, one-line representation of oligosaccharides [23], in which the IUPAC sugar symbols are replaced by one-letter codes, and brackets instead of parentheses are used as branch delimiters. An example O-glycan is shown in Fig 1.
We identified 25 distinct enzyme activities in which these common monosaccharides are transferred during GalNAc-linked glycosylation, which are shown in Table 2. The O-glycan structure indentifiers enable us to write the reactions catalysed by these enzymes more succinctly. For instance, the ST3Gal-I reaction, CMP-N-acetylneuraminate + N-acetyl-α-neuraminyl-(2 → 3)-β-d-galactosyl-(1 → 3)-N-acetyl-d-galactosaminyl-R = CMP + N-acetyl-α-neuraminyl-(2 → 3)-β-d-galactosyl-(1 → 3)-[N-acetyl-α-neuraminyl-(2 → 6)]-N-acetyl-d-galactosaminyl-R can be represented in the current notation as
CMP - S + [ Sa 3 Lb 3 ] VT = CMP + [ Sa 6 ] [ Sa 3 Lb 3 ] VT
where CMP-S is the donor and [Sa3Lb3]VT is the acceptor. Table 2 shows the enzyme reactions using a shorthand form based on the formal language. For simplicity, the stereochemical information (a/b) will be omitted within the text, based on the known specificities of the enzymes. For the enzymes considered in this model, all of the fucosyltransferases and sialyltransferases produce α-linked structures. The galactosyltransferases and N-acetylglucosaminyltransferases will be assumed to form β-linked products, unless indicated otherwise, while N-acetylgalactosaminyltransferases will be assumed to form α products. Hence, without ambiguity, we can rewrite the reaction equation above as
CMP - S + [ S 3 L 3 ] VT = CMP + [ S 6 ] [ S 3 L 3 ] VT
A consequence of the formal grammar is that any residue added to the base GalNAc is treated as a branch. Therefore [L3]VT is written instead of L3VT, and [S6][S3L3]VT instead of S3L3[S6]VT. While we could write [Y3[Y6]L4Y3]VT to represent GlcNAcβ1-3(GlcNAcβ1-6)Galβ1-4GlcNAcβ1-3GalNAc, by convention we will write such structures as [[Y6][Y3]L4Y3]VT, even though both are valid according to the grammar. Branches at the same level are written from right to left in ascending linkage order, as shown in Table 2.
We introduce a formal grammar [24], Γ = (ΣN, ΣT, P, S), where ΣN is a set of nonterminal symbols and ΣT is a set of terminal symbols. ΣN and ΣT are disjoint sets, meaning that they share no members in common. S defines a starting symbol and P is a set of production rules, each element of which maps a single non-terminal symbol to a string of one or more symbols drawn from ΣT∪ΣN, or to the null (empty) string, ϵ.
Σ N = { Z , A , B , C , m , d , l } Σ T = { 2 , 3 , 4 , 6 , 8 , a , b , f , s , K , L , N , S , T , V , Y , [ , ] } P = Z → A T A → ϵ | B B V B → ϵ | [ C m l d ] C → ϵ | C m l d | C [ C m l d ] m → f | s | K | L | N | S | V | Y d → 2 | 3 | 4 | 6 | 8 l → ϵ | a | b S = Z
The grammar generates a language L by the successive substitution of nonterminal symbols with the right-hand sides of production rules in P. The set ΣT ∪ ΣN is the alphabet of L, and strings of symbols generated by Γ are the words of the language. We define a structure identifier as a word of L that contains only symbols drawn from ΣT.
The following sequence of strings serves as an example of a derivation within the grammar. For brevity, some steps are the result of several simultaneous applications of production rules.
Z A T { Z → A T } B B VT { A → B B V } [ C m l d ][ C m l d ] VT { B → [ C m l d ] } [ C m l d ][ m l d m l d ] VT { C → C m l d , C → ϵ } [ m l d ][ m l d m l d ] VT { C → ϵ } [S l d ][ S l d L l d ] VT { m → S , m → L } [S6][ S 3 L 3 ] VT { d → 6 , d → 3 , l → ϵ }
The final string in the list is a word in Γ denoting disialylated T antigen, commonly known as “diST”, a core-1 O-glycan.
The linear string identifiers described in this work can be used to draw glycan structures in the manner of turtle graphics [26]. Reading the identifier from right to left, the drawing method acts according to the current symbol: if the symbol is an element of the set {f,K,L,N,S,V,Y,s}, it draws the symbol corresponding to the monosaccharide at the current drawing position; if the string character is a right bracket,], the current position and orientation information are pushed onto a stack, and are popped from the stack on meeting a left bracket. A two-pass approach is employed, with the bond framework being drawn on the first pass, and the sugar symbols drawn on the second.
A suite of Perl scripts was written for the generation of structure identifiers by enzyme simulation, for parsing, and rendering as Scalable Vector Graphics (SVG) image files. A library of functions was written as a Perl module, which enabled (i) the translation of structure identifiers to and from the IUPAC condensed-form one-line notation; (ii) identification of common epitopes, such as Lex, based on regular-expression patterns; (iii) parsing of O-glycan strings by an LL(1) parser based on a simplified version of Γ; (iv) rendering of string identifiers as SVG, in either UOXF or CFG styles.
Not all of the structures encoded by the formal grammar are feasible, in that structures such as [S3][L3]VT are syntactically correct, but chemically impossible, since it describes a sialic acid (S) and galactose (L) both 3-linked to the same N-acetylgalactosamine (V). In order to generate a set of biologically relevant O-glycans, therefore, a set of regular-expression based substitution rules was developed to mimic the actions of each of the enzymes shown in Table 2; throughout this work, numbers in bold face refer to the corresponding activities in this table. The rules were incorporated into a Perl script, which took a single O-glycan identifier as the initial substrate, and applied each of the substitutions in turn to output a set of products. The initial structure defaulted to the non-glycosylated site, ‘T’, but any valid glycan structure could be supplied by the user as a starting point. The process was applied iteratively, such that each new product formed was presented as a substrate to every enzyme upon the next iteration. Where an enzyme rule could match at more than one position, as in the case of diantennary O-glycans, the identifier was split, using the current regular expression, and then each part substituted according to the same rule, before reassembling the parts, with the new string being added to the pool of possible products. Branching level and extension by poly-N-acetyllactosamine repeating units could be controlled by placing an optional limit on the total number of GlcNAc residues incorporated. Restrictions could be placed on individual enzyme activities by conditionals employing Boolean logic. The program could also be limited to use a subset of the enzymes. Simulations terminated after a prescribed number of iterations, or after any iteration in which no new products had been generated. The output of the program for three iterations of the method is shown in Fig 2. A web-application front end to the enzyme simulator (see Methods) is available online at http://www.boxer.tcd.ie/glycologue.
The enzymes of Table 2 can be divided into five main classes of activity: initiation (2), core formation (5,6,8,9), branching and extension (1,7,10,12,19), sugar modification (20–22) and termination (3,4,11,13–18,23–25). The terminal residue of an oligosaccharide is the monosaccharide appearing at its non-reducing end. In the current model, the two methods of termination were fucosylation or sialylation of a terminal galactose. Sulfation was the only type of non-glycosyltransferase modification that was considered. Oligosaccharide chains can be of type 1 (ending in Galβ1-3GlcNAc-) or type 2 (ending in Galβ1-4GlcNAc-).
The enzyme rules were reversed, so that a single monosaccharide was removed at each step of the simulation. Any O-glycan structure supplied as an initial substrate to the reversed enzyme simulator was considered to be predictable, or deducible, within the system if its final step was the removal of the terminal GalNAc from the protein by the enzyme ppGalNAc-T, according to VT -- ppGalNAc-Ts --> T. If the simulation ended with no new products formed, and without reaching the non-glycosylated site, the glycan was considered non-predictable within the system.
The reaction data provided by the method described earlier, and depicted in Fig 2, were used to generate network graphs in GraphViz (www.graphviz.org), with O-glycan identifiers as nodes and with edges representing enzyme-catalysed reactions, colour-coded according to the monosaccharide being transferred. The enzyme simulator also allowed enzymes to be knocked out in silico, either individually or in groups, with each knockout resulting in a different reaction network. A web application, O-Glycologue (see Methods) was developed in order to view the structures obtained for a particular set of knockouts, and compare them with the structures obtained for the “wild-type” network, defined as the network obtained with all 25 of the enzymes active. The method is illustrated with an example taken from a study on salivary MUC7 glycans [45], a triantennary core-2 structure with the structure identifier [S3L4[f3][s6]Y6][[S3L4[f3][s6]Y6][S3L4[f3][s6]Y3]L3]VT (Fig 4A). The reversed reaction network is shown in Fig 4B, which successfully removed all monosaccharides in 17 iterations using the nine enzyme activities 1, 2, 5–7, 11, 16, 19 and 20. The network of reactions produced when the enzyme simulator was run in the forward direction with only these enzymes active is shown in Fig 4C.
With all 25 enzyme activities enabled, 18 iterations of the method generated 13,127,561 unique O-glycans, in 34,215,049 reactions. All structure identifiers generated by the enzyme simulations were shown to be valid according to the parser. Different epitopes could be determined from the terminal sequences of the identifier string, and were counted as percentages of the total number of glycans formed: Lewis A ([L3[f4]Y, 13.2%), Lewis X ([L4[f3]Y, 25.0%), sialyl-Lewis A ([S3L3[f4]Y, 4.2%), sialyl-Lewis X ([S3L4[f3]Y, 8.4%), Lewis B ([[f2]L3[f4]Y, 4.3%), Lewis Y ([[f2]L4[f3]Y, 8.2%), H antigen ([[f2]L3Y, 9.4%), A ([V3[f2]L3[f4]Y, 1.9%), B ([La3[f2]L, 17.5%), Sda/Cad ([S3[Vb4]L, 12.7%) and other (24.7%).
Depending on the degree of branching, several different epitopes could appear together on the same O-glycan. Overall, 227 different pattern combinations of recognised epitopes could be distinguished, such as Lewis A with the H antigen.
As a consequence of the method used to produce the network, in which the products at iteration n + 1 are dependent only upon those arising from iteration n, the growth function can be approximated by a discrete logistic map, ν(n + 1) = bν(n), b > 1, with solution ν(n) = abn. Although the total population is therefore expected to grow exponentially, by setting a limit on the maximum number of GlcNAc residues incorporated in each O-glycan, it was possible to close the networks, so that eventually no further structures were added to the glycan pool (Fig 5B).
Under the assumption of irreversibility of each reaction, the network can be viewed as a rooted, directed acyclic graph G = (V,E), where V and E are sets of nodes and edges, respectively, with each node representing a distinct O-glycan and edges representing enzyme-catalysed reactions in which O-glycans appear as substrates or products. The degree of a node is defined as the number of its immediate neighbours to which it is connected by an edge. For a directed graph, the number of incoming nodes is called the in-degree, and the number of outgoing nodes is defined as the out-degree. An important network property is the degree distribution, which is frequently expressed in terms of the probability, P(k), that a randomly selected node will be of degree k. Many real networks possess the property of hierachical clustering of nodes [46] with a degree distribution that displays a power-law tail, P(k)∼k − λ. In contrast, our reaction network displayed a Poisson-like distribution that is characteristic of random networks [47]. After 14 iterations, the average degree of the network, 〈k〉, was calculated to be 4.36, with the in-degree and out-degree averages each equal, at half of this value. A bilog plot of the degree-distribution of the network (node degree frequency vs degree) is non-linear, as shown in Fig 5C, indicating that the network is not self-similar [48], or scale-invariant. That the degree distribution of a reaction network arising from a fully deterministic system has the characteristics of a random network may be a natural outcome of the method that was used to generate the glycan structure libraries. Since this method is essentially combinatoric, in that every possible acceptor-product is discovered from every substrate, we conjecture that its degree distribution can be described by a binomial function. Newman et al.[49] have shown that networks with a binomial degree distribution become Poisson when the number of nodes is large.
Quantitative measures of the connectedness of the reaction network are provided by the α, β and γ indices [50]. The β index is the ratio of the number of edges, e, to the number of nodes, v:
β = e v (1)
The definitions of the non-planar versions of the α and γ indices, which allow for edges to cross at non-nodal positions in the plane, are
α = ( e - v ) v ( v - 1 ) / 2 - ( v - 1 ) (2)
and
γ = 2 e v ( v - 1 ) . (3)
The α index represents the number of cycles in a graph to the maximum number of possible cycles, and will take a value between 0 and 1. The γ index is the ratio of the number of edges to the total number of edges in the fully connected network, v(v − 1). Local clustering coefficients were also computed, and averaged across the complete reaction network [51]. The clustering coefficient, Ci, provides a measure of the fractional degree to which nearest neighbours of node i are connected to each other. Let ki be the number of immediate neighbours of node i. Since there can be at most ki(ki − 1) edges between ki nodes, for a directed graph, the value of Ci is defined as
C i = E i k i ( k i - 1 ) (4)
where Ei is the number of existing edges between the neighbours of node i. An average network clustering coefficient, 〈C〉, was defined over the whole reaction network. The values of β and 〈C〉, which were calculated at each iteration of the enzyme simulation, showed an increase overall, monotonically above the iteration 7, while the non-planar γ index decayed uniformly from unity (Fig 5D). The increase in β index approximated to linearity above iteration 8.
We simulated the effects of knocking out individual enzymes, observing the changes incurred in the topology of this reaction network. O-Glycan heterogeneity was most strongly influenced by the activities of Gcnt2, C2/4Gn-T, β3Gn-T2/3/4/5/7, β3Gn-T6 and β4Gal-T4, as quantified by the changes in the indices in Fig 6A–6C. Changes to local clustering coefficients were also noticeable, although they were not as marked. In the absence of enzyme β3Gn-T2/3/4/5/7 (10), the network closed after 20 iterations, and in the absence of β4Gal-T4 (1), the network was closed after 14 iterations, since no further extension of antennae was possible in the absence of either of these activities. Changes to the α and γ indices were notable only for these two enzymes (Fig 6B).
Changes to the distributions of common epitopes are given in Table 3. The occurrences of each epitope, expressed as a percentage of the total number of unique O-glycans, were obtained for 12-iteration networks with the enzyme knocked out as indicated, and from which the sulfotransferases (20–22) had been omitted. Excluded from the results are ppGalNAc-Ts and the knockouts of the sialyltransferases 17 and 18, which showed no alteration from “wild type” (wt). Since more than one epitope can be expressed on a single O-glycan, the numbers on each line in the table need not sum to 100. The β4Gal-T4 knockout was found to eliminate all glycans expressing Lex, SLex, Ley and Sda antigens, indicating that it is an essential component of their biosynthesis; an increase in the percentage of O-glycans bearing the B antigen was also observed. The greatest decrease in the total number of glycans formed was observed with this knockout (not shown). Single-enzyme knockouts of the N-acetylglucosaminyltransferases did not affect the distributions of these epitopes so markedly, as might be expected from their functions in core formation, elongation and branching, rather than termination. Knocking out the β-1,3-galactosyltransferase activity eliminated only O-glycans expressing the B antigen.
The predictive power of the enzyme simulator was tested by comparing the in-silico generated O-glycans against fifteen published collections of such structures that had been identified experimentally: mucin O-glycans from human colon [52, 53]; structures of MUC1 mucin glycoforms obtained from normal and cancerous breast epithelial cell lines [54]; poly-N-acetyllactosamine extended structures of leukosialin glycoprotein obtained from promyelocytic and myelogenous leukaemia cell lines [55]; leukosialin O-glycans expressed in T-lymphocytic leukemia [56] and erythroid, myeloid, and T-lymphoid cell lines [57]; O-glycans from salivary MUC7, a major component of mucin glycoprotein 2 (MG2) [45]; O-glycans of Tamm-Horsfall glycoprotein [58]; sulfated core-2 and core-4 oligosaccharides obtained from mucins associated with chronic bronchitis [59]; bovine serum fetuin, human serum IgA1 and secretory IgA, human neutrophil gelatinase B and glycophorin A O-glycans [60]; extended core-1 and core-2 O-glycans from Chinese hamster ovary (CHO) cells transfected with β3Gn-T3 [61]; MUC1 and MUC4 O-glycans from bovine and human milk [62], normal human serum [63] and a human gastric adenocarcinoma cell line (MKN45) [64]; mucin from normal descending colon [65]; recombinant mucins from engineered CHO cells [66]. In all, 244 unique O-glycan structures were collected from these studies and assigned structure identifiers. Multiple identifiers were assigned where a number of different configurations was possible. For example, the monosialylated forms of Galβ1-3(Galβ1-4GlcNAcβ1-6)GalNAc-R [64] were represented by the separate identifiers [L4Y6][S3L3]VT and [S3L4Y6][L3]VT.
Each member of the set of experimentally determined O-glycans was supplied to the reversed enzyme simulator as the starting substrate, and tested for predictability within the system. Overall, 87% of the unique O-glycan structures were predicted by the method, which was able to reproduce any of the extended branched core 1–4 structures, with sialyl-Lewis X, Lewis Y, Lewis A or -B terminals and their 3′- and 6-sulfated variants. Table 4 lists the O-glycans determined experimentally that appeared in more than one of the studies, and thus independently verified, in descending order of frequency. Shown are the structure identifier, the supporting literature and a check next to those structures that were predicted in silico. Of the 45 oligosaccharides most commonly occurring, 44 were predicted by the model, giving a coverage of 98%.
From analysis of the grammar, and the results of the enzyme simulations, we predict that a highly heterogeneous population of mucin O-glycans is likely to result if even a limited subset of the enzyme activities of Table 2 is expressed. In-silico enzyme knockouts have identified β4Gal-T4 as a key regulator of the complexity of O-glycosylation networks, in keeping with our earlier observations on the influence of this enzyme on N-linked glycosylation in engineered Chinese hamster ovary cells [67].
The number of iterations was chosen according to the type of in-silico experiment: trends in the changes to the indices were discernable by iteration 15, hence this value was chosen for the enzyme-knockout studies; 18 is the maximum number of iterations of the basic model that were possible within the available memory (32 GB), with all 25 enzymes active and no limitations placed on the number of GlcNAcs. Not all of the enzymes in the current model will be present in all species, or active at all times. The full network is therefore a chimeric construct, but one which could be tailored for specific cases as needed, by considering only the enzymes known to be expressed in a particular organism or tissue. The O-Glycologue web application, described in Methods, provides an easy way to experiment with the effects of knockouts or knock-ins of the enzymes of O-glycosylation.
The transferase activities leading to cores 5 through 8 are as yet uncharacterized [1], but could be added in future to account for such structures as are occasionally found in colonic tissues. The O-glycan structure [L4Y3L4[f3]Y6][L3]VT was also not predicted by the current model (Table 4). Although its appearance could be the result of a wider acceptor specificity of β3Gn-T2/3/4/5/7 (10) that would allow this enzyme to act according to the pattern *[Lb4[fa3]Y*T, it could also be the result of fucosylation of an inner GlcNAc by one of the several known α1,3-fucosyltransferase variants, such as FUT4 [68]. The pattern corresponding to the substrate acceptor in such a case would be *Lb4Y*T. An additional α1,3-fucosylation pattern that was evident from this data set is the sequence *L4[f3]Y6*, evident in ten of the non-predicted glycans from two studies [60, 62], and in the sole non-predicted structure of Table 4. It is likely that a fucosyltransferase activity exists that is yet to be characterized, and which acts on type-2 chains with a preference for the 6-linked GlcNAc of core-2 or core-4 O-glycans. In the future, these reactions, as well as those of other fucosyltranserases that are distinguished by different substrate specificities, could be incorporated into the simulator either as additional rules or as refinements of the existing rule (11).
Some structures that were not predicted may also have been mischaracterised. For example, the non-predicted glycan structure described by Podolsky [52], to which we assigned the identifier [S6][[S6L3Y6][S6L3Y3]L4Y3]VT, is in the same paper identified as a type-2 structure, which could be predicted. Our validation study therefore provides a lower bound on the number of structures that can be predicted. Certain poly-6-sialylated structures, including [S6][S6L3Y3[S6]L4Y3]VT, were not predicted. It is possible that a sialyltransferase activity exists in colon that recognises galactose at a distance from the non-reducing end of an oligosaccharide; for instance, an alternative reaction of ST6GlcNAc-I (18) might be CMP-S + *Y3Lb4Y*T = CMP + *Y3[Sa6]Lb4Y*T.
Our analysis of the monosaccharide content of O-glycans extracted from the CFG database revealed that the frequency of occurrence of Neu5Ac was between two and three times the total of the remaining monosaccharides of lesser occurrence: Glc, GlcA, Kdn, and Neu5Gc. Of these, Neu5Gc, or N-glycolylneuraminic acid, is of particular interest because it is immunogenic in humans as a result of the silencing of CMP-N-acetylneuraminate monooxygenase (EC 1.14.18.2). This enzyme, which is active in other mammalian species, adds a single oxygen to CMP-N-acetylneuraminate to form CMP-N-glycolylneuraminate. Neu5Gc obtained in the diet can become incorporated into the cell surface glycome, especially that of cancerous tissue, making it a potential target for immunotherapy [69]. Sialic acids entering the cell via endocytic pathways become activated by the nuclear enzyme CMP-sialate synthase (EC 2.7.7.43, N-acylneuraminate cytidylyltransferase) [70]. Together with the observation that CMP-Neu5Gc can readily substitute for the native donor in reactions catalysed by the sialyltransferases from other species [71], a reasonable assumption is that Neu5Gc is incorporated into human glycoforms by this means. Thus, while Neu5Ac may be the dominant component of the sialylated epitopes expressed in O-linked and N-linked glycoproteins, a portion of such glycans generated by the enzyme simulator could be considered as terminating in Neu5Gc. If the sialyltransferase activities of Table 2 were allowed to act with CMP-Kdn as donor, an additional six structures from the validation study could be predicted by the model, increasing coverage of the data set to 89%.
The notation we have described provides a succinct way to encode structural information for both graphical representation and modelling. Other linear string representations of carbohydrates exist, such as LINUCS [72] and Linear Code [31], which are broader in scope than O-GalNAc glycosylation, and are supported by established glycoinformatic software tools, such as GlycoWorkbench [73]. An advantage of the modelling language described in this work is that it is able to encode the sialic acid Neu5Gc, which cannot be expressed in Linear Code. A more general, and widely supported carbohydrate encoding format is GlycoCT [32]. More recently, the Web3 Unique Representation of Carbohydrate Structures (WURCS) formalism was introduced with an even wider scope [74]. The GlycoForm web application, described in the methods, is able to output any O-glycan structure identifier as both IUPAC, Linear Code and GlycoCT condensed formats, making it interoperable with other software and databases. For the purposes of modelling and display, however, the advantages of the structure identifiers presented in this work are twofold; first, adherence to a strictly one-letter system for the monosaccharides reduces the memory requirements, which can be large when all enzymes of the model are allowed to act; second, the lexical analysis is simplified, since in the drawing algorithm each character can act as a single instruction.
The method could be adapted to other systems, depending on the intended application. For instance, other enzyme activities could be included to account for branch termination by α-GlcNAc, as has been observed in porcine gastric mucins [10], but not commonly on human glycoproteins [42]. The formal grammar could be modified to describe N-glycans, such as those expressed on immunoglobulins [75], the hypermannosylated glycans produced by yeasts [76], or glycans initiated through O-linked fucose [77] or mannose [78]. Additional reaction rules could be supplied, as needed, to support the enzyme activities of galactose 6-O-sulfotransferase and α-2,8-sialyltransferase. A limitation of the current implementation is that not all routes to a product may be included: for example, the simulated activity of Core-2 forming enzyme (5) does not recognise a 3-linked sialic acid on the lower arm of Core 1. The alternative route to [Y6][S3L3]VT could be accommodated by including sialic acid as an option to the reaction pattern, similar to the case for reactions that allow sulfation of Gal or GlcNAc.
Although we have restricted our subject to the enzymes of O-glycan biosynthesis, the actions of glycosidases, which are involved in O-glycan degradation, may have an important regulatory role. For example, it is known that α-l-fucosidase (EC 3.2.1.51) is downregulated in certain types of colorectal cancer [79], from which we infer that an increase in Lewis-type epitopes might be the result of both increased fucosyltransferase activity in Golgi and decreased fucosidase activity in either tissue or plasma. In the future, therefore, this model could be extended to include enzymes involved in the catabolism of O-linked glycoproteins. A quantitative analysis of O-linked glycosylation, incorporating the kinetic parameters of the enzymes involved, would be a natural extension, and development along these lines is proceeding.
The web application, O-Glycologue, provides a convenient way to draw O-glycan structures from the identifiers used in this work, and to explore the wide variety of possible oligosaccharide structures formed by the activities of several known enzymes of O-glycosylation. While a MATLAB-based system for modelling N- and O-linked glycosylation has recently appeared [15], the system described in this article requires neither installation by the user nor a commercial software license. To our knowledge, O-Glycologue is the first tool capable of testing the effects of knockouts of the enzymes of O-linked glycosylation on glycoform heterogeneity. As a knowledge-based system, it should be useful to glycobiologists interested in predicting the biosynthetic pathways forming particular O-glycans. Given that the glycoslation of mucins is known to change during cancer progression [7, 69], the software may be an aid to discovering the enzyme activities most responsible for the formation of particular cancer biomarkers.
In conclusion, we have presented a method for encoding and displaying mucin-type O-glycans, and a method for generating reaction networks from enzymes known to act in O-glycosylation. The formal grammar and the enzyme reaction rules of Table 2, together with an initial glycan identifier as an axiom, comprise the deductive apparatus of a formal system for the modelling and display of these O-glycans. Through an analysis of the reaction networks, we predict that β4Gal-T4 is a key regulator of mucin-type O-glycan heterogeneity, along with β3Gn-T2/3/4/5/7, Gcnt2, C1Gal-T, C2Gn-T and CHST4/6. A comparison of the output of the model with experimentally derived glycans suggests the existence of several novel activities. This approach, which has been validated by structure predictions and the effects of enzyme removal, is intended to form a basis for future kinetic evaluations, and extensions to accommodate other types of glycan structure.
|
10.1371/journal.pgen.1005472 | Multicopy Single-Stranded DNA Directs Intestinal Colonization of Enteric Pathogens | Multicopy single-stranded DNAs (msDNAs) are hybrid RNA-DNA molecules encoded on retroelements called retrons and produced by the action of retron reverse transcriptases. Retrons are widespread in bacteria but the natural function of msDNA has remained elusive despite 30 years of study. The major roadblock to elucidation of the function of these unique molecules has been the lack of any identifiable phenotypes for mutants unable to make msDNA. We report that msDNA of the zoonotic pathogen Salmonella Typhimurium is necessary for colonization of the intestine. Similarly, we observed a defect in intestinal persistence in an enteropathogenic E. coli mutant lacking its retron reverse transcriptase. Under anaerobic conditions in the absence of msDNA, proteins of central anaerobic metabolism needed for Salmonella colonization of the intestine are dysregulated. We show that the msDNA-deficient mutant can utilize nitrate, but not other alternate electron acceptors in anaerobic conditions. Consistent with the availability of nitrate in the inflamed gut, a neutrophilic inflammatory response partially rescued the ability of a mutant lacking msDNA to colonize the intestine. These findings together indicate that the mechanistic basis of msDNA function during Salmonella colonization of the intestine is proper production of proteins needed for anaerobic metabolism. We further conclude that a natural function of msDNA is to regulate protein abundance, the first attributable function for any msDNA. Our data provide novel insight into the function of this mysterious molecule that likely represents a new class of regulatory molecules.
| Multicopy single-stranded DNA (msDNA) is a unique molecule consisting of both an RNA and DNA portion. This molecule is produced by a reverse transcriptase and has no known natural function despite more than 30 years of study. We report that msDNA is important for both Salmonella Typhimurium and an enteropathogenic E. coli, two pathogens that cause diarrhea in susceptible hosts, to survive in the intestine. Using mutant strains incapable of producing msDNA, we show that msDNA is needed for Salmonella to grow in the absence of oxygen. Mutants grown in oxygen-deficient conditions have substantial changes in overall protein composition, including numerous proteins known to be important for anaerobic metabolism and growth in the intestine. Our findings link msDNA to the ability of Salmonella to thrive in an oxygen-deficient environment similar to the conditions inside the gut. We report that msDNA regulates the quantity of proteins, the first natural function attributed to this molecule. msDNA may represent a new class of regulatory molecules.
| Retron reverse transcriptases (RT) in bacteria were first described in Myxococcus xanthus [1] and E. coli [2] in the 1980s and are now known to be widely distributed in the genomes of eubacteria and archaea (reviewed in [3]). All retrons contain three regions essential for production of msDNA: msr (RNA primer for reverse transcription), msd (template sequence), and a reverse transcriptase (RT). The retrons of pathogens, such as Salmonella Typhimurium (STm), may also encode an additional ORF of unknown function [4]. The product of the ‘retron’ is a small covalently linked RNA-DNA hybrid molecule called multicopy single-stranded DNA (msDNA) that is predicted to form complex secondary structures [5]. The predicted secondary structures of msDNA from enteric pathogens including STm, enteropathogenic E. coli and Vibrio spp. are similar [4] but the reverse transcriptase amino acid sequence from these enteric pathogens share little identity. The location of the retron as well as the number of retrons in each species varies. These observations suggest that retrons have been horizontally acquired by convergent evolution to function in a fashion that is specific to the biology of the host bacterium.
Although the molecular details of the production of msDNA have been heavily studied, no natural function has been attributed to this mysterious molecule despite 30 years of study (reviewed in [3]). A critical obstacle to elucidating the natural function of msDNA was the lack of any phenotype for mutants unable to make this molecule. We have shown that the retron reverse transcriptase encoded by STM3846 is essential for Salmonella Typhimurium (STm) to colonize the calf intestine [6], a natural model of enteric salmonellosis that recapitulates the earliest stages of human non-typhoidal Salmonella (NTS) infection. This was the first reported phenotype for a mutant lacking a retron reverse transcriptase.
NTS are major threats to global animal and human health, causing more than 90 million cases of gastroenteritis in people worldwide [7]. Human enteric salmonellosis is characterized by inflammatory diarrhea containing primarily neutrophils. To efficiently colonize the host, NTS use the type 3-secretion system 1 (T3SS-1) encoded on Salmonella Pathogenicity Island-1 (SPI-1) to invade the intestinal epithelium [8,9] and to promote the characteristic neutrophilic inflammatory response. The host inflammatory response gives Salmonella a competitive advantage over resident microflora. Within the intestinal lumen, the product of the neutrophilic oxidative burst generates tetrathionate from oxidation of thiosulfate [10]. Salmonella uses tetrathionate as a terminal electron acceptor within the anaerobic conditions of the intestinal lumen to gain a competitive advantage over resident microflora. Effectors of the TTSS-1 may directly activate epithelial production of inducible nitric oxide synthase (iNOS) thereby creating nitrate, an additional terminal electron acceptor [11]. The relative importance of nitrate during infection is illustrated by the fact that it is a powerful chemoattractant for Salmonella during anaerobiosis [12]. In addition, Salmonella uses host-derived nutrients such as ethanolamine [13] during intestinal inflammation. These strategies facilitate the growth of Salmonella in the complex microbial community of the intestine.
We used the enteric pathogen, Salmonella Typhimurium, to dissect the function of msDNA. In the work described here, we report that mutants lacking msDNA produced by the STM3846 reverse transcriptase are defective for colonization of the intestine using murine models of salmonellosis. This colonization defect is due, in part, to a growth defect for these mutants in anaerobic conditions. We show that mutants lacking msDNA have altered abundance of over 200 proteins in anaerobiosis, many of which are known to be required for growth in anaerobic conditions and for the pathogenesis of STm during enteric infection. Inappropriate abundance of proteins encoding alternate terminal electron acceptor reductases results in an inability of mutants lacking msDNA to utilize these compounds, inhibiting anaerobic growth in vitro. The mutants lacking msDNA can only utilize nitrate as an anaerobic terminal electron acceptor. Mutants lacking msDNA fail to colonize portions of the intestine lacking substantial neutrophilic inflammation, likely due to the ability to only utilize nitrate to support anaerobic growth. Finally, we report a similar defect in intestinal persistence for an enteropathogenic E. coli lacking its retron reverse transcriptase suggesting that msDNA is critical for enteric pathogens to thrive in the intestine of mammalian hosts. Thus, we report a role in regulating protein abundance for msDNA, the first reported natural function for any msDNA. msDNA may represent a new class of bacterial regulatory molecules.
Retron reverse transcriptases, including the STM3846 reverse transcriptase of the St-85 retron, use msr to prime reverse transcription of the msd template sequence to produce msDNA [14] (Fig 1A). We generated a non-polar deletion of msd to establish that msDNA, and not some other potential product of the STM3846 RT, mediates STm colonization of the intestine. Neither the ΔSTM3846 mutant nor the Δmsd mutant produce msDNA and its production can be restored in both mutants by complementation in trans (Fig 1B). The additional ORF, STM3845, is dispensable for msDNA production.
We used the murine colitis model [15], which responds to NTS infection with profound neutrophilic inflammation in the cecum, to dissect the function of the retron in intestinal colonization. We confirmed the requirement for STM3846 in colonization of the inflamed intestine in this model (Fig 1C). In addition, both the Δmsd and ΔSTM3846 mutants have indistinguishable phenotypes, suggesting that the effect of deletion of the RT is mediated by the msDNA itself. The ability of each of these mutants to colonize the intestine is rescued by complementation in trans (Fig 1C and 1D). In cell culture, only the Δmsd mutant invades epithelial cells at a level mildly reduced compared to the isogenic wild type (S1 Fig) suggesting that reduced tissue invasion is unlikely to be the cause of the phenotype that we observed during infection of animal models. Our findings definitively link msDNA to the ability of Salmonella to colonize the intestine.
The intestine is a specialized and highly diverse niche. Oxygen tensions within the lumen decline from the stomach to the colon [16,17], and there is a gradient of increasing oxygen tension from the center of the lumen towards the epithelium [18]. Enteric pathogens must replicate in this hypoxic setting using both aerobic and anaerobic metabolic pathways [19,20] and express genes necessary for virulence in order to compete with resident microflora and colonize the host efficiently.
To determine whether the intestinal colonization defect of the STm msDNA mutants could be due to an inability to grow in oxygen limited conditions, we measured the growth of our mutants in the absence of oxygen, a condition where the retron is highly expressed [21]. Both mutants unable to produce msDNA have severe growth defects in rich media in anaerobic conditions (Fig 2A and 2B, S2 Fig), while the growth of these mutants in the presence of oxygen is similar to the isogenic WT in both rich and minimal media (Fig 2C–2F). The necessity for msDNA during anaerobic growth is consistent with the inability of msDNA-deficient mutants to efficiently colonize the intestine.
We hypothesized that msDNA might act as a trans regulator of gene expression for two reasons. First, small RNAs are well known to have regulatory properties through base pairing with DNA or mRNA transcripts [22]. Second, substantial over-expression of msDNA from one strain of E. coli in a heterologous strain lacking a retron resulted in small changes in the proteome [23]. To determine whether the msDNA produced by the St-85 retron might have regulatory properties, we evaluated the proteome of the WT and msDNA-deficient mutants (ΔSTM3846 and Δmsd) at late exponential phase, a time when the retron is expressed and msDNA is produced (Fig 1B and [24]), in both the presence and absence of oxygen. Of the 1504 total proteins identified, no significant differences in protein abundance between the WT and mutants in the presence of oxygen were detected (Fig 3 and S1 Table). This finding is consistent with previous findings that mutants lacking msDNA grow indistinguishably from the wild type organism in standard laboratory conditions (Fig 2C–2F). In addition, we noted that very few proteins differ in abundance between the ΔSTM3846 and Δmsd mutants, consistent with the hypothesis that the reverse transcriptase and msDNA operate in the same biological pathway.
In anaerobic conditions however, we identified 238 proteins that differed in abundance between the wild type and msDNA-deficient mutants (Fig 3 and S1 Table). Forty-three percent of proteins with reduced abundance in the mutant were involved in amino acid and carbohydrate transport/metabolism and energy production/conversion (Table 1). Twenty-five percent of all proteins of altered abundance did not belong to a functional grouping (Table 1). The abundance of proteins encoded on SPI-1 was unchanged in the absence of msDNA (Fig 3). Proteins necessary for motility were increased in abundance in anaerobically grown msDNA-deficient strains (Fig 3). However, this apparent increase did not result in a change in swimming motility of these strains in anaerobic conditions compared with the WT (S3 Fig). The abundance of numerous proteins known to be important for anaerobic growth and intestinal colonization was significantly reduced (Fig 3 and S1 Table), including proteins for 1,2 propanediol utilization [25], ethanolamine utilization [13], anaerobic sn-glycerol-3-phosphate utilization [26], anaerobic vitamin B12 biosynthesis [27], and serine/threonine degradation [28].
Numerous proteins involved in reduction of anaerobic electron acceptors [29] were altered in abundance between msDNA mutants and wild type bacteria during anaerobic growth (Fig 3). Proteins important for the reduction of thiosulfate (PhsAB) and sulfide (AsrC) were of low abundance (Fig 4 [adapted from [30]] and S1 Table). In addition, proteins necessary for the reduction of DMSO (DmsA, STM4305.s) and fumarate (FrdA) were in low abundance in mutants lacking msDNA, although they did not meet our stringent criteria for statistical significance. Expression of genes necessary to utilize alternate electron acceptors is often induced by the presence of the electron acceptor [29] so the absence of a statistically significant reduction in some of these proteins is not surprising because these compounds were not present in the growth conditions we used. Interestingly, NapA, encoding the periplasmic nitrate reductase [29], was one of the proteins that was present in increased abundance in msDNA deficient mutants compared to the WT, and there was no change in the abundance of NarGH, one of the two other nitrate reductase complexes (Fig 4 and S1 Table). These data are consistent with the growth defect of our mutants in anaerobic conditions, and suggest that msDNA-deficient mutants have a severe dysregulation of proteins necessary for reduction of terminal electron acceptors needed during anaerobiosis.
Our proteomic data predict that msDNA is critical for STm to produce proteins necessary for reduction of terminal electron acceptors critical for metabolism during anaerobic conditions. In order to confirm that the reduced abundance of anaerobic terminal electron acceptor reductases, as indicated by our proteomic data, has functional consequences, we tested the ability of the addition of various terminal electron acceptors to rescue anaerobic growth of the STM3846 mutant. We found that providing the alternate electron acceptors fumarate, DMSO, or thiosulfate to the culture media during anaerobic growth failed to restore growth of the strain lacking msDNA to WT levels (Figs 5B–5F, 6B and 6C). This finding makes sense, as our proteomic data suggest that the enzymes that transfer electrons to these terminal electron acceptors during anaerobic growth, thiosulfate reductase, sulfide reductase, fumarate reductase, and two DMSO reductases, are reduced in abundance in mutants that lack msDNA. However, the addition of nitrate to culture medium rescued the anaerobic growth of the reverse transcriptase mutant (Figs 5A and 6A). These data are consistent with our proteomic data showing that mutants lacking msDNA have adequate NarG and an increased amount of NapA allowing these strains to use nitrate as a terminal acceptor for electrons during anaerobic growth.
In the presence of an intact T3SS-1, NTS induce an inflammatory response that includes recruitment of luminal neutrophils and induction of inducible nitric oxide synthase as part of the inflammatory response [9,10,31], resulting in generation of tetrathionate and nitrate as available terminal electron acceptors in the inflamed intestine. To determine whether the colonization defects we observed were dependent on a functional T3SS-1 and host neutrophilic inflammatory response, we performed competitive infection experiments between the virulent WT and the ΔSTM3846 mutant both in the presence and absence of SPI-1 (Fig 7A). We observed that a ΔSTM3846 mutant colonizes the intestine poorly and associated organs. The modest colonization defect may be due to an inability to utilize carbon and amino acid sources within the inflamed intestine [13], or due to poor growth compared with WT prior to the host inflammatory response. Interestingly, the colonization defect of the ΔSTM3846 mutant in the mouse cecum was exacerbated in the absence of a functional T3SS-1, suggesting that a robust inflammatory response partially rescues mutants unable to produce msDNA (Fig 7A). Consistent with this finding both the small and large intestines, which lack appreciable neutrophilic inflammation (Fig 7C), are poorly colonized with the ΔSTM3846 mutant in mice inoculated with this strain alone (Fig 7B). In murine models that do not develop a neutrophilic infiltrate in the intestine in response to infection (murine typhoid model), the ΔSTM3846 mutant also colonizes poorly after oral infection (Fig 8A and 8B). Our results suggest that STM3846 is essential for STm to colonize the intestine, a defect that is partially rescued in the presence of a profound host inflammatory response, supporting the necessity for intact anaerobic metabolic pathways in intestinal colonization.
The msDNA of STm is similar in predicted secondary structure to msDNA of other enteric pathogens including enteropathogenic E. coli (EPEC; [4]), a close relative of STm. EPEC attaches to the epithelial surface causing characteristic attaching and effacing lesions and a malabsorptive diarrhea [32]. Despite the fact that the pathology caused by NTS and EPEC is distinct, both organisms colonize the intestine and cause diarrheal illness in susceptible hosts. We hypothesized that the RT of EPEC O127:H6, a serotype previously shown to produce msDNA [33], is necessary for this organism to colonize the gut. To test this hypothesis, we generated a non-polar deletion of the retron RT (ΔE2348C_3890) and performed competitive infections between this mutant and the WT. We found that an EPEC mutant lacking the RT fails to persist within the intestine of mice, both in the luminal contents and adherent to tissue (Fig 9). This defect was reversed by complementation in trans (S4 Fig). These data suggest that the importance of retron reverse transcriptases during intestinal infection is not restricted to salmonellae, and thus are likely to be more broadly applicable to enteric pathogens.
The natural function of msDNA has remained elusive despite more than 30 years of study [1,2,5,34–38]. We describe the first phenotypes for any mutant lacking msDNA. Using the enteric pathogen S. Typhimurium, we show that msDNA produced by a retron reverse transcriptase is critical for efficient colonization of the mammalian intestine. In STm, msDNA is critical for the ability to grow in the absence of oxygen. Identification of these phenotypes creates the first opportunity for detailed studies of the molecular function of msDNA since the discovery of these unique molecules. We further showed that STm msDNA directs colonization of the intestine through regulation of the abundance of proteins necessary for central anaerobic metabolism. Thus, our data suggest that the natural function of msDNA may be to control protein abundance, the first natural function to be ascribed to any msDNA molecule.
In STm, msDNA is produced by the STM3846 reverse transcriptase using msd as a template sequence and msr as a primer. The msDNA from STm has a predicted 85-nucleotide DNA stem with no mismatched base pairs and a 4-nucleotide loop, and an RNA portion with two predicted smaller imperfect stem loop structures [4]. The RNA and DNA portions of msDNA are covalently joined by a unique 2’5’ phosphodiester linkage on a conserved guanine [39]. It is unclear whether the entire msr RNA sequence remains in the mature STm msDNA. Consistent with prior reports [34,39], we showed that both the RT and msd are requirements for production of msDNA. The intervening ORF, STM3845, is dispensable for msDNA production. This is perhaps not surprising as the presence of another ORF in addition to the RT in retrons is relatively rare, and appears to be more common on retrons borne by pathogens [4,40,41]. It has been suggested that the retron RT could produce a variety of different cDNA molecules if the sequence of the mRNA transcript is identical to that of the 5’ end of msr [42]. However, we observed similar defects in intestinal colonization and anaerobic growth of mutants lacking either the RT or msd. These data suggest that it is msDNA, and not some other potential product of the RT, that mediates intestinal colonization of STm.
Previous attempts to evaluate the function of msDNA have used artificial systems, failing to identify phenotypes for mutants lacking msDNA and to definitively identify the natural function of these molecules [23,43–47]. When an msDNA from one strain of E. coli with mismatched base pairs in the predicted DNA stem region is significantly overexpressed in a heterologous strain of E. coli lacking its own retron, the frequency of spontaneous mutation was increased due to sequestration of mismatch repair proteins [43,44,46]. Thus, the production of msDNA was thought to increase mutation frequency. However, no previous work has demonstrated that deletion of msDNA from a bacterium naturally producing msDNA decreases mutation frequency. We hypothesize that substantial over-expression of any mismatched DNA could increase mutation frequency by the same mechanism. Thus, this previous finding may not illuminate the true function of msDNA in the cell.
When mutants lacking the ability to make msDNA are grown without oxygen, 15% of all proteins we could identify were in altered abundance. However, no dysregulated proteins were identified during aerobic growth, consistent with the lack of identifiable phenotypes in the presence of oxygen. Our proteomic data were generated using cultures grown for the same duration of time under varying growth conditions. Some of the differences in protein abundance may result because wild type and mutant that cannot make msDNA grow differently during anaerobic conditions. However, our growth data suggest that the growth phase of the wild type and mutants unable to make msDNA are not dramatically different at the times we chose to collect samples for our analysis. Furthermore, the differences in protein abundance between the msDNA mutant and the wild type during anaerobic growth that we re-tested appear to be functionally significant. We show that the growth of mutants that cannot make msDNA, and that have reduced abundance of several alternate electron acceptor reductases needed during anaerobic growth, cannot be rescued by addition of the cognate alternate electrons. Furthermore, the msDNA mutant overproduces periplasmic nitrate reductase (NapA) and a wild type level of a second nitrate reductase (NarG). We show that these proteins and thus this pathway are functional, as the exogenous addition of the terminal electron acceptor nitrate rescues the anaerobic growth of mutants unable to make msDNA. Prior reports suggest that the DNA portion of msDNA can be engineered to act as a regulatory molecule by creating an antisense sequence in the DNA loop [45]. Our data suggest that the natural function of msDNA may be to act as a regulatory molecule although we have not yet identified specific regulatory targets.
There are two known master regulators of anaerobic metabolism in facultative anaerobes: fnr and arcA [48]. The transcriptional and protein profiles of anaerobically-grown Salmonella mutants deficient in fnr and arcA are established [49,50]. With few exceptions, the proteins of altered abundance in our proteomic data align poorly with genes regulated by either fnr or arcA. However, it is difficult to draw meaningful comparisons across our proteomic data and published transcriptional profiles of mutants grown in the absence of oxygen, because of protein profiles with transcript abundance are not directly comparable. Our data raise the possibility that regulation by msDNA may represent an additional pathway to regulate the abundance of proteins necessary for anaerobic metabolism. Further mechanistic study of the anaerobic regulation of gene and protein expression is critical to understanding the behavior of Salmonella in intestinal colonization.
Recent evidence suggests that Salmonella exploits the host inflammatory response to gain a competitive advantage in the intestinal lumen [10–13]. Reactive oxygen species produced by neutrophils oxidize thiosulfate to tetrathionate, a compound that Salmonella, but not resident microflora, uses as a terminal electron acceptor [10]. Epithelial-derived nitrate also contributes to the growth of Salmonella in the anaerobic conditions of the intestine by acting as a preferred electron acceptor in these conditions [31]. Some nutrients, such as ethanolamine, are used only during the neutrophilic inflammatory response [13]. We have shown that some of these processes in STm are altered in mutants unable to produce msDNA, along with many other proteins with less clearly defined roles in pathogenesis.
We also observed a defect in intestinal persistence of an EPEC mutant lacking its retron RT. Salmonella enterica and E. coli are close phylogenetic relatives and both cause diarrheal illness in susceptible hosts, but there are critical differences in the retron between organisms. The RT of EPEC O127:H6 is located in a different genomic context than the retron of STm and has a GC content of 51.8%, similar to the average GC content of 50.6% [51] suggesting that this gene was not acquired recently. This GC content in the EPEC retron is in contrast to the GC content of the retron of STm, 30.6% compared with the average GC content of 52.4% [4,52]. Unlike STm, the retron of EPEC O127:H6 lacks an additional ORF. The predicted secondary structures of msDNA from EPEC and STm are similar, however EPEC msDNA is predicted to have mismatched base pairs in the DNA stem [4]. Despite these differences, we report that EPEC mutants lacking the retron RT also have a phenotype during colonization of the intestine.
Critical differences also exist between the pathogenesis of EPEC and STm diarrheal diseases. In the intestine, Salmonella lives both in the lumen and invades the epithelium, replicating intracellularly and inducing a profound neutrophilic inflammatory diarrhea [53]. In contrast, EPEC attaches to the intestinal epithelium below the intestinal mucus in these regions of the gastrointestinal tract and remains extracellular [54]. Although our understanding of the molecular mechanism of the development of diarrhea during EPEC infection is incomplete, this infection causes a secretory diarrhea [32]. Thus, the mechanism of EPEC-induced diarrhea is substantially different than the inflammatory diarrhea caused by non-typhoidal salmonellae. Our data suggest that retron RTs are critical for colonization of the intestine by both of these pathogens, yet the phenotypes of these mutants in Salmonella versus EPEC during infection are different. While the role of retron reverse transcriptases and msDNA in intestinal colonization by enteric pathogens is likely to be ubiquitous, we hypothesize based both on our data and on the differences in diseases between these two organisms, that the processes regulated and the regulatory targets themselves are likely to be different.
We show that a natural function of msDNA is to regulate protein abundance, the first reported natural function of any msDNA molecule. STm mutants unable to make msDNA poorly colonize the murine intestine. This colonization defect is due to altered abundance of numerous proteins, including those necessary for central anaerobic metabolism, a process known to be necessary for the ability of STm to colonize the intestine of mammals. We observed that an EPEC mutant lacking its retron reverse transcriptase has a reduced ability to persist in the murine intestine, suggesting that the presence and function of msDNA may be broadly applicable to other enteric pathogens. Retrons are also widespread in non-pathogenic eubacteria (Reviewed in [3]) including most isolates of the environmental bacterium Myxococcus xanthus [35]. msDNA is present in high copy per cell [1], suggesting that the regulatory function of this molecule is critical for the lifestyle of the host bacterium. It is puzzling that this molecule appears to have a function under only certain conditions despite the fact that it is produced in abundance. One possible explanation for this phenomenon is that msDNA may sense environmental changes in order to regulate gene expression. This hybrid RNA-DNA molecule represents an exciting new class of bacterial regulatory molecules with broad application to the understanding of the lifestyles of pathogens and non-pathogens alike.
All bacterial strains, plasmids, and primers used for mutant construction are listed in S2 Table, S3 Table). All Salmonella strains are derivatives of ATCC 14028s. Enteropathogenic E. coli O127:H6 strain E2348/69 [55], a generous gift of M. Donnenberg, is the genetic background for all EPEC mutants described here. Mutants were constructed using a modification of the lambda-red recombination technique and antibiotic resistance cassettes removed as previously described [56,57] [58]. All Salmonella mutations were moved into a clean genetic background by P22 transduction [59]. Standard cloning protocols were used to generate complementing plasmids [60].
All bacterial cultures were grown at 37°C aerobically with vigorous agitation or standing in an anaerobic chamber with internal atmosphere of 5% H2, 5% CO2, and 90% N2 (Bactron I, ShelLab). For anaerobic growth experiments, bacteria were grown overnight aerobically then transferred into the anaerobic chamber and diluted 1:100 into media pre-equilibrated for at least 18 hours. Alternate electron acceptors (Sigma-Aldrich) sodium nitrate, sodium fumarate, sodium thiosulfate, and sodium tetrathionate were added to LB to a final concentration of 40mM. Sodium chloride (Sigma-Aldrich) at a final concentration of 40 mM served as a negative control. DMSO (Sigma-Aldrich) was added to LB to a final concentration of 0.1% (v/v). Bacteria were grown in Luria-Bertani (LB) broth or LB or MacConkey (Difco) agar supplemented with the following antibiotics as appropriate: kanamycin (50 mg/L), nalidixic acid (50 mg/L), carbenicillin (100 mg/L), streptomycin (100 mg/L), and chloramphenicol (20 mg/L).
All experiments were performed on at least three separate occasions. Bacterial generation number was calculated using the following equation: [log10(CFU final)—log10(CFU start)]/log10(2).
Ethics Statement: This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The Institutional Animal Care and Use Committees of Texas A&M University and North Carolina State University approved all animal experiments (protocol numbers 2012–084 and 2011–167 (TAMU) and 14–132-B (NCSU)). All experiments that utilized mice were performed using 8–12 week old female C57BL/6J mice (Jackson Laboratories). For competitive infection experiments, mice were infected by gavage with an equivalent ratio of WT and mutant bacteria. The competitive index was determined by dividing the ratio of WT to mutant bacteria in the selected organ by that ratio in the inoculum. For single infections, mice were infected with either WT or mutant bacteria. The harvested tissue was weighed, homogenized, and CFU was determined per gram of tissue collected.
Salmonella infections were performed as previously described [15]. For the murine colitis model, mice were administered 20 mg streptomycin in 75 μL sterile water by gavage. Twenty-four hours after treatment, mice were infected with approximately 108 CFU of Salmonella in 100 μL volume by gavage. Feces were collected 24 hours after infection. Mice were euthanized by carbon dioxide asphyxiation at 96 hours post-infection and organs harvested, homogenized, serially diluted, and plated on LB agar with appropriate antibiotics for enumeration of CFU. For the murine typhoid model, mice were treated with 75 μL sterile water by gavage. Mice were then infected and euthanized as above.
EPEC mouse infections were performed essentially as previously described [61]. Mice were infected with approximately 108 CFU in 100 μL volume by gavage. Feces were collected every other day for 9 days. Mice were euthanized 10 days post-infection. The aboral 5 cm of small intestine, the entire cecum, and the entire colon were collected. Intestinal contents were exposed through a longitudinal incision. The intestinal segment was placed into sterile PBS and vigorously agitated to remove intestinal contents. Intestinal tissue was washed in sterile PBS to remove remaining ingesta. Intestinal contents and tissue were homogenized separately, serially diluted, and plated on MacConkey agar and LB agar with appropriate antibiotics to enumerate CFU.
Samples from mouse ileum, cecum, and transverse colon were collected 96 hours post-infection and fixed in formalin. All tissues were routinely processed and stained with hematoxylin and eosin. All histologic analyses were performed by a veterinary pathologist blinded as to infection group. Tissues were scored (0–4) for each of the following parameters: polymorphonuclear cell (PMN) infiltration, mononuclear leukocyte infiltration, crypt abscess, submucosal edema, villus blunting, and epithelial damage as described [13,15,62,63].
msDNA was isolated from aerobic late log phase cultures normalized by OD600. Bacteria were lysed as for plasmid isolation (Qiagen Mini-prep) and msDNA isolated from the filtered fraction with subsequent ethanol precipitation. msDNA was visualized using a native polyacrylamide gel with in-gel ethidium bromide staining.
Cell lines were purchased from American Type Culture Collection (ATCC) and used within 15 passages. HeLa cells (human cervical adenocarcinoma epithelial, ATCC CCL-2) were grown as recommended by ATCC. HeLa cells were seeded in 24-well plates at 5 x 104 cells/well approximately 24 h prior to infection.
Late-log phase cultures were prepared by inoculating 10 ml LB broth with 0.3 ml overnight shaking culture. Flasks were grown at 37°C with agitation for 3 hours. Bacteria were collected by centrifugation at 8000 x g for 90 seconds, resuspended in an equal volume of Hanks’ buffered saline solution (HBSS, Mediatech) and added directly to mammalian cells seeded in 24-well plates for 10 minutes. The multiplicity of infection was approximately 50. Non-internalized bacteria were removed by aspiration. Monolayers were washed three times in HBSS and were then incubated in growth media until 30 min post-infection. Thereafter, gentamicin was added at 50 μg/ml from 30–90 min p.i. to kill extracellular bacteria and reduced to 10 μg/ml from 90 min post-infection For enumeration of intracellular bacteria, monolayers were washed once in phosphate-buffered saline, and then solubilized in 0.2% sodium deoxycholate and serial dilutions were plated on LB agar.
Swimming motility was performed as previously described [64]. Swimming was assayed on plates containing 0.3% Difco Bacto Agar (LB agar base 25g/L). Plates were incubated either in open air or in the anaerobic chamber overnight prior to use for swimming assays. Overnight cultures of bacterial strains were grown at 37°C with agitation and cell numbers normalized by optical density. An aliquot of each normalized culture was transferred into the anaerobic chamber. The WT, ΔSTM3846, and Δmsd mutants (3 μl each) were spotted onto the same swimming agar plate and incubated at 37°C aerobically or anaerobically for 5 hours. The diameter of the cell spread was measured and compared with that of the WT on the same plate. Each assay was performed in triplicate on three independent occasions (anaerobic) or in four replicates on two independent occasions (aerobic).
Statistical analysis was performed using GraphPad Prism 6. All data were log transformed prior to analysis. Statistical significance was set at P < 0.05 and was determined using a t-test or ANOVA where indicated.
Aerobic overnight cultures of the wild type and the ΔSTM3846 and Δmsd mutants were diluted 1:100 and incubated either aerobically or in an anaerobic chamber (Coy) for 4 hours on three independent occasions. Bacteria were pelleted and supernatants discarded. Cell pellets were resuspended in 100 mM NH4HCO3, pH 8.0 and lysed by vigorous vortexing in the presence of 0.1 mm silica/zirconia beads. Proteins were denatured and reduced with 8M urea and 5 mM dithiothrietol, respectively, for 30 minutes at 60°C. The proteins underwent enzymatic digestion for 3 hours at 37°C with 1/50 enzyme/protein (w/w) ratio of sequencing-grade trypsin. The resultant peptides were desalted for mass spectrometric (MS) analysis using C18 solid phase extraction cartridges (50 mg, 1 mL, Discovery, Supelco). The cartridges were activated with methanol, followed by equilibration with 0.1% TFA before loading the samples. The cartridges were then washed with 5% acetonitrile (ACN)/0.1% TFA and eluted with 80% ACN/0.1% TFA. Eluted peptides were concentrated in the vacuum centrifuge and diluted to a concentration of 0.5 mg/mL with water for the MS analysis.
Digested peptides were loaded into capillary columns (75 μm x 35 cm, Polymicro) packed with C18 beads (3 μm particles, Phenomenex) connected to a custom-made 4-column LC system [65]. The elution was performed using the following gradient: equilibration in 5% B solvent, 5–8% B over 2 min, 8–12% B over 18 min, 12–35% B over 50 min, 35–60% min over 27 min and 60–95% B over 3 min. (solvent A: 0.1% FA; solvent B: 90% ACN/0.1% FA) and flow rate of 300 nL/min. Eluting peptides were directly analyzed either on an Orbitrap (LTQ Orbitrap Velos, Thermo Scientific, San Jose, CA) mass spectrometer using chemically etched nanospray emitters [66]. Full scan mass spectra were collected at 400–2000 m/z range and the ten most intense ions were submitted to low-resolution CID fragmentation once (35% normalized collision energy), before being dynamically excluded for 60 seconds.
Tandem mass spectra were searched with MSFG+ against Salmonella enterica serovar Typhimurium 14028s and common contaminant sequences (downloaded from NCBI, all in forward and reversed orientations), using the following parameters: (i) partial tryptic digestion, (ii) 50 ppm parent mass tolerance, (iii) methionine oxidation as a variable modification. The peptides were filtered with a MSGF probability score [67] ≤ 1x10–9. Peak areas for each peptide were retrieved using the MultiAlign tool [68], and to ensure the quality of peptide-to-peak matching, the data was filtered with a Statistical Tools for AMT tag Confidence (STAC) score ≥ 0.7 and uniqueness probability ≥ 0.5 [69]. Additionally, proteins were required to have at least 2 peptides and at least one peptide with STAC ≥ 0.9. Peptide abundance values were log transformed and rolled-up into proteins using Qrollup tool, available in DAnTE [70]. Abundance values for each protein across all 32 conditions (WT, mutants, anaerobic, aerobic conditions, biological replicates, and technical replicates) were used to calculate a Z-score for each measurement where missing values were filled with 19.5. The Z-score transformation enables comparisons of trends across conditions and proteins to identify relevant abundance changes.
|
10.1371/journal.pgen.1000718 | A Robust Approach to Identifying Tissue-Specific Gene Expression Regulatory Variants Using Personalized Human Induced Pluripotent Stem Cells | Normal variation in gene expression due to regulatory polymorphisms is often masked by biological and experimental noise. In addition, some regulatory polymorphisms may become apparent only in specific tissues. We derived human induced pluripotent stem (iPS) cells from adult skin primary fibroblasts and attempted to detect tissue-specific cis-regulatory variants using in vitro cell differentiation. We used padlock probes and high-throughput sequencing for digital RNA allelotyping and measured allele-specific gene expression in primary fibroblasts, lymphoblastoid cells, iPS cells, and their differentiated derivatives. We show that allele-specific expression is both cell type and genotype-dependent, but the majority of detectable allele-specific expression loci remains consistent despite large changes in the cell type or the experimental condition following iPS reprogramming, except on the X-chromosome. We show that our approach to mapping cis-regulatory variants reduces in vitro experimental noise and reveals additional tissue-specific variants using skin-derived human iPS cells.
| Most complex traits likely result from a combination of genetic polymorphisms. The normal variation in gene expression is thought to be an important contributor. In order to examine a wide range of personalized tissue types from a given individual, we developed a robust method for detecting regulatory variants genome-wide in human induced pluripotent stem (iPS) cells. By having a platform capable of mapping regulatory variants despite large biological and experimental noise, and by being able to use in vitro differentiation to derive multiple human tissue types, our approach should enable the identification of large numbers of regulatory variants genome-wide using minimally invasive skin biopsies from a large number of human subjects.
| The recent advances in whole genome association studies (GWAS) have uncovered multiple genetic loci linked to common human diseases and traits. In addition to the more interpretable coding sequence changes, a large number of identified loci are in the non-coding region, suggesting that inheritable regulatory polymorphisms may play an important role [1]–[3]. Expression quantitative trait loci (eQTL) studies can reveal both cis- and trans-regulatory variants that can be mapped to a specific genetic region [4],[5]. However, it requires a large sample size to reach the statistical power necessary to observe subtle changes in gene expression due to noise, ‘batch effects’ and other confounding factors [1],[6]. Current mapped eQTL loci account for only a small fraction of the overall genetic risk for a given trait, suggesting that the weak effects from multiple genetic loci may play an important role.
Although eQTL loci in different tissues can overlap [7]–[11], the range of cell types available for study still poses a problem since many regulatory pathways are tissue-specific [1],[12]. Given the potential of eQTL for elucidating genetic causes of complex traits and diseases, an ambitious effort has been launched to collect various tissue types from a large number of individuals (i.e. Genotype-Tissue Expression project). However, the existing approaches to tissue sampling, including the use of surgical and tumor specimens, are complicated by social, medical and legal issues in addition to artifacts associated with tissue collection and processing [1]. In addition, it is difficult to follow up with a functional assay in the same individual and evaluate the biological effect of regulatory variants in the absence of a viable experimental system (i.e. cell lines).
Induced pluripotent stem cells [13]–[16] can be derived from skin, hair or blood [17],[18], using transduction of reprogramming factors (i.e. OCT4, SOX2, KLF4 and MYC). They can be used to derive a number of tissues and cell types in vitro without resorting to invasive biopsy, and differentiation of iPS cells can theoretically allow for tissue-specific eQTL studies. However, the difficulty in observing pure and/or consistent in vitro differentiation can result in significant experimental variability and mask subtle regulatory variants given the practical limits on the sample size. An alternative approach may be to compare the expression level between two heterozygotic parental genes using ‘reporter’ SNPs (expression SNPs) in the exon [19]–[26]. Allele-specific gene expression (ASE) results from cis-regulatory differences in transcription (i.e. upstream activating sequences, DNA methylation, core promoters) or processing (i.e. alternative splicing, miRNA) [27],[28]. As such, the ASE ratio can control for the effect of experimental variations on gene expression, which function predominantly in trans [29],[30].
Here, we used padlock probes and high-throughput sequencing for digital RNA allelotyping to map tissue-specific expression regulatory variants in human iPS cells and their derivatives and showed that allele-specific expression analysis could overcome experimental noise and artifacts. Current approach will allow in vitro experiments on individualized iPS cell lines to map additional tissue-specific and context-dependent regulatory variants.
The Personal Genome Project (PGP) is a repository for pre-consented phenotype and genetic data as well as cell lines, including iPS cells. We derived primary skin fibroblast lines from two participants in the Personal Genome Project (PGP), using two partial depth skin biopsy samples obtained from both arms (Bx1 and Bx2). Clonal populations of PGP1 and PGP9 primary skin fibroblasts (named PGP1Bx1F and PGP9Bx1F) were isolated by routine subcloning. Non-clonal populations of primary fibroblasts (named PGP1Bx2F and PGP9Bx2F) were derived from a second biopsy site. The PGP1 and PGP9 fibroblast populations were transduced with retrovirus expressing pluripotency reprogramming factors (OCT4, SOX2, KLF4 and MYC) [31]. The isolated iPS clones expressed pluripotency markers (Figure 1A) and formed tetratomas containing normal derivatives of all three germ layers (Figure 1B), confirming their functional pluripotency.
In order to harness the accuracy of high-throughput sequencing for quantitative allele-specific RNA analysis, we designed padlock probes targeting 27,000 common exonic SNPs (minor allele frequency > 0.07), representing 10,345 unique genes, based on the hg18 annotation (UCSC Genome Browser) (Table S1). The padlock probes were synthesized on an Agilent array in a massively parallel manner, and they were then PCR amplified and processed to generate single-stranded DNA molecules [19],[32]. The pool of single-stranded padlock probes was annealed to the double-stranded cDNA and/or the genomic DNA, followed by a 9-bp fill-in and ligation reaction to circularize the annealed probes [33],[34]. The circularized products containing the captured sequence were amplified and sequenced on Illumina GAII. On average, we obtained 6.4±2.0 million sequencing reads per sample, and we were able to map 69.8±17.2% of the reads against the RefSeq sequences used for the padlock probe design (Table 1). Approximately 19,000 (70.4%) out of 27,000 SNPs were covered at least 20 times with a mean coverage of 250 reads for each SNP, of which 25% were heterozygous calls (Table 2). Genotyping calls made using Affymetrix 500K and digital allelotyping showed a concordance rate of 98% for >20x coverage and 99% for >50x coverage (Table 3). Among the heterozygous SNPs, the ratio between reference and alternative alleles was symmetrically distributed around 0.51 (Figure 2A), and the distribution of sequencing reads was nearly identical between the two alleles (Figure 2B), suggesting little or no bias in capturing and mapping the reads.
For RNA allelotyping, we amplified the singled stranded cDNA from 50 ng total RNA using linear displacement amplification (NuGen) and generated the double stranded cDNA using random hexamer priming (Invitrogen). We confirmed that the padlock probes captured both + and - strands with a similar efficiency, 51.6% and 48.4% respectively (Table 4). Typically, we observed ∼1,300 (25%) heterozygous expression SNPs out of ∼5,200 total expression SNPs. As expected, large ASE deviations were associated with SNPs having a small number of reads (<100), indicating the presence of biological and/or technical noise (Figure 3A). However, the allele-specific expression ratio was highly reproducible between the total RNA replicates (R2 = 0.7994 with <100 reads and R2 = 0.905 with >100 reads) (Figure 3C and 3D). In order to validate our method, we compared digital RNA allelotyping to quantitative Sanger sequencing, which showed a high correlation between the two methods among the 12 heterozygous expression SNPs in PGP1 samples (R2 = 0.825) (Figure 3B).
We then asked whether the total number of reads for each SNP might reflect the gene expression level. We compared the mean number of sequencing reads from probes targeting the same transcript and normalized the values against the number of sequencing reads from the genomic DNA. We then compared these values against the relative gene expression levels as measured by Illumina BeadChip Human Ref-8, revealing only a weak correlation (R2 = 0.1684) (Figure 4A). We also asked whether we were capturing only those genes that were highly expressed. When we compared a list of genes captured using our method and compared it to their relative gene expression level, 159 out of 1124 (14%) captured SNPs were associated with the genes below the detection limit on the BeadChip platform (Figure 4B). These results suggested that digital RNA allelotyping was capable of detecting rare transcripts and that the absolute read counts did not necessarily reflect the overall gene expression level, possibly due to differences in probe hybridization, abundance and/or amplification.
In our previous study, we showed that human fibroblasts, lymphoblastoid cell lines and primary keratinocytes all demonstrated tissue-specific ASE (4.3–8.5% of heterozygous SNPs), using a different probe library design (CES22k-3.2) [19]. When adjusted for the false discovery rate in biological replicates, the percentage of SNPs with tissue-specific ASE was between 2.3–6.5%. Using a new probe design (CES27k-9bpV3), we looked for tissue-specific ASE in PGP1 fibroblasts and lymphoblastoid cell lines (Dataset S1 and Dataset S2). We observed that 3.8% (31/807) of the SNPs showed tissue-specific ASE reproducibly in both replicates. Between iPS clones and primary fibroblasts, the number of reproducible tissue-specific ASE loci increased to 9.8% (107/1091), while it was 6.9% (71/1036) between iPS cells and embryoid bodies (EBs) (Table 5). These findings suggested that up to 10% of ASE showed reproducible tissue-specificity and that they were more numerous in complex and/or heterogeneous tissue samples.
In order to explore the relationship of ASE ratios across a wide range of tissue types, we used 186 expression SNPs that were universally present in multiple cell types from PGP1 and 9 and hierarchically clustered them using un-normalized ASE ratios directly (Figure 5). The sample correlation between the biological replicates was 0.983 (PGP1Bx2 F1 and F2) and 0.987 (PGP1Bx1 iPS1a and iPS1c), while the correlation between primary fibroblasts and lymphoblastoid cells was 0.980 (PGP1Bx2 fibroblasts versus lymphocytes). The differentiated PGP1Bx1 iPS1 derivatives were related to each other with a lower correlation of 0.969. In contrast, the ASE ratio between PGP1 and PGP9 samples had a correlation of only 0.542. We have shown previously that genetic similarity was highly correlated with allelic ratio similarity (R2 = 0.91) [19], and the current result confirmed this conclusion and further suggested that allele-specific expression from human iPS cells were remarkably similar to other cell types from the same individual, despite differences in their epigenetic states [35].
We then normalized direct allelic ratios from the cDNA with those from the genomic DNA in order to reduce probe-specific effects on ASE measurements. To correct for a normalization bias, we calculated the mean ASE ratio across all the samples and used the distance from the mean for hierarchical clustering (Figure 6). Using the relative change in the ASE ratio across multiple cell types, we observed a consistent correlation between fibroblasts (0.31 correlation), lymphocytes (0.39 correlation) and iPS cells (0.24 correlation), while the sample correlation of fibroblasts versus lymphocytes and iPS cells was 0.27 and −0.0093, respectively. Finally, the correlation coefficient between the PGP1 and PGP9 samples was −0.26, indicating a significant difference between the two individuals. From these results, we concluded that the structure of cis-regulatory variants was largely genotype-dependent and that the allelic architecture in gene expression changed to a much smaller degree from cell type to cell type.
Strictly speaking, the ASE ratio was a quantitative measure that reflected the relative abundance of different RNA alleles. However, any detectable differences in ASE alone could also be used as an indicator of functional regulatory variants nearby. In order to assign a confidence score to ASE-mapped genes, we used a chi-squared test (cDNA-to-genomic DNA alleles; χ2>6.64). Since miniscule ASE could be called ‘significant’ solely due to the large number of sequencing reads, we required that the ASE ratio be >0.60 or <0.40. Therefore, our digital ASE calls addressed whether a cis-regulatory variant could be confidently mapped to a gene locus, not whether ASE showed a biologically meaningful allelic imbalance. When examining 427 digital ASE-positive SNPs out of 1822 total SNPs in technical replicates, the correlation coefficient of ASE ratios increased from 0.8672 to 0.9766 (Figure 7A), suggesting that much of the measurement noise had been eliminated due to a large number of observations. Using technical replicates, we also estimated the false discovery rate of digital ASE calls to be 1.6% (Figure 7B), and when all the samples were adjusted for the false discovery rate, 27±4.7% of the heterozygous expression SNPs were ‘confidently’ mapped in any given sample (Table 6). In order to show that digital ASE calls did not depend solely on the number of observations, we compared digital ASE-positive and negative calls and looked at the number of cDNA and genomic DNA reads as well as the average ASE deviation. We observed that the number of cDNA and genomic DNA reads were ∼45% higher, whereas the average allelic ratio deviation was ∼250% higher in the ASE-positive calls (Table 7). We also examined the ASE calls between PGP1 and PGP9 in order to see if they reflected the difference in allele-specific expression (Dataset S1 and Dataset S2). While the allelic deviation was ∼90% higher in the ASE-positive calls, the number of genomic DNA reads was also ∼120% higher. These results indicated that our method for mapping ASE-associated regions was influenced by all three parameters, as expected.
In order to visualize tissue-specific ASE loci associated with high confidence scores, we examined 1522 heterozygous expression SNPs in 20 PGP1 and PGP9 samples, out of which 317 SNPs were shared among at least 80% of the samples. When these digital ASE calls were hierarchical clustered, there were able to discriminate different tissue types and individuals (Figure 8A). A possible explanation of why digital ASE calls reflected tissue-specificity was that higher tissue-specific expression resulted in higher cDNA observation counts. However, we previously demonstrated that there was no appreciable difference in the number of cDNA reads between ASE-positive and -negative calls in a variety of tissues (Table 7). In addition, the average number of sequencing reads correlated poorly with the absolute gene expression level (Figure 4A), suggesting that the differences in read counts alone did not explain tissue-specific ASE mapping. We also examined individual-specific ASE-positive clusters with the sample correlation of 0.7223 in PGP1 (29/317). Interestingly, a large fraction of PGP1-specific clusters were characterized by consistent ASE calls across all cell types (Figure 8B), indicating that approximately 1/3–1/2 of the mapped cis-regulatory variants were cell type and context-independent.
So far, we attempted to map cis-regulatory variants using the gene transcripts that were universally present among various cell types and found that up to 10% of the genes might be influenced by tissue-specific regulatory variants. However, we expected that other cis-regulatory variants would only be detected using tissue-specific transcripts. In order to capture these variants, we compared different cell types with a similar sequencing depth (5.3–7.4 million reads) and counted the number of ASE-positive calls that were specific to that tissue. We were able to examine between 1,500 to 1,900 heterozygous expression SNPs in primary fibroblasts, immortalized B-lymphocytes, iPS cells and iPS-derived embryoid bodies (EBs) from PGP1 (Table 8). The number of expression SNPs unique to each cell type was 34 (2.2%) and 49 (3.2%) for fibroblasts and lymphocytes, respectively. In contrast, we observed 126 (7.8%) and 287 (14.9%) tissue-restricted expression SNPs in iPS cells and EBs, respectively. This suggested that iPS cells and EBs expressed many transcripts absent in primary cell lines. In addition, we found that the percentage of ASE-positive SNPs was generally lower in fibroblast- and lymphocyte-specific transcripts (∼24%) as compared to iPS and EB-specific transcripts (∼38%). Overall, the number of ASE-positive loci mapped using primary fibroblasts alone was 391, which increased to 562 (44% increase) using iPS cells and limited in vitro differentiation. We estimated that more than 12% of all heterozygous SNPs were associated with ‘mappable’ functional regulatory variants using our approach. We expect this number to increase when other differentiated cell types are examined.
Dosage compensation in mammalian somatic cells is achieved by randomly silencing one of the transcriptionally active X-chromosomes [36]. Random X-inactivation in mouse ES cells is tightly coupled to cell differentiation and the silenced X-chromosome can be re-activated by somatic nuclear transfer [37]. In order to determine how ASE might be affected by re-activation of the silenced X-chromosome after iPS reprogramming, we used a clonal population of female primary fibroblasts to generate two iPS cell lines (PGP9Bx1 iPS1 and PGP9Bx1 iPS2). We then examined 66 heterozygous expression SNPs that were present on the X-chromosome. We observed 14 genes (21%) that were expressed and captured in the two iPS cell lines from PGP9. The ASE ratios of these genes were highly reproducible (R2 = 0.98), including 6 out of 14 SNPs (42%) showing a near mono-allelic preference (Figure 9A). We also observed that eight X-chromosomal expression SNPs were shared between PGP9Bx1 F1 and PGP9Bx1 iPS2. Surprisingly, their ASE ratios were proportionately reversed with a negative linear correlation of R2 = 0.52 (Figure 9B). In contrast, the autosomal ASE ratios in the same pair of cell lines demonstrated a positive linear correlation (R2 = 0.63) (Figure 9C). When we examined a polyclonal population of primary fibroblasts (PGP9Bx2F1), their X-chromosomal ASE ratios were near 0.5, likely due to the population averaging of random X-chromosomal inactivation (Figure 9D). These results indicated that both complete and partial inversions of X-chromosomal ASE ratios occurred during iPS reprogramming and that our method was sensitive and robust enough to detect true changes in allele-specific expression due to reasons other than cis-regulatory polymorphisms.
We then examined ASE in undifferentiated and differentiating iPS cells. When considering only the ASE-positive SNPs, we observed that the correlation between iPS biological replicates (R2 = 0.94) was similar to that of technical replicates (R2 = 0.98) (Figure 10A). When iPS cells were treated with 100-µM trans-retinoic acid for 12 hours, the ASE ratio showed a reduction in correlation between replicates (R2 = 0.62), likely due to the heterogeneity of the colony size and the differentiation environment (Figure 10B) [38]. When the iPS cells were further differentiated into embryoid bodies (EBs) for 7 days, we similarly observed a reduction of correlation between replicates (R2 = 0.59) (Figure 10C). We also found that up to 5–13% of the ASE-positive expression SNPs switched the allelic preference during transient and long-term iPS differentiation (Figure 11A), indicating that parental isoforms could be alternately expressed during developmental transitions. While this phenomenon could be due to random stochastic noise, we showed that the ASE ratio was highly reproducible between biological and technical replicates, even among the rare gene transcripts falling below the traditional detection limit. This suggested that ASE switching was due to the biological heterogeneity of stem cell differentiation and not random measurement noise alone. Finally, changes in autosomal ASE did not affect all chromosomes equally during iPS differentiation (N = 6 samples). We observed that Chromosome 6 displayed lower ASE variance that was statistically significant (p-value: 0.022), possibly due to the amount of stable gene imprinting present on Chromosome 6 (Figure 11B). This observation also supported the idea that the variability in ASE during iPS differentiation was not solely due to random noise.
Studying subtle and/or normal variations in gene regulation requires a sensitive and robust method for measuring true genetic effects. Such effects should be measured in a wide range of human tissues, whether by using human tissue samples or in vitro cell culture, both of which can introduce many confounding factors and experimental artifacts. By combining alleles-specific expression analysis together with human pluripotent stem cell reprogramming, we were able to achieve both objectives with high sensitivity and reproducibility. Despite extreme variations in the cell types, the epigenetic status, cell derivation and reprogramming methods and cell differentiation protocols, we were able to detect a subtle allelic imbalance as small as 60:40 and map approximately 27% of the expression SNPs in a given cell line, of which 3–10% were tissue-specific. We also demonstrated that 1/3–1/2 of mappable ASE loci were reproducible regardless of the cell type used and that they were strongly dependent upon the genotype. We also showed that differentiated iPS cells expressed >40% more transcripts associated with ASE and that more should now be mappable using directed in vitro differentiation. Finally, xwe demonstrated two examples of dramatic ASE changes during X-chromosomal inactivation and during iPS differentiation, showing that our approach can successfully detect global changes in allele-specific gene regulation during development.
The reproducibility of ASE loci across many different cell types was reassuring, but it also pointed to the possibility of having a systematic bias throughout all the samples. Thus, we asked whether we could find an example of ASE changes that was both expected and biologically interpretable. We found that the X-linked ASE ratio was proportionately inversed after iPS reprogramming, including those that were partially silenced. It was known that up to 25% of the X-linked genes could escape X inactivation in human cell lines [39], and indeed, we observed 7/23 and 4/16 X-linked SNPs that were only partially silenced in PGP9 iPS cells and fibroblasts, respectively. Our study demonstrated that these genes were still influenced by X inactivation and that the effect remained proportionately similar even after random chromosomal silencing. While nuclear reprogramming has been reported to reset random X-inactivation in cloned mouse embryos [37] and in mouse iPS cells [40], it was not known whether human iPS cells reached an embryonic ground state. However, we showed that human iPS cells from clonal primary fibroblasts possessed an inverted X-chromosome inactivation pattern, suggesting that human iPS reprogramming can indeed completely erase the somatic X-inactivation memory, a property associated with the embryonic ground state.
Conceptually, allele-specific expression is a direct result of functional cis-regulatory mutations or variations. However, it is also caused by random stochastic events [41],[42] and gene imprinting/silencing [43] as well as allele-specific methylation [22]. Because iPS reprogramming is accompanied by a high degree of cell clonality and epigenetic changes, it offered us an unprecedented opportunity to study how allele-specific expression was affected by such factors. Using a genome-wide allele-specific expression analysis on multiple cell types derived from the same individual, our study conclusively showed that the mappable ASE loci were not dramatically affected by the cell clonality, the methylation status and/or the pluripotency reprogramming and that they were highly individual-specific. It indicated that allele-specific expression might be a good surrogate for indicating the presence of functional cis-regulatory variants. The next logical step will be to determine whether this mappable ASE loci are in fact inheritable and that they can combine in the offspring to produce a gene expression phenotype that is much more dramatic and biologically significant.
While it is tempting to use the ASE ratio as a quantitative trait for association mapping, most ASE loci may not produce a strong phenotype in heterozygous individuals. However, allele-specific expression may exert a more direct influence when combined with other functional variants to generate a mixture of functionally altered protein isoforms. With full diploid genome sequencing, it may now be possible to measure the frequency of allelic combinations that may produce measurable effects on the protein function as well as the signaling and/or transcriptional pathways in an allele-specific manner. Our study showed that as many as 5–13% of the mapped ASE loci changed their preference of parent-specific gene expression during early iPS differentiation and development. It will be fascinating to examine whether alternating patterns of parent-specific gene expression associated with functional coding variants can give arise to subtle variations in parent-specific cellular and tissue organization during different phases of the human development.
While the most straightforward cause of allele-specific expression is differential transcription factor binding on the promoter, other mechanisms such as alternative splicing and methylation-mediated repression may also play an important role. We are currently developing technologies for examining additional molecular features beyond gene transcription to explore allele-specific processes during gene expression and processing. While functional haploid cells and organisms have greatly enhanced our understanding of various molecular pathways in simple organisms, especially in conjunction with mutagensis screening, such approaches are not possible in higher eukaryotes such as mice and humans. However, an allele-specific readout such as ASE allows one to study the effect of haploid elements and variations in fully functional cell lines, enabling one to design experiments to dissect the phenotypic consequence using family of cell lines with different genetic combinations. Therefore, the real power of ASE and other analyses may not necessarily reside in their ability to map of regulatory variants, but to determine the mechanism of allelic combinations that can contribute to the development of a complex inheritable phenotype.
While the use of iPS cells and allele-specific expression analysis for expression trait mapping shows much promise, there are limitations to this approach. The iPS reprogramming, and the propagation and differentiation of iPS cells can be laborious and do not scale up easily. It also does not distinguish among various possible mechanisms for allele-specific expression (i.e. promoter activation, alternative splicing, sequence-specific degradation). In order to bypass these bottlenecks, we are engaged in an effort to automate cell immortalization/iPS reprogramming as well as allele-specific expression assays in order to examine a large population of human volunteers with extensive phenotype and genotype data (Personal Genome Project). Leveraging the power of full genome sequencing technology, our approach of using padlock probes will enable one to examine thousands of samples simultaneously, providing a way to explore cis-regulatory variants in many different tissues in thousands of living study volunteers cost-effectively. We are currently also targeting potential regulatory variants using zinc finger nuclease-mediated homologous recombination in iPS cells to alter their ASE profile and the gene expression level. This and other similar efforts to map and understand numerous functional variants in the vast stretches in the non-coding region and integrating it with experimental biology in a high-throughput manner will likely yield a potent insight into the person-specific regulation in gene expression, cellular biology and ultimately, personalized medicine.
Personal Genome Project (PGP) obtained informed consent from human volunteers who have agreed to release both genetic and tissue samples to the research community. All protocols relating to the collection and processing of human data and samples have been approved by Harvard Institutional Review Board (IRB).
The primary fibroblasts were maintained in 15% NCS (Hyclone) D-MEM/F12 (Gibco) supplemented with 10 ng/ml hEGF (R&D Systems), non-essential amino acid (Gibco), Pen/Strep and L-Glutamine (Gibco). The iPS cells were maintained in 20% KO-Serum (Invitrogen) KO-DMEM (Invitrogen) supplemented with 4 ng/ml bFGF (BD Biosciences), β-ME (Gibco), non-essential amino acid, Pen/Strep and L-Glutamine on a γ-irradiated MEF layer (GlobalStem).
Briefly, pMIG containing OCT4, SOX2, KLF4 and MYC along with VSV-G and Gag-Pol vectors were transiently transfected into 293T cells. We collected retrovirus-containing medium and passed through a 0.45-micron filter unit, followed by ultracentrifugation. We added each virus at multiplicity of infection (MOI) of 5 to human primary fibroblasts (passage number <8). We found that clonally derived PGP1Bx1 fibroblasts were more difficult to reprogram, and it required SV40 large T and NANOG to achieve functional pluripotency [13]. By day 21–30 post-infection, hES cell-like flat colonies started to appear, and they were picked manually and propagated on a freshly prepared MEF layer.
The total RNA was prepared using RNeasy (Qiagen). The RNA sample was then linearly amplified and synthesized into a single-strand cDNA using a whole transcriptome amplification method (NuGen). The linearly amplified single-stranded cDNA is then converted into double-stranded cDNA fragments using random hexamers and E. coli DNA polymerase at 16°C for 2.5 hours. Of note, we did not observe a significant difference in read counts between the first strand and the second strand (Table 4).
Circularization was performed in 20-ul reactions containing 400 ng genomic DNA or 200 ng ds-cDNA, 0.5 pmole padlock probes (total concentration), 2U AmpLigase (Epicenter), 2U AmpliTaq Stoffel fragment (Applied Biosystems), 0.1 µM dNTP in 1x AmpLigase buffer. The reactions were incubated at 95°C for 5 minutes, 60°C for 48 hours. The reactions were then denatured at 94°C for 1 minutes, cooled down to 37°C, then digested with Exonuclease I (10U) and Exonuclease III (100U) for 2 hours at 37°C, and finally heat inactivated at 94°C for 5 minutes. Post-capturing PCR reactions were performed in 100-ul reactions including 10-ul circularization products, 0.4x SYBR Green I, 0.4 µM forward and reverse PCR primers in 1x iProof PCR master mix. The parameter for real-time PCR was 98°C 30 seconds; followed by 3 cycles of 98°C 15 seconds, 53°C 20 seconds, 72°C 10 seconds; then <15 cycles of 98°C 15 seconds and 72°C 20 seconds. We terminated the reactions when the amplification curves went up close to the plateau stage. The 154-bp amplicon was purified with a 6% TBE polyacrylamide gel (Invitrogen), and sequenced with Illumina Genome Analyzer II.
We designed the padlock probes to ensure that the captured sequences are uniquely mappable to the genome using UCSC BLAT. We mapped sequencing reads (25–41 bp) to the sequences by NCBI BLAST using the word size of 8–12 depending on the read length, considering the variant site as degenerate (NCBI Short Read Archive #SRA008291.1). For any sequences that had more than one hit, we required that the second hit had an e-value 5-fold higher than the top hit. In contrast, Maq-based mapping could not handle degenerate sequences, and it was consistently biased towards the reference allele. We made genotyping calls using the “best-P” method on SNPs that were sampled at least 20 times. For each SNP we performed both the test of homozygosity (assuming the allelic ratio of (1-e)/e where e is the sequencing error) and the test of heterozygosity (assuming 50:50 allelic ratio), and determined the genotype based on the one that giving a higher p-value. We used chi-squared test to identify expressed SNPs that exhibit RNA allelic ratios significantly different from the genomic allelic ratios (see Table S1, Dataset S1, Dataset S2). Hierarchical clustering and image viewing were done on Cluster and TreeView.
|
10.1371/journal.pcbi.1006172 | On the role of sparseness in the evolution of modularity in gene regulatory networks | Modularity is a widespread property in biological systems. It implies that interactions occur mainly within groups of system elements. A modular arrangement facilitates adjustment of one module without perturbing the rest of the system. Therefore, modularity of developmental mechanisms is a major factor for evolvability, the potential to produce beneficial variation from random genetic change. Understanding how modularity evolves in gene regulatory networks, that create the distinct gene activity patterns that characterize different parts of an organism, is key to developmental and evolutionary biology. One hypothesis for the evolution of modules suggests that interactions between some sets of genes become maladaptive when selection favours additional gene activity patterns. The removal of such interactions by selection would result in the formation of modules. A second hypothesis suggests that modularity evolves in response to sparseness, the scarcity of interactions within a system. Here I simulate the evolution of gene regulatory networks and analyse diverse experimentally sustained networks to study the relationship between sparseness and modularity. My results suggest that sparseness alone is neither sufficient nor necessary to explain modularity in gene regulatory networks. However, sparseness amplifies the effects of forms of selection that, like selection for additional gene activity patterns, already produce an increase in modularity. That evolution of new gene activity patterns is frequent across evolution also supports that it is a major factor in the evolution of modularity. That sparseness is widespread across gene regulatory networks indicates that it may have facilitated the evolution of modules in a wide variety of cases.
| Modular systems have performance and design advantages over non-modular systems. Thus, modularity is very important for the development of a wide range of new technological or clinical applications. Moreover, modularity is paramount to evolutionary biology since it allows adjusting one organismal function without disturbing other previously evolved functions. But how does modularity itself evolve? Here I analyse the structure of regulatory networks and follow simulations of network evolution to study two hypotheses for the origin of modules in gene regulatory networks. The first hypothesis considers that sparseness, a low number of interactions among the network genes, could be responsible for the evolution of modular networks. The second, that modules evolve when selection favours the production of additional gene activity patterns. I found that sparseness alone is neither sufficient nor necessary to explain modularity in gene regulatory networks. However, it enhances the effects of selection for multiple gene activity patterns. While selection for multiple patterns may be decisive in the evolution of modularity, that sparseness is widespread across gene regulatory networks suggests that its contributions should not be neglected.
| Many biological systems are modular. That is, their interactions occur predominantly within groups of elements and rarely between groups [1–3]. Modularity seems to be associated to important attributes of distinct biological systems. For example, theoretical and experimental efforts have associated a modular organization to structural robustness in RNA [4] and proteins [5] and to the resilience of metapopulations [6]. The relevance of modularity extends beyond the scope of fundamental research in biology. The reason is that modularity confers design and functional benefits to distinct classes of systems. Therefore, for disciplines as diverse as evolutionary robotics [7, 8], artificial intelligence [9, 10], neuroscience [11, 12] and synthetic biology [13, 14], it is relevant to study the effects of a modular organization and how to construct modular systems. Advances in this direction may lead to new useful therapeutical and technological applications. Here, however, I focus on the role that modularity has in development and evolution.
Several researchers have underscored modularity in developmental mechanisms as an important property that facilitates evolution [1, 15–22]. The main underlying reason is that, in modular systems, perturbations of an element are often contained within its module and have few, little or no effects on the rest of the system [23]. It is thus possible to optimize a module without disturbing the functions of other modules. By allowing independent modification of different traits or functions, modularity increases the range of phenotypes that random genetic change can access. For example, the existence of distinct separate gene sets, i.e. modules, for beak width and length may have been an important factor in the evolution of a wide range of beak shapes in the Darwin finches’ adaptive radiation [24, 25]. That the production of distinct aspects of the colour pattern in different parts of butterfly wing blades depends on small and unique sets of genes has also indicated that modularity plays an important role in the evolution of a wide diversity of wing patterns in Heliconius butterflies [26].
Gene regulatory networks play fundamental roles in guiding developmental processes [27, 28]. They consist of sets of genes that cross-regulate their expression through regulatory interactions. Such interactions depend mainly on the production of transcription factors that bind cis-regulatory regions in other genes [28]. Upstream factors, be them genes external to the network, molecular signals coming from neighboring cells, environmental cues, or maternal factors, define a network’s initial state, in which some genes may be active and other genes may be inactive. Then, the genetically encoded interactions guide a dynamic process in which some genes change their activity state. Network dynamics eventually settles in a developmental end state, a gene expression pattern that indicates the network’s commitment to a particular task [29]. I will refer to such a terminal expression pattern as the network’s gene activity phenotype (GAP). The same gene network may yield different GAPs when subject to different signals that produce distinct initial states of gene activity. Thus, gene networks produce the different GAPs that distinguish tissues, organs and cell types. Evolution of new GAPs, through modification of gene regulatory networks, has produced many evolutionary innovations across the tree of life [30–32].
Given the relationship between modularity and evolvability, it is paramount to evolutionary biology to understand the origins of modularity. Because modularity refers only to how mechanisms are structured, and not directly to the phenotypic output of such mechanisms, modularity does not increase fitness by itself. Therefore, to understand how modules evolve we need to study how a modular arrangement relates to other properties [2]. Different groups have proposed distinct mechanisms to explain the evolutionary origins of modularity [2, 33–35].
Modularity may arise whenever interactions between distinct sets of genes obstruct adaptation [36]. In that case, selection would disfavour organisms with such deleterious interactions thus decreasing the number of interactions between sets. There are several hypotheses for what makes interactions between different sets of genes consistently deleterious. G.P. Wagner and collaborators suggested that interactions between genes associated to two different characters may be detrimental when one of such characters is subject to stabilizing selection and the other to directional selection. In this case, interactions between these groups of genes would obstruct the action of either form of selection [1, 17, 37]. In a different scenario, network modularity evolves when selection fluctuates recurrently between two selection regimes. In one regime, selection favours the performance of one complex task that combines two subfunctions. In the second regime, selection favours a different task that combines the same subfunctions in a different manner [38]. Hence, phenotypic optima vary in a modular manner. In this scenario, excessive interactions between elements involved in distinct subfunctions are deleterious. The reason is that networks that have different sets of elements assigned to different subfunctions and that combine subfunctions through few interactions, require less mutations to alternate from producing one optimum to producing the other. Thus, such modular networks have an evolutionary advantage. Congruently with this hypothesis, bacteria living in environments where fluctuations are more frequent tend to have metabolic networks with higher modularity scores [39].
Evolution of new gene activity phenotypes concerns another scenario where interactions between distinct sets of genes are selected against [40]. Specifically, modularity may increase if selection favours networks that produce an additional GAP, besides the network’s ancestral GAP. This scenario requires two sets of genes: one in which each gene has the same activity state in the ancestral and new GAPs (set A), and one in which each gene is inactive in one of the GAPs but active in the other (set B). Here, interactions between genes in the two sets obstruct adaptation: A gene in set A regulated mainly by genes in set B would likely present different activity states in the two GAPs. Alternatively, a gene in set B with a strong influence from set A would tend to show the same activity state in the new and ancestral GAPs. In other words, a gene that is more heavily influenced by genes in the other set than by genes in the same set would not comply with selection. In support of this scenario comes the observation that evolution of new gene activity phenotypes is frequent in an evolutionary scale. Such new GAPs are associated, for example, to the evolution of new cell types or to the specialization of serial homologues [30]. It is noteworthy that sister cell types [41–43] or specialized serial homologues [30, 44] tend to share the activity state of some genes but differ in that of others, thus conforming to this scenario’s requirements. Moreover, modularity is not lost once new GAPs evolve [40]. This contrasts with the ‘modular fluctuations’ scenario, in which modularity drops when fluctuations between selection regimes stop [38].
An alternative influential hypothesis for the origins of modules does not consider that interactions between specific sets of genes may be detrimental. Rather, it considers that any interaction is slightly deleterious and that network sparseness underlies modularity [45]. Clune and collaborators studied the evolution of feed-forward boolean networks as a model for the evolution of biological networks. They found that when connections come at a fitness cost and connectivity decreases, modular networks evolve. Importantly, many biological regulatory networks are sparse [46]. Moreover, later contributions suggested that sparseness is also associated to a hierarchic organization [47] and, in artificial neural networks, to an enhanced ability to learn new tasks [9].
How modularity is assessed is specially germane to discussions regarding the sparseness hypothesis for the origin of modules. Given a network structure, assume that there is an a priori proposal for a network partition P that assigns nodes to non-overlapping sets that serve as presumptive modules. The score QP reflects how interactions are concentrated within the modules that P assumes [3] (see Methods). However, when studying a network structure, one often has no preconception regarding the number or composition of modules. A common strategy is then to use an algorithm to search for the network’s best partition into modules [3, 48, 49]. That is, these algorithms attempt to find a partition that maximizes the modularity score. I will call Qopt to the optimized modularity score that results from such a search. Because of local fluctuations in connection density, even networks that allocate their interactions in a random fashion, without any bias, may have random islands of nodes with connection densities greater than in the rest of the network [50–53]. It is noteworthy that random islands of a high connection density appear more easily in sparse networks [50, 51, 54]. For example, Guimerà et al found that Qopt decreases with the number of connections in random Erdös-Rényi networks [50]. Thus, sparse random networks seem more modular than denser networks. These observations prompt the question of what is exactly the role of sparseness in the evolution of modularity in developmental gene regulatory networks. Is it only that sparse random networks have wider fluctuations in connection density or are there other effects? The question is specially relevant, considering the importance of modularity for development and evolution and the ubiquity of sparseness in biological networks.
To dissect the role of sparseness in the evolution of modularity, it may be useful to separate the effects that sparseness alone has on modularity from effects caused by other factors. One may do so by comparing a network’s raw modularity scores, Qopt or QP, with those of networks devoid of constraints other than preserving the number of connections and degree distribution. To attain such a comparison, I will use normalized scores Q P N and QN, that equal 0 when a network’s raw modularity score equals that expected for random networks with the same number of interactions (see Methods). A positive normalized score means that a network is more modular than networks with the same number of interactions.
As a means of illustration, Fig 1 shows a comparison for two sets of random networks that differ in the number of interactions. Fig 1 first approaches modularity in these networks without any preconception of how they are partitioned into modules. Thus, modularity is assessed in Fig 1A using an algorithm that identifies partitions that optimize modularity [55]. This panel shows that, as indicated by the raw optimized modularity score Qopt, sparse networks seem to be more modular than networks with more interactions. Fig 1 also presents an analysis of the same sets of networks, but using the normalized modularity score QN, that compares the raw score Qopt to the expected value of the same score in networks with the same connectivity attributes. Fig 1B shows that the distribution of QN is practically the same for random networks with contrasting connectivities.
Fig 1C shows a comparison of the same sets of networks, but now in terms of QP, the raw modularity score associated to a specific partition P that, for this example, I chose arbitrarily. The distribution of QP is centered at 0 for both sparse and dense networks. However, Fig 1C also shows that sparse and dense networks differ in how easily a random network’s QP score tends to deviate from 0. Notwithstanding, Fig 1D shows that, if we evaluate the modularity associated to partition P using the normalized score Q P N, random networks with contrasting connectivities present very similar means and spread. In sum, normalized modularity scores may be a better indicator of the deviation in modularity with respect to random expectation, regardless the number of connections. They allow to discard the effect that sparseness alone has on modularity.
One may distinguish three non-exclusive possible manners in which sparseness could contribute to modularity in developmental gene regulatory networks: i) Sparseness may suffice to explain modularity in such networks. If this were the case, gene regulatory networks would be as modular as random networks with the same number of interactions. ii) Sparseness may be necessary for the evolution and consolidation of modules. In this case, evolution of modularity would be impossible in dense networks. iii) Sparseness may enhance the effects of other mechanisms. To assess the role of sparseness in the evolution of modularity, I designed computer simulations of the evolution and dynamics of gene networks and analyses of the structure of biological regulatory networks. I report that sparseness is neither sufficient nor necessary to explain modularity in gene regulatory networks. Notwithstanding, I also found that sparseness has a positive effect on the evolution of modularity when it is combined with additional mechanisms. In sum, this study suggests that, despite its positive contribution, sparseness is not a critical factor for the evolution of modularity. Moreover, the analyses that I put forward are consistent with the proposal that modularity evolves easily under selection for new additional gene activity phenotypes, where interactions between distinct sets of genes are deleterious. That gene regulatory networks are indeed sparse facilitates the evolution of a modular arrangement.
The work that I present here considers both the development and evolution of gene regulatory networks. In terms of gene networks, development entails the production of stable GAPs through regulatory interactions among a network’s genes.
To model a gene regulatory network’s developmental dynamics, I consider a set of N nodes that represent cross-regulating genes. A vector st describes the system state at a given time t by listing the activity state of each gene at that moment. At time t, a gene i may be active (s i t = 1) or inactive (s i t = 0).
In nature, an organism’s genome defines interactions between regulatory genes, mainly through specification of cis-regulatory regions that transcription factors bind [28]. In the model, who regulates whom is defined in a matrix G that represents an organism’s genotype (Fig 2A). A positive entry in matrix G, gij > 0, indicates that gene j favours the expression of gene i. In contrast, a negative entry gij < 0 means that gene j inhibits the activity of gene i. The change in gene activity depends on:
s i t + 1 = σ i [ ∑ j = 1 N g i j s j t - θ i ] (1)
where σi is a step function given by
σ i ( x ) = { 1 , if x > 0 s i t , if x = 0 0 , if x < 0 (2)
Thus, a gene i becomes active if the sum of the influence of i’s active regulators surpasses a threshold θi specific for gene i. I consider that the value of θi depends on gene i’s regulators. Specifically, θ i = ∑ j = 1 N g i j 2. This value of θi guarantees the existence of combinations of activity states of i’s regulators that can switch gene i on or off. Moreover, with this definition of θi and σi, the model is equivalent to a model of gene network dynamics first used by Wagner to study evolutionary properties of gene regulatory networks [56, 57]. Variations of this model have allowed researchers to address successfully several diverse important questions on the evolution of regulatory networks [58]. Importantly, previous research [59] has shown that this model has properties that allow the generalization of results that one may find for one GAP to other easily recognizable GAPs (see details in section 1.1 of S1 Text).
In nature, a network’s dynamic trajectory describes the changes in gene activity that a cell experiences until gene activity settles. The network has then reached a developmental end state, that is, a GAP. Network dynamics starts from an initial state, defined by factors external to the network. In the model, the system starts its dynamics from an initial system state s0 and network genes update their activity state iteratively according to Eq 1. Eventually, because the system is discrete, it attains a state that it has visited previously. In the absence of perturbations, the system then follows the same dynamic trajectory and thus becomes locked in a sequence of k distinct system states that represents a GAP (Fig 2B). The network has then reached a developmental end-state [60]. If k = 1, the GAP is stationary. In it, a gene activity pattern self-maintains (Fig 2B). If k > 1, the GAP is a limit cycle in which a particular system state reappears every k iterations. The same network genotype may attain distinct GAPs when it starts its dynamics from different initial conditions (Fig 2B). In nature, such different initial conditions may arise in distinct parts of an organism subject to different molecular signals.
To assess an individual’s phenotype and evaluate its fitness (Fig 2C), I consider the network’s ability to produce, from different initial conditions, T distinct reference GAPs that serve as evolutionary goals. I will refer to them as target GAPs. I consider that each individual is composed of 100 cells and assigns K = 100/T of them to each of the T functions that the distinct target GAPs optimally perform. A cell required to produce target GAP X starts its dynamics from an initial condition that I obtain by flipping with probability κ each entry in the target GAP X. κ thus reflects how often the sequences of changes in gene activity are perturbed by changes in the initial system state. In nature, such perturbations may arise from developmental noise, for example in the form of stochastic fluctuations in the number of protein molecules, or from random disturbances in a cell’s environment. A cell’s dynamic trajectory stops once the network attains a GAP Y, according to the procedure described in the preceding section. A cell’s contribution to fitness depends on how different Y is from the target GAP X in the following manner:
w = ( 1 - S ) D ( X , Y ) (3)
where S is the selection coefficient, that calibrates how deleterious are deviations from a target GAP. D ( X , Y ) measures how different are Y and X (see section 1.2 in S1 Text).
I define the contribution of a target GAP X to organismal fitness as the arithmetic mean of the fitness contribution of the K cells required to produce X. Organismal fitness is the product of the fitness contribution of each target GAP.
A population is first created as copies of a founder network. To create such a founder network, I built random networks with γN2 interactions until one network appears that produces, in the absence of perturbations of the initial system state, the target GAP that selection favours when evolution starts. Then, I subject network populations to rounds (i.e. generations) of selection and mutation to simulate an evolutionary process. To produce the next generation, I first evaluate the fitness of each network by assessing how similar are the GAPs that it yields to target GAPs, as explained in the previous section. Then, I pick randomly (with replacement) M networks with roulette wheel selection, in which the probability to pick a network is proportional to its fitness. Each gene in a network is then subject to mutation with probability μ.
Mutation changes an entry in the matrix G that lists gene interactions. It thus changes the number of regulators that a gene has (Fig 2D). Gene duplication and other kinds of mutations are not considered in this contribution. As described in detail in S1 Text (section 1.3), the propensity to gain interactions, γ, modulates how likely it is that a new mutation leads to a new interaction and not to the loss of an existing interaction. The probabilities that a gene i loses or acquires an interaction also depend on the number of regulators that i already has. Acquiring or losing an interaction is more likely when the number of i’s regulators is low or high, respectively. In this setup, mutation tends to pull the number of a gene’s regulators towards Nγ. Thus, tuning γ allows to define the expected number of interactions per network γN2.
The analyses that I present here concern those populations in which maximum fitness surpassed a threshold of 0.9. For each of such populations, I chose randomly one network among those with the highest fitness in that population. Notwithstanding, the results are qualitatively the same when including all populations, regardless whether they adapted successfully or not (S1 Fig), and when I consider average population values in successfully adapted populations (S2 Fig). Unless stated otherwise, the parameters that I use in the simulations presented here are: N = 10, κ = 0.05, γ = 0.3, S = 0.4, μ = 0.01 and M = 200. With this choice of parameter values, the optimal phenotype evolved in most cases, without incurring in excessive computational cost. Indeed, even in scenarios where adaptation is slower, evolution of modularity had effectively come to a halt by the end of the simulations (see section 2.1 in S1 Text). Many of the evolutionary scenarios that I address include two distinct stages. The procedures described above are followed in both stages, albeit perhaps with different target GAPs, or with a different value for a specific parameter that affects evolution. The first stage merely sets an ancestral population. Comparing populations before and after evolution under different conditions allows addressing if those conditions increase modularity relative to the ancestral state.
Given a directed network and a specific partition P of the network’s nodes into non-overlapping sets, the network’s modularity, i.e. the degree with which interactions are concentrated within the partition sets, is described with:
Q P = ∑ m [ l m L - d m i n d m o u t L 2 ] (4)
where m refers to one of the sets of nodes that compose partition P, L is the total number of regulatory interactions in the network, lm is the number of interactions within module m, d m i n is the sum of incoming interactions of all nodes in module m and d m o u t is the same but for outgoing interactions [55]. Because random networks that differ in the number of connections or in other networks properties vary in how easily they deviate from random expectation, I compare a network’s QP with that of networks with the same number of nodes, interactions and degree distribution but devoid of any other constraints. Hence, I measure the QP score for the same partition P in a set of 103 networks that allocate their interactions randomly but that preserve the same degree distribution and the same number of nodes and edges as the original network. To create randomized networks I follow the ‘switching’ algorithm [61]. Thus, I assess a normalized version of the QP score that measures modularity after discarding the modularity due to sparseness alone:
Q P N = Q P - Q ^ P S D Q , P (5)
where Q ^ P and SDQ,P refer to the mean and standard deviation of QP for networks in the randomized set.
In the lack of an hypothesis for the composition of modules in the network, an usual strategy is to search for a partition that maximizes QP [3, 48, 49]. Here, I will call Qopt to the modularity score that results from such an optimization procedure. In this paper, I use the spectral method proposed by Newman, known for its advantages in performance and computational cost [49]. Specifically, I use Leicht and Newman’s version of the method that allows its application to directed networks [55].
Because even random networks can have a high Qopt value, it is convenient to compare a network’s Qopt with the random expectation for networks with the same properties. For each network under study, I built a set of 103 randomized networks with the same number of nodes, edges and degree distribution. I then apply Leicht and Newman’s spectral method to each network in the set. Next, I use a z-score as a normalized modularity score:
Q N = Q o p t - Q ^ S D Q (6)
where Q ^ and SDQ refer to the mean and standard deviation of Qopt for networks in the random set. In order to enable comparisons with previous work, section 2.2 in S1 Text provides the main results in this article, but from the perspective of the raw optimized score Qopt.
Previous research supports that network modularity evolves when networks become sparser [9, 45, 47, 62]. However, one must take into account how modularity is regularly assessed. In the absence of an hypothesis for which nodes are integrated into modules, a network’s modularity is usually evaluated by finding one network partition that maximizes the modularity score [48, 49, 63]. I refer to such an optimized score as Qopt. Because the search algorithms look for the best partition, even random networks can have a high Qopt score [50–53]. Sets of nodes with a connectivity (locally) higher than in the whole network appear more easily in sparse than in densely connected random networks. As explained previously, normalized modularity scores allow fairer comparisons of a network’s modularity in terms of how it deviates with respect to random networks with the same number of connections. Therefore, hereafter I will report a network’s modularity using normalized scores. I will use score Q P N whenever I consider a mechanism that predicts the composition of modules. In this case, Q P N tells whether interactions concentrate specifically according to the proposed partition P. If different sets with high internal connectivity were to arise, they should not validate the mechanism under evaluation for its capacity to create and consolidate modules. When an hypothesis for the composition of modules is lacking, I will use the score QN that compares an optimized score in a focal network to similarly optimized scores in randomized networks.
One open possibility is that modularity in developmental regulatory networks is not higher than expected by chance for networks of the same size and connectivity. If that were the case, what we perceive as modules would be the result of local fluctuations in connection density that theory predicts in sparse random networks [50, 51, 54]. Then, the normalized modularity score of developmental networks would follow a distribution similar to that in Fig 1B, with (nearly) equal probability of being positive as that of being negative. To assess this possibility, I revisited 12 recent studies of developmental regulatory networks. Out of the many possible sources of information, I took the network structures from modelling studies sustained on experimental evidence. The reason is that the confidence is high that a model that successfully reproduces a developmental process takes into account all critical factors and interactions involved in the process. Each of these networks can attain any of several stable gene activity phenotypes (GAPs), as they participate in developmental decisions in plants and animals (Table 1). The sample contains studies in plants [64–66], insects [67], nematodes [68], echinoderms [69] and mammals [70–75]. The networks vary greatly in size (ranging from 5 to 94 nodes) and number of interactions (13-209). Connection density does not span a wide interval (1.59 to 3.24 interactions per node), coinciding with the observation that regulatory networks are usually sparse [46]. I found that all twelve networks have a positive normalized modularity score QN (Table 1). In other words, the modularity of each network is higher than the average for random networks of the same size, number of interactions and degree distribution. Moreover, in most cases QN is not close to 0, which is the null expectation. Had QN in these developmental networks come from a symmetric distribution centered on 0, the probability that all twelve scores were positive would be 0.512 ≈ 2.4 × 10−4. These observations support that networks in development tend to be more modular than expected for random networks with the same connectivity. Hence, sparseness alone is unlikely to explain modularity in gene regulatory networks. Moreover, that developmental networks that produce multiple GAPs have a positive normalized modularity score is consistent with the proposal that evolution of new additional gene activity phenotypes may be an important factor for the evolution of modularity [40].
I next used a model of gene network dynamics (Fig 2) to study how modularity may evolve. The model allows to follow the changes in gene activity according to the regulatory interactions that the network specifies. Eventually network dynamics reaches a GAP, a pattern of gene activity in which the system settles. Importantly, a network may produce different GAPs, for example, when it starts its dynamics from different initial system states that may occur in different parts of an organism (Fig 2B). Because gene interactions guide developmental dynamics, the network structure defines the GAPs that a network is able to produce. Therefore, changes in gene interactions may produce changes in the GAPs that a network attains. I subjected network populations to cycles of mutation and selection. On the one hand, a mutation changes an interaction at random (Fig 2D). On the other hand, selection favours those networks that were able to produce gene activity phenotypes similar to predefined target GAPs that are assumed to perform a biological function optimally. Thus, the networks that had greater chances of leaving offspring for the next generation were those that produced the target GAPs even under perturbations of the network dynamics (see Fig 2C and Methods).
Simulations of the evolution of gene regulatory networks have already suggested that networks that evolve to produce both a new and an ancestral GAPs from different initial conditions tend to be more modular than ancestral networks [40]. As explained in the introduction, the reason would be that interactions between distinct sets of genes obstruct the production of either the old or the new GAP. Specifically, previous research had shown that evolution of new GAPs increases QN [40]. Hence, I first assessed whether the modularity score Q P N associated to a partition P also increases in this scenario. To address this question, I considered an ancestral target GAP I, a new additional beneficial target GAP II and two sets of genes A and B of equal size. Each gene in set A has the same activity state in target GAPs I and II. In contrast, each gene in set B has a different activity state in target GAPs I and II (Fig 3A). Thus, the specific partition that I consider, P, allocates genes in set A to one module and genes in set B to a second module, as previous research predicts [40]. Then, I followed the evolution of 500 populations, each with M = 200 individuals (i.e. networks), for 2,000 generations of selection for target GAP I. This first stage of evolution sets ancestral populations before the evolution of a new gene activity pattern. In a second stage of evolution, lasting 8,000 generations, selection favours those individuals that yield GAP I in half of their cells and that produce, from different initial conditions, GAP II in the remaining cells. At the end of each selection regime, I measured Q P N in a network with the highest fitness in each population. 475 populations successfully adapted to both selection regimes. The mean Q P N value for networks after selection for I is -0.14 (SD = 1.009). After selection for both target GAPs, mean Q P N equals 2.569 (SD = 0.846). The difference is highly significant according to a Wilcoxon signed-rank test (Fig 3B; W = 112, 990; p < 2.2 × 10−16). The consolidation of two distinct modules, each comprising genes in either set A or B, is also evident by looking at how interactions are distributed at the end of each selection regime. Fig 3C shows that when selection favours target GAP I alone interactions between any pair of genes appear with similar frequency. In contrast, after selection for both target GAPs I and II, interactions are concentrated within sets A or B (Fig 3D). These observations are consistent with the hypothesis that selection for additional GAPs increases modularity.
Sparseness may still play a role in the evolution of modularity, even if it is not sufficient to increase modularity to the extent observed in gene regulatory networks. Thus, I addressed whether sparseness has a role in increasing modularity beyond the level expected in random networks. To pursue the answer, I evolved 3,000 populations of networks, 500 with each of six different values for the propensity to gain interactions γ. As explained in Methods, the expected number of interactions in evolving networks increases linearly with increasing γ. Network populations evolved in the same scenario as populations in Fig 3: 2, 000 generations under selection for target GAP I and then 8, 000 generations under selection for target GAPs I and II (Fig 3A). These simulations show that a greater number of interactions makes adaptation more difficult in this evolutionary scenario, as the number of populations that successfully adapted decreased with γ. The number of populations in which networks surpassed a fitness threshold of 0.9 at the end of the two selection regimes was 485, 475, 459, 441, 391 and 293 when γ equaled 0.2, 0.3, 0.4, 0.5, 0.6 and 0.7, respectively. I assessed the modularity score Q P N, referring to a partition P into sets A and B, in networks with the highest fitness in populations that adapted successfully. The increase in modularity is less pronounced when interactions accumulate more easily (Fig 4). Notwithstanding, there is a substantial increment in modularity even in networks with a propensity to acquire interactions as unrealistically high as 0.7. In this case, mean Q P N equals 0.065 (SD = 1.013) after selection for target GAP I and it equals 1.292 (SD = 0.964) after selection for I and II. This increase in modularity is also statistically significant (W = 38, 418; p < 2.2 × 10−16).
Even though sparseness is not sufficient, it may well be necessary to increase modularity above levels expected in random networks. If this were the case, the increase in modularity would only occur in networks with few connections, even in regimes where propensity to gain interactions is high. This possibility may seem supported by the observation that adaptation to produce two target GAPs and the increase in modularity are more likely when the probability to gain interactions is low. However, this is not the case. The expected number of interactions for networks evolving in the absence of selection is 20, 30, 40, 50, 60 or 70 when γ equals 0.2, 0.3, 0.4, 0.5, 0.6 or 0.7, respectively. Similarly, the average ± SD numbers of interactions in networks successfully adapted to produce target GAPs I and II are, respectively, 21.38 ± 3.62, 30.1 ± 4.34, 39.69 ± 4.67, 50.25 ± 4.97, 60.78 ± 4.79 and 70.63 ± 4.54. One may think that the number of interactions is so close to the number of expected interactions only because mutation drives too strongly the number of interactions towards values close to γN2 (see Methods and section 1.3 in S1 Text). Nevertheless, even when mutation is not biased towards a specific number of regulators per gene, the number of interactions does not decrease beyond random expectation after selection for an additional GAP (S3 Fig). All these results indicate that networks do not tend to lose connections when evolving modularity under selection to produce two different target GAPs. Consequently, modularity can evolve in non-sparse networks. Moreover, although computational cost prohibits exploration of the evolution of very large networks, simulation of the evolution of slightly larger networks in the same evolutionary scenario are also consistent with these observations (S4 Fig).
It is noteworthy that the greater increase in modularity in networks with a lower propensity to acquire interactions is also observed after selection for I and II when the model of network dynamics is modified substantially. Specifically, this effect is observed when the gene activity threshold θi (Eq 1) is set to zero for all genes (S5 Fig) and when the entries in genotype matrices G are continuous (S6 Fig). These results suggest that the effect of sparseness on the evolution of modularity after selection of an additional target GAP does not depend on the model’s details.
For reasons explained in section 1.1 in S1 Text, what is true for networks that evolve to produce target GAPs I and II will also be valid if another pair of stationary GAPs had taken the place of I and II in the simulation. The only restriction is that the two GAPs should differ in the same number of activity states as I and II. Notwithstanding, I tested whether a low value of γ is associated to a greater increase in modularity when the pairs of target GAPs that selection favours present more and less differences in gene activity than I and II. S7 and S8 Figs show that this effect is also observed in simulations in which the two target GAPs differ in the activity of three and seven genes, respectively. Therefore, these observations are not particular to networks evolving GAPs I and II, but extend to networks evolving a wide variety of pairs of target GAPs.
The simulations that I have already presented suggest that sparseness contributes to the evolution of modularity, despite being neither necessary nor sufficient. An additional analysis supports this interpretation. I evolved 500 network populations under selection to produce target GAPs I and II for 2 × 104 generations. In the course of the first 104 generations, the propensity to gain interactions was relatively high (γ = 0.4). In the last 104 generations γ equaled 0.2. Although Q P N was already high after the first 104 generations of evolution (mean: 2.17; SD: 0.874), it rises substantially after evolution with a low γ (mean: 2.885; SD: 0.75). This increment is also statistically highly significant (W = 83, 538; p < 2.2 × 10−16).
It is noteworthy that the positive effects that sparseness has on modularity do not occur under any selection regime. I studied the evolution of networks under selection for a single target GAP, specifically GAP I. As in the previous analyses, I considered two epochs, each lasting 104 generations. The propensity to gain interactions, γ, equaled 0.4 and 0.2 in the first and second epoch, respectively. Because of the lack of an hypothesis for a partition in this selection regime, I assessed the score QN instead of Q P N. At the end of each of the two epochs modularity was lower than that of random networks with the same connectivity. Mean QN equaled -0.44 (SD = 0.98) when γ had a high value. After evolution under a low γ, mean QN equaled -0.63 (SD = 0.88). Indeed, a Wilcoxon signed-rank test does not support that modularity increased after decreasing the value of γ (W = 51, 492;p = 0.999).
These results together support that sparseness may enhance the positive effects that selection has on taking modularity further above values expected in random networks. Remarkably, these positive effects do not appear in any selection regime. They do appear when interactions between distinct sets of genes obstruct adaptation. Otherwise, sparseness does not seem to contribute to the appearance and consolidation of modules beyond random expectation.
Are there any other factors that, like sparseness, modulate the increase in modularity that other mechanisms produce? I considered that non-genetic perturbations in a network’s initial system state may also play this role. That is, it may be that greater probabilities of altering the dynamic trajectory that leads to a GAP may contribute to increased modularity. The reason would be that the detrimental effects of some interactions between modules may appear only under some combinations of gene activity states but not in others. Thus, trying more and different trajectories may facilitate revealing these conditionally deleterious interactions. In nature, such perturbations in dynamic trajectories may come from developmental noise or from environmental disturbances. In the model, I simulate them by perturbing initial states in network dynamics with probability κ (see Methods).
I evolved network populations under different probabilities κ of perturbing initial system states. As in previous simulations, the first 2 × 103 generations selection favoured networks that produced target GAP I, while the remaining 8 × 103 generations selection benefitted networks yielding target GAPs I and II. Similarly as with sparseness, Fig 5 shows that as perturbation rate takes higher values, the increase in Q P N is greater after selection for two GAPs. Another similarity with sparseness is that a value as extreme as κ = 0, that implies complete absence of non-genetic perturbations of the initial condition, is not able to stop the evolution of modularity after selection for GAPs I and II. One may think that these similarities would easily be explained if evolution under frequent perturbations produces an increase in sparseness and, consequently, on Q P N. This would require that the number of interactions were lower in populations evolving in the face of more perturbations. However, the expected number of interactions (30) is clearly in the bulk of the distribution of interactions in networks evolved under each of the three distinct values of κ that I tested. The mean ± SD number of interactions when κ equals 0, 0.025 and 0.05 is 28.65 ± 4.94, 30.58 ± 4.64 and 30.1 ± 4.34, respectively. These observations support that the effect that perturbation rate has on the evolution of modularity is not contingent on connectivity.
I also performed additional simulations in which selection favoured target GAPs I and II throughout evolution for 2 × 104 generations. In these simulations, the value of κ changed from 0.01 to 0.05 after 104 generations of evolution. Thus, networks were subject to more frequent perturbations of initial system states in the second stage of evolution. After evolution with the higher perturbation rate, the Q P N score increased significantly (W = 97, 200; p < 2.2 × 10−16). Namely it increased from a mean value of 2.01 (SD = 0.92) to 2.64 (SD = 0.78). Together, the results that I present in this section support that a high perturbation rate, like sparseness, amplifies the effects on modularity that other mechanisms produce.
In this article I have addressed the role that sparseness plays in the evolution and consolidation of modules in gene regulatory networks that participate in development. The first possibility that I considered was that sparseness alone could explain modularity in such networks. This hypothesis requires gene regulatory networks to be as modular as random networks with the same number of interactions. My results suggest that this is not the case in developmental gene regulatory networks. A sample of 12 experimentally sustained regulatory networks involved in a wide variety of developmental processes suggests that such networks are more modular than expected for random networks with the same connectivity distribution. Whether this is a common property of other networks in biology remains an open question. A hint in this direction is that metabolic networks, represented as non-directed graphs, are also more modular than equally sparse random networks [76]. Here, however, my focus is on developmental regulatory networks. The results that I put forward suggest that factors other than sparseness must be invoked to explain modularity in developmental regulatory networks. One such factor may be selection for multiple gene activity phenotypes, as proposed previously [40].
Next I assessed whether sparseness collaborates with other factors to take modularity beyond the level expected in random networks. Because developmental regulatory networks have the ability to produce more than one GAP and also a modularity higher than that in random networks (Table 1), I decided to check how sparseness combines with selection to yield new GAPs. The analyses suggest that this selection regime increases modularity even when networks are very densely connected. Thus, sparseness is not absolutely required to make networks more modular than as predicted for random networks. Still, I found that sparseness amplifies the increase in modularity that selection for new activity phenotypes produces. Note that this effect does not depend on the random islands of high density that sparseness creates [50, 51]. After all, such analyses considered modularity with respect to the expectation in random networks with the same connectivity and referring to a specific partition P that selection favoured. Moreover, I found that my observations still hold after changes in the specifications of the model, in the identity of the target GAPs that selection favours and in parameter values (S1–S8 Figs).
That adaptation to produce new additional GAPs occurs less easily in dense networks suggest one reason why sparseness facilitates selection’s role in allocating interactions preferentially within modules. The mere abundance of inter-module connections may obstruct selection’s efforts to remove them. In addition, selection may also have lower incentives to remove interactions between modules in dense networks. The reason is that dense networks will also have more interactions within modules that counteract the pernicious effects of connections between modules.
The analyses that I present here support that, while sparseness facilitates the evolution of modularity, modularity can still evolve in non-sparse networks. One scenario that does suffice to obtain modules is that in which phenotypic optima vary in a modular manner, the so-called ‘modularly-varying goals’ scenario [38]. The ubiquity of environmental fluctuations raise the possibility that this scenario explains modularity in many organismal traits. For example, metabolic traits in bacteria that live in less stable environments tend to be more modular than those of bacteria in stable environments [39]. However, this may not be a general case. The metabolic networks in diverse fly or mammal species do not follow the same pattern [77]. Moreover, this scenario demands specifically ‘modular’ fluctuations, since fluctuations between random phenotypic optima do not promote modularity [38]. Another limitation is that the increase in modularity that this scenario produces is not stable. Once fluctuations stop, modularity drops abruptly [38].
Selection for new additional GAPs is a different option that also suffices to obtain modules in gene regulatory networks. In an evolutionary scale, the acquisition of new activity phenotypes has been a regular means to produce innovations across lineages [30, 41, 42]. Such new GAPs arise, for instance, when new cell types or body structures evolve. Indeed, sister cell types [41–43] and serially homologous structures [30, 44] share the activity of some genes, that allow us to recognize their relationship, but differ in the activity of other genes that underlie their distinct identities and functions. Therefore, nature frequently encounters the conditions that this scenario requires for the evolution of modularity in gene regulatory networks.
When selection combines genes that perform one function with other gene activity states, a new GAP emerges. This requires severing interactions that obstruct this new arrangement in gene activity. Thus, it is expected that modules reflect the selection process, as predicted previously [17]. Specifically, modules should group genes with activity states that are highly correlated in developmental end-states. This correlation will be positive or negative depending on whether selection rewards or punishes, respectively, their joint expression. For example, the algorithm by Leicht and Newman [55] identifies three modules in the Arabidopsis floral organ determination network [64]. The first module comprises meristem identity genes that determine whether a meristem is vegetative or gives rise to a flower. Such genes are active in one kind of meristem but not in the other. The second module concerns genes involved in deciding, within the flower, whether reproductive (carpels, stamens) or perianth (sepals, petals) organs are produced. Lastly, the third module concerns the decision between B-gene activity (petals, stamens) and no B-gene activity (sepals, carpels). Note that, in this framework, one module does not necessarily correspond to one function [78–80]. Two or more functions may be associated to the same module if genes assigned to distinct functions have correlated activity states in different parts of an organism or if the functions share a genetic basis.
That selection for multiple functions produces modular mechanisms may be valid even outside of the framework of gene regulatory networks and gene activity phenotypes. Modularity also appears in other kinds of systems where interactions between distinct sets of elements interfere and obstruct the performance of multiple beneficial functions. For example, modularity increases when selection favours networks that recognize two different patterns [45], RNA molecules that produce distinct structural units [4], gene networks that evolve to produce segments and differentiate them [81], and robots that grow and move [7] or that acquire steering and propulsion abilities [8].
Current data suggests that sparseness is widespread in gene regulatory networks. Its ubiquity does not seem to require an elaborate explanation. It stems naturally from the fact that it is easier for random genetic change to eliminate existing interactions than to create new ones [82]. The analyses that I put forward in this contribution suggest the existence of two distinct effects of sparseness in the evolution of modularity. Regarding the first effect, sparseness creates random islands where connection density is greater than in the rest of the network [50, 51]. The second is contingent on other factors, like some form of selection that makes interactions between specific sets of genes deleterious. I contend that the latter effect is more important in the evolution of gene regulatory networks for two main reasons. First, gene regulatory networks seem to be more modular than random networks with the same number of connections. Second, by amplifying the effects of selection, sparseness may facilitate adaptive evolution.
In this perspective, selection for new gene activity phenotypes may take outstanding importance. Besides producing stable increases in modularity [40], the ability to produce different GAPs seems associated with modularity in biological regulatory networks (Table 1). Moreover, the appearance of new gene activity phenotypes has been a frequent factor in the evolution of many organisms [31]. It may well have been a relevant factor in the evolution of modular gene regulatory networks.
|
10.1371/journal.pntd.0007462 | A single-dose ChAdOx1-vectored vaccine provides complete protection against Nipah Bangladesh and Malaysia in Syrian golden hamsters | Nipah virus (NiV) is a highly pathogenic re-emerging virus that causes outbreaks in South East Asia. Currently, no approved and licensed vaccine or antivirals exist. Here, we investigated the efficacy of ChAdOx1 NiVB, a simian adenovirus-based vaccine encoding NiV glycoprotein (G) Bangladesh, in Syrian hamsters. Prime-only as well as prime-boost vaccination resulted in uniform protection against a lethal challenge with NiV Bangladesh: all animals survived challenge and we were unable to find infectious virus either in oral swabs, lung or brain tissue. Furthermore, no pathological lung damage was observed. A single-dose of ChAdOx1 NiVB also prevented disease and lethality from heterologous challenge with NiV Malaysia. While we were unable to detect infectious virus in swabs or tissue of animals challenged with the heterologous strain, a very limited amount of viral RNA could be found in lung tissue by in situ hybridization. A single dose of ChAdOx1 NiVB also provided partial protection against Hendra virus and passive transfer of antibodies elicited by ChAdOx1 NiVB vaccination partially protected Syrian hamsters against NiV Bangladesh. From these data, we conclude that ChAdOx1 NiVB is a suitable candidate for further NiV vaccine pre-clinical development.
| Nipah virus was discovered in 1998 after an outbreak in Malaysia. Since then, several outbreaks have been reported in Bangladesh and India. Although most outbreaks are relatively small, a very high case-fatality rate is reported (75%). Furthermore, human-to-human transmission has been reported. Currently, no approved vaccine or countermeasure exist. In this manuscript, we discuss a vaccine based on a chimpanzee adenovirus. Importantly, the vaccine vector (ChAdOx1) is in clinical trials. In the work presented here, we show that this vaccine is fully protective against both genotypes of Nipah virus. Furthermore, we observe partial protection against Hendra virus, a related virus. Antibodies produced upon vaccination with our vaccine alone are partially protective against Nipah virus. This is an important step forwards towards the development of an approved vaccine for Nipah virus.
| Nipah virus (NiV) is a highly pathogenic emerging virus in the family Paramyxoviridae, genus Henipavirus. The virus was first detected in 1998 when it caused an outbreak of severe, rapidly progressing encephalitis in pig farmers in Malaysia and Singapore, with a case-fatality rate of 38% [1]. Pigs were most likely infected after eating fruit contaminated by infected fruit bats of the genus Pteropus, the animal reservoir of NiV [2]. This was then followed by pig-to-human infection, but only very limited human-to-human transmission was reported [3]. Since then, NiV has caused near annual outbreaks in Bangladesh and India. Reported outbreaks in these countries involve fewer patients, with a higher case-fatality rate (average 75%) [4] and, in contrast to the Malaysia outbreak, there are well-documented cases of human-to-human transmission [5]. It has been hypothesized that the source for NiV infection in index cases in India and Bangladesh is palm sap contaminated with bat urine [6–9]. The most recent outbreak of NiV occurred in the Indian state Kerala in May 2018 where 19 patients were infected resulting in 17 fatalities. Although this state had not seen NiV infections before 2018, bats of the genus Pteropus are prevalent in this area. Sequence analyses have demonstrated that isolates originating from Malaysia and Bangladesh represent two different genetic lineages [10–12].
A second member of the Henipavirus genus is Hendra virus (HeV), which is characterized by a similar pathology and has caused infections in humans in Australia [13].
NiV-caused disease is characterized by the onset of non-specific symptoms such as fever, headache, dizziness, and myalgia. Hereafter patients may develop severe encephalitis and pulmonary disease. Pulmonary disease is observed more frequently in patients infected with NiV Bangladesh. A unique potential complication is the late onset or relapsing encephalitis, which has been documented up to 11 years after NiV infection [14].
The host range of NiV is broad, facilitated by the use of the conserved ephrin-B2 and–B3 as cellular receptors [15], raising the possibility of further outbreaks resulting from transmission from infected livestock or domestic animals. The current lack of licensed vaccines or treatments has prompted the WHO to identify NiV as a pathogen requiring urgent investment into development of countermeasures [16]. Given the sporadic nature of NiV outbreaks, the aim is to develop a vaccine that demonstrates protective efficacy in animal models and acceptable safety and immunogenicity profiles in phase I and II clinical trials. Vaccines which meet these criteria will be stockpiled and may then be used in the event of an outbreak, following clinical trial protocols for use prepared in advance.
Apart from demonstrating efficacy against challenge in animal models, other desirable characteristics for a vaccine to be stockpiled are the availability of large-scale manufacturing processes, thermostability, and safety in all sections of the population including the youngest, oldest, and immunocompromised patients [17].
ChAdOx1-vectored vaccines fulfil all these requirements, making this a promising platform. The ChAdOx1 vector is a replication-deficient simian adenovirus vector which has been used to produce several vaccines which are now in clinical development. A common feature of these vaccines is their low reactogenicity, strong immunogenicity, and the absence of vector replication after immunization, an important safety feature. In pre-clinical studies a single dose of ChAdOx1 vectored vaccines has been shown to be protective against infection with Rift Valley Fever Virus, Middle East respiratory syndrome coronavirus, Mycobacterium tuberculosis and Zika virus [18–21]. Large scale manufacturing has been performed for replication-deficient adenoviral vectored vaccines for Ebola, with one vaccine now licensed and another in advanced clinical development [22, 23]. Further, a simple thermostabilization process allows for vaccine storage at ambient temperatures [24], removing the need for a cold chain for storage and shipping. We now report on pre-clinical immunogenicity and efficacy testing of ChAdOx1 NiVB.
Animal experiment approval was received from the Institutional Animal Care and Use Committee (IACUC) at Rocky Mountain Laboratories. Experiments were performed in an Association for Assessment and Accreditation of Laboratory Animal Care-approved facility by certified staff, following the guidelines and basic principles in the NIH Guide for the Care and Use of Laboratory Animals, the Animal Welfare Act, United States Department of Agriculture and the United States Public Health Service Policy on Humane Care and Use of Laboratory Animals (Protocol # 2017-033E and 2018-035E). The Institutional Biosafety Committee (IBC) approved work with infectious NiV and Hendra virus (HeV) strains under BSL4 conditions and sample inactivation was performed according to IBC-approved standard operating procedures for removal of specimens from high containment.
Henipavirus isolates were obtained from the Special Pathogens Branch of the Centers for Disease Control and Prevention, Atlanta, GA or Public Health Agency, Winnipeg, Canada. NiV Bangladesh (GenBank no. AY988601), NiV Malaysia (GenBank no. AF212302), and HeV (GenBank no. AF017149) have been passaged three, four, and three times in VeroE6 cells respectively. All virus propagation in this manuscript was performed in VeroE6 cells in Dulbecco’s modified Eagle’s medium (DMEM, Sigma) supplemented with 2% fetal bovine serum (Gibco), 1 mM L-glutamine (Gibco), 50 U/ml penicillin (Gibco), and 50 μg/ml streptomycin (Gibco) (2% DMEM). VeroE6 cells were maintained in DMEM supplemented with 10% fetal bovine serum, 1 mM L glutamine, 50 U/ml penicillin and 50 μg/ml streptomycin.
The glycoprotein (G) gene from Nipah virus (Bangladesh outbreak 2008–2010, Genbank accession number: JN808864.1) was codon optimized for humans and synthesized by GeneArt (Thermo Fisher Scientific). The synthesized G gene was cloned into a transgene expression plasmid comprising a modified human cytomegalovirus immediate early promoter (CMV promoter) with tetracycline operator (TetO) sites and the polyadenylation signal from bovine growth hormone (BGH). The resulting expression cassette was inserted into the E1 locus of a genomic clone of ChAdOx1 using site-specific recombination [25]. The virus was rescued and propagated in T-REx-293 cells (Invitrogen). Purification was by CsCl gradient ultracentrifugation, and the virus was titered as previously described [26]. Doses for vaccination were based on infectious units (IU).
Female Golden Syrian hamsters (4–6 weeks old) were purchased from Envigo. Animals were vaccinated I.M. with 50 μl of 108 IU of vaccine or injected I.M. with 50 μl of saline, in each thigh (100 μl total volume). For the homologous challenge vaccine experiment, animals were vaccinated at D-70 and/or D-42. For the heterologous challenge experiment, animals were vaccinated at D-28. Three days prior to vaccination and virus challenge animals were bled via orbital sinus puncture. All animals were challenged with 1000LD50 of virus in 500 μl DMEM via I.P. inoculation: NiV Bangladesh = 5.3 x 105 TCID50; NiV Malaysia = 6.8 x 104 TCID50; HeV = 6.0 x 103 TCID50. We chose the I.P. route as a uniformly lethal challenge route and to be able to compare with previously conducted vaccine experiments [27]. For each study group, 10 hamsters were utilized. Of these, four animals were euthanized 4 (HeV) or 5 (NiV) days post inoculation and the remaining six animals were followed for 28 days post challenge. Weight was recorded daily up to 10 days post infection, and oropharyngeal swabs were taken daily up to 7 days post inoculation in 1 mL of DMEM. Animals were euthanized when >20% of weight loss was recorded, or severe disease signs (e.g. difficulty breathing or paralysis) were observed. Upon euthanasia, blood and tissues were collected and subsequently analyzed for virology and histology as approved by IACUC.
Female Golden Syrian hamsters (4–6 weeks old) were purchased from Envigo. Fifteen animals were vaccinated with either ChAdOx1 NiVB or ChAdOx1 GFP as described above at 56 and 28 days before serum collection. Serum was collected via cardiac puncture, pooled per vaccine group and IgGs were purified using the MAbtrap kit (Sigma) according to manufacturer’s instructions from 10 mL of serum. Purified IgGs were filtered through an 0.45μm filter and diluted to 4.5 mL in sterile PBS. Ten hamsters were immunized via I.P. injection using 400 μl per hamster. All animals were challenged as described above one day post treatment. For each study group, 10 hamsters were utilized. Of these, four animals were euthanized 5 days post challenge and the remaining six animals were followed for 56 days post challenge. Weight was recorded daily up to 10 days post challenge, and oropharyngeal swabs were taken daily up to 7 days post inoculation in 1 mL of DMEM. Animals were euthanized when >20% of weight loss was recorded, or severe disease signs (e.g. difficulty breathing or paralysis) were observed. Upon euthanasia, blood and tissues were collected and subsequently analyzed for virology and histology as approved by IACUC.
Virus titrations were performed by end-point titration in VeroE6 cells, which were inoculated with tenfold serial dilutions of virus swab media or tissue homogenates. After 1hr incubation at 37°C and 5% CO2, tissue homogenate dilutions were removed, washed twice with PBS and replaced with 100 μl 2% DMEM. Cytopathic effect was scored at 5 dpi and the TCID50 was calculated from 4 replicates by the Spearman-Karber method [28].
Sera was heat-inactivated (30 min, 56°C) and two-fold serial dilutions were prepared in 2% DMEM. Hereafter, 100 TCID50 of NiV was added. After 1hr incubation at 37°C, virus was added to VeroE6 cells and incubated at 37°C and 5% CO2. At 5 dpi, cytopathic effect was scored. The virus neutralization titer was expressed as the reciprocal value of the highest dilution of the serum which still inhibited virus replication.
NiV-G Malaysia (residues E144—T602, gene accession number NC_002728) was cloned into the pHLSEC mammalian expression vector [29] and NiV-F Malaysia (residues G26—D482, gene accession number AY816748.1) was cloned into the pHLSEC vector containing a C-terminal GCNt trimerization motif [30]. The constructs were transiently expressed in human embryonic kidney (HEK) 293T cells in roller bottles, as described previously [29]. Supernatant was harvested 96 hours after transfection and diafiltrated using the AKTA FLUX system (GE Healthcare) against either PBS, pH 7.4 (NiV-G) or buffer containing 10 mM Tris and 150 mM NaCl, pH 8.0 (NiV-F). The proteins were further purified by Ni-NTA immobilized metal-affinity chromatography using His-Trap HP columns (GE Healthcare) followed by size exclusion chromatography. NiV-G was purified using a Superdex 200 10/300 Increase GL column (GE healthcare) equilibrated in PBS pH 7.4 and NiV-F was purified using a Superose 6 Increase 10/300 GL column (GE healthcare) equilibrated in 10 mM Tris and 150 mM NaCl pH 8.0.
Maxisorp plates (Nunc) were coated overnight at 4°C with 5 μg of G or F protein per plate in Carb/Bicarb binding buffer (4.41 g KHCO3 and 0.75 g Na2CO3 in 1 L distilled water). After blocking with 5% milk in PBS with 0.01% tween (PBST), serum (2x serial diluted starting at 100x dilution) in 5% milk in PBST was incubated at RT for 1 hr. Antibodies were detected using affinity-purified antibody peroxidase-labeled goat-anti-hamster IgG (Fisher, 14-22-06) in 5% milk in PBST and TMB 2-component peroxidase substrate (Seracare) and read at 450 nm. All wells were washed 3x with PBST in between steps. Prior to using F and G proteins based on NiV Malaysia, we established that cross-reactivity with NiV Bangladesh antibodies was sufficient for usage in ELISA by testing sera known to be positive for NiV Bangladesh antibodies.
Necropsies and tissue sampling were performed according to IBC-approved protocols. Harvested tissues were fixed for a minimum of 7 days in 10% neutral-buffered formalin and subsequently embedded in paraffin. Hematoxylin and eosin (H&E) staining and in situ hybridization (ISH) were performed on tissue sections and cell blocks. Detection of NiV and HeV viral RNA was performed using the RNAscope FFPE assay (Advanced Cell Diagnostics Inc., Newark, USA) as previously described [31] and in accordance with the manufacturer’s instructions. Briefly, tissue sections were deparaffinized and pretreated with heat and protease before hybridization with target-specific probes for NiV or HeV. Ubiquitin C and the bacterial gene, dapB, were used as positive and negative controls, respectively. Whole-tissue sections for selected cases were stained for NiV and HeV viral RNA, UBC and dapB by the RNAscope VS FFPE assay (RNAscopeVS, Newark, USA) using the Ventana Discovery XT slide autostaining system (Ventana Medical Systems Inc., Tucson, USA). A board-certified veterinary anatomic pathologist evaluated all tissue slides.
Statistical analysis was performed by the Log-rank (Mantel-Cox) test to compare survival curves, and by Welch-corrected one-tailed unpaired student’s t-test to compare infectious virus titers in tissue. SEM was calculated for all samples. P-values < 0.05 were significant.
To determine efficacy of the ChAdOx1 NiVB vaccine, we vaccinated groups of 10 hamsters with either a single dose at D-42 or a prime-boost regime at D-70 and D-42. As control groups, we either injected hamsters with ChAdOx1 GFP at D-70 and D-42 or saline at D-42 (Fig 1A). Virus neutralizing antibodies could be detected after a single dose of ChAdOx1 NiVB and increased upon a secondary dose (average VN titer ± SEM = 30.5 ± 5.7 after single dose, 91 ± 21 after boost). In contrast, no neutralizing antibodies could be detected in serum obtained from the control groups (Fig 1B). All hamsters were challenged with a lethal dose of NiV Bangladesh (1000 LD50) via intraperitoneal inoculation on D0 (Fig 1A). All vaccinated animals survived challenge and did not show signs of disease, such as weight loss, at any stage throughout the experiment. This was in contrast to the control groups in which all animals succumbed to disease between D6 and D10 and exhibited weight loss (Fig 1C and 1D), as well as respiratory and/or neurological signs, including labored breathing and paralyzed hind legs. Statistical analysis demonstrated that survival in the vaccinated groups was significant compared to both control groups (P < 0.0001).
Oropharyngeal swabs were taken daily and assessed for infectious virus by limiting dilution titrations. None of the vaccinated animals shed virus at any timepoint. In contrast, control animals from both groups were found to shed virus at D5 and D6 (Fig 1E).
Four animals of each group were euthanized at D5 and lung and brain tissue were harvested. Infectious virus could only be detected in lung tissue of animals from both control groups (average titer ± SEM = 3.3 x 104 ± 2.5 x 104 TCID50/g of tissue) and was not detected in any tissue of the vaccinated animals (Fig 1F). We did not observe any differences between the two control groups.
Lung and brain tissue harvested at D5 were then evaluated for pathological changes. None of the vaccinated animals displayed pulmonary pathology and no viral RNA was detected in lung tissue by ISH. Control animals developed pulmonary lesions that were indistinguishable between the two groups. These hamsters developed bronchointerstitial pneumonia that was characterized by multifocal inflammatory nodules that were centered on terminal bronchioles and extend into adjacent alveoli. The nodules were composed of large numbers of foamy macrophages and fewer neutrophils and lymphocytes admixed with small amounts of necrotic debris. In most cases hemorrhage, fibrin and edema admixed with inflammatory cells was observed. Edema and fibrin often were extended into surrounding alveoli. Alveoli that were adjacent to areas of inflammation were thickened by fibrin, edema and small numbers of macrophages and neutrophils as previously observed in NiV infected hamsters [32]. There was abundant viral RNA demonstrated by ISH in areas of inflammation (brown staining). The viral RNA was predominantly found in type I pneumocytes but was also multifocally present in vascular and bronchiolar smooth muscle and endothelial cells (Fig 2).
To determine efficacy of ChAdOx1 NiVB against NiV Malaysia and HeV, groups of 10 hamsters were vaccinated with a single dose of ChAdOx1 NiVB or a single dose of ChAdOx1 GFP at D-28 (Fig 3A). As before, virus neutralizing antibodies could be detected after vaccination with ChAdOx1 NiVB but not upon injection with ChAdOx1 GFP (Average VN titer ± SEM = 68.6 ± 13.6) (Fig 3B). Subsequently, hamsters were challenged with either NiV Malaysia or HeV (1000 LD50) via intraperitoneal inoculation on D0 (Fig 3A). All vaccinated animals challenged with NiV Malaysia survived with no signs of disease such as weight loss at any stage throughout the experiment. In contrast, animals challenged with NiV Malaysia that received ChAdOx1 FGP all succumbed to infection between D5 and D6. These animals experienced weight loss and respiratory and neurological signs (Fig 3C and 3D). Statistical analysis demonstrated that survival in the vaccinated group was significantly different from the control group (P = 0.0012).
Oropharyngeal swabs were taken daily and assessed for infectious virus. None of the vaccinated animals challenged with NiV Malaysia shed virus at any timepoint. In contrast, control animals challenged with NiV Malaysia were found to shed virus at D5 and D6 (Fig 3E).
Four animals from both groups were euthanized at D5 and lung and brain tissue were harvested. Infectious virus could only be detected in lung and brain tissue of animals from the control group (average virus titer lung ± SEM = 1.5 x 105 ± 5.2 x 104 TCID50/g, brain ± SEM = 6.8 x 101 ± 4.4 x 101 TCID50/g) and was not detected in any tissue of the vaccinated animals (Fig 3F).
Four out of six vaccinated animals challenged with HeV succumbed to disease between D5 and D7. The two survivors showed minimal weight loss (<2%) and no signs of disease. Animals that received ChAdOx1 FGP all succumbed to HeV infection between D4 and D6. These animals showed weight loss as well as respiratory and neurological signs (Fig 3C and 3D). Log-rank (Mantel-Cox) test demonstrated that survival in the vaccinated group was significant (P = 0.0476) compared to the control group.
Oropharyngeal swabs were taken daily and assessed for infectious virus. None of the vaccinated animals challenged with HeV shed virus at any timepoint. In contrast, control animals challenged with HeV were found to shed virus at D4, D5 and D6 (Fig 3E).
Four animals from both groups were euthanized at D4 and lung and brain tissue were harvested. Infectious virus was detected in three out of four lungs of the vaccinated animals and all lungs of the control animals (average virus titer ± SEM = 5.2 x 105 ± 3.6 x 105 and 4.4 x 106 ± 2.2 x 106 TCID50/g tissue for vaccinated and control animals, respectively). No statistical difference in infectious virus titer was found between the two groups using an unpaired one-tailed Student’s t-test (P = 0.0674). Infectious virus was only detected in brain tissue of animals from the control group (average titer ± SEM = 4.6 x 102 ± 2.0 x 102 TCID50/g) and not in vaccinated animals (Fig 3F).
Harvested lung tissue was then evaluated for pathological changes. All four groups of hamsters developed pulmonary lesions. All animals challenged with HeV and control animals challenged with NiV Malaysia developed bronchointerstitial pneumonia which was indistinguishable from the lesions described for the control animals in the homologous challenge study. Vaccinated hamsters challenged with NiV Malaysia developed mild to moderate bronchointerstitial pneumonia and did not display any evidence of pulmonary edema, fibrin or hemorrhage. ISH demonstrated viral RNA predominantly in type I pneumocytes and rarely in vascular and bronchiolar smooth muscle and endothelial cells in animals challenged with HeV and control animals challenged with NiV Malaysia. In vaccinated animals challenged with NiV Malaysia, however; there was very little RNA present and only in type I pneumocytes in areas of inflammation (Fig 4).
Finally, we wanted to assess the protective effect of antibodies elicited after ChAdOx1 NiVB vaccination. Two groups of 15 hamsters were either vaccinated with ChAdOx1 NiVB or injected with ChAdOx1 FGP at D-56 and D-28. All animals were bled at D0 and we collected 13 and 15 mL respectively. IgG was purified from 10 mL pooled serum. Ten animals per group were then injected peritoneally with purified IgG. Animals were challenged with a lethal dose of NiV Bangladesh (1000 LD50) one day post passive transfer (Fig 5A). We were unable to detect neutralizing antibodies in serum obtained at D5 from four hamsters from each group. However, serum from animals treated with NiV antibodies was positive by ELISA against NiV G protein, albeit with a lower reciprocal titer than antibodies in serum obtained from single-dose vaccinated animals (Fig 5B). One out of six animals treated with NiV antibodies succumbed to disease on D11. No weight loss was observed, however the animal showed severe neurological signs. None of the other NiV antibody-treated animals experienced weight loss or signs of disease. Four out of six animals treated with GFP antibodies succumbed to disease between D6 and D8. These animals showed weight loss and respiratory or neurological signs. The two surviving animals did not show any signs of disease throughout the experiment. One of these animals did not seroconvert as measured by ELISA against NiV F and G protein, and it was suspected this animal was not infected. Therefore, this animal was excluded from the survival curve. The log-rank (Mantel-Cox) test demonstrated that survival in the treated group was significant (P = 0.0168) compared to the control group (Fig 5C and 5D).
Oropharyngeal swabs were taken daily and assessed for infectious virus. Shedding was minimal and found in one animal treated with NiV antibodies on D5, and five animals treated with GFP antibodies between D4 and D6 (Fig 5E).
Four animals from both groups were euthanized at D5 and lung and brain tissue were harvested. Infectious virus could only be detected in lung tissue of animals treated with GFP antibodies and was not detected in any tissue of the animals treated with NiV antibodies (Fig 5F).
Lung tissue harvested at D5 was then evaluated for pathological changes. Both groups of hamsters developed pulmonary lesions similar to those described in the homologous challenge study, however; the NiV antibody-treated hamsters developed mild to moderate pulmonary lesions whereas the control animals developed severe lesions. Additionally, none of the NiV antibody-treated hamsters displayed any pulmonary fibrin, edema or hemorrhage. ISH demonstrated viral RNA in type I pneumocytes in areas of inflammation. Abundance of viral RNA was notably less in animals treated with NiV antibodies (Fig 6).
NiV is a re-emerging infectious disease which causes outbreaks with a high case-fatality rate.
No licensed vaccine against NiV currently exists, and it is therefore key that a safe and effective vaccine be developed. Several vaccine candidates have been explored in different animal models. These can be categorized as subunit vaccines or live-vectored vaccines that target the NiV outer membrane proteins G and/or F. Protection against disease and lethality has been shown in hamsters [27, 33], pigs [34, 35], African green monkeys [36–38], cats [39], and ferrets [40, 41]. Efficacy is thought to be mediated by neutralizing antibodies, as passive transfer of antibodies in naive animals also results in protection against disease [27, 42]. These approaches are promising, but no vaccine candidates have so far been moved into clinical trials.
In the studies presented here, we tested the efficacy of a vaccine based on NiV Bangladesh G protein in a replication-deficient simian adenovirus vector in Syrian hamsters. A prime-only as well as a prime-boost regime protected Syrian hamsters against challenge with a lethal dose of NiV Bangladesh and NiV Malaysia, and partially protected against HeV challenge. Furthermore, antibodies elicited by vaccination alone provided partial protection against a NiV Bangladesh challenge.
Two genetic lineages of NiV have been described; NiV Malaysia and NiV Bangladesh [10–12]. Although NiV Malaysia has not caused an outbreak in humans since 1999, the virus was isolated from Pteropus vampyrus, Pteropus hypomelanus and Pteropus lylei in Malaysia and Cambodia [43–45] and another spillover event could occur. Having one vaccine that protects against both lineages of NiV virus would be the easiest and cheapest countermeasure. A single-dose vaccination with ChAdOx1 NiVB, which is based on NiV Bangladesh, fully protected Syrian hamsters against lethal disease caused by NiV Malaysia. The G proteins of the NiV strains used in this study are 95.5% pairwise identical on the amino acid level, with 27 amino acid differences scattered throughout the protein. Although we did not see sterile protection against NiV Malaysia, none of the vaccinated animals showed signs of disease and all were protected against lethal disease. These results suggest that ChAdOx1 NiVB could protect against both lineages of NiV.
Like NiV, HeV is a species in the Henipavirus genus and thus we investigated cross-protection of ChAdOx1 NiVB against a lethal challenge with HeV in Syrian hamsters. The G protein of the HeV strain used in this study was 78.2% identical to the ChAdOx1 NiVB G protein; 133 amino acids differ between the two proteins. ChAdOx1 NiVB only protected partially against HeV challenge; four out of six animals did not survive challenge. We observed a non-significant decrease in infectious HeV titer in lung and brain tissue of vaccinated animals compared to control animals. It is possible that disease progression in vaccinated animals is delayed compared to control animals. This is supported by the delay in time to death; whereas the average time to death is 5 days in control animals, it is 6 days in vaccinated animals.
Cross-protection of NiV or HeV vaccines has been studied by other groups as well. An adeno-associated virus vaccine expressing NiV G protein offered 50% protection against a lethal challenge with HeV in hamsters [46]. In contrast, vaccines based on HeV provide full protection against NiV in the ferret and NHP model [36, 41, 47]. Likewise, high levels of cross-protective antibodies were found in sera from HeV-infected individuals, whereas cross-protective antibodies were limited in NiV-infected individuals [48]. This might be caused by induction of a more robust and cross-reactive immune response by native HeV protein compared to NiV protein, as suggested by Bossart et al. [48].
Human cases of HeV are associated with direct contact with infected horses, the intermediate animal host of HeV, and direct contact with bats or their products has not yet been associated with HeV infection in humans [49]. It is therefore likely that prevention of HeV in horses will completely prevent human cases. Currently, a HeV vaccine (Equivac) is available for horses and fully protects against HeV [50]. Furthermore, the total number of human cases that contracted HeV is relatively low at 7 [13]. Thus, the requirement of a human vaccine for HeV is therefore less urgent than that of a NiV vaccine.
Previous work has shown that the humoral immune response to NiV vaccination is sufficient to protect Syrian hamsters against a lethal challenge with NiV [27, 42]. Likewise, administration of a human neutralizing monoclonal antibody (m102.4) provided full protection against both HeV and NiV in multiple animal models [51, 52]. Administration of purified IgG obtained from ChAdOx1 NiVB vaccinated hamsters provided partial protection against NiV challenge. Furthermore, infectious virus could only be detected in the lungs of control animals and not in the lungs of vaccinated animals, and thus as in previous studies, ChAdOx1 NiVB-elicited antibodies are able to provide protection against a lethal challenge with NiV. Although we were able to detect NiV G protein-specific antibodies in serum obtained from NiV antibody-treated animals, the reciprocal titer was much lower than that detected in serum from Syrian hamsters after a single dose of ChAdOx1 NiVB. It is possible that administering a higher dose of IgG would have led to uniform protection.
Two animals treated with IgG purified from animals which received injections with ChAdOx1 FGP survived a lethal challenge with NiV Bangladesh. Occasional survival has been observed in the Syrian hamster model [33]. The increased survival rate might however also reflect a non-specific effect of treatment with IgG, which has been reported previously [53]. As the survival rate was significantly different between the NiV IgG-treated group and the control IgG-treated group, the passive transfer experiment shows that antibodies elicited by ChAdOx1-NiVB are sufficient for protection against a lethal challenge with NiV.
Animals in the passive transfer experiment were observed for 56 days, to ensure that the two animals that survived would not succumb to disease after 28 days.
The Syrian hamster is a suitable initial small animal model to investigate the efficacy of NiV vaccines, followed by the African green monkey [54]. The immune system of African green monkeys is more like humans than that of hamsters and is therefore seen as a more relevant animal model to test NiV vaccines. Based on the results presented in the current manuscript, future studies are planned to test ChAdOx1 NiVB in African green monkeys, supported by the Coalition for Epidemic Preparedness Innovations (CEPI).
We show that ChAdOx1 NiVB provides complete protection against lethal disease in Syrian hamsters challenged with NiV Bangladesh. Furthermore, ChAdOx1 NiVB vaccination results in complete survival but with limited evidence of viral replication after NiV Malaysia challenge, and partial protection against HeV. Passive transfer of antibodies elicited by ChAdOx1 NiVB vaccination provide partial protection against lethal challenge with NiV Bangladesh.
|
10.1371/journal.ppat.1000117 | Legionella Eukaryotic-Like Type IV Substrates Interfere with Organelle Trafficking | Legionella pneumophila, the causative agent of Legionnaires' disease, evades phago-lysosome fusion in mammalian and protozoan hosts to create a suitable niche for intracellular replication. To modulate vesicle trafficking pathways, L. pneumophila translocates effector proteins into eukaryotic cells through a Type IVB macro-molecular transport system called the Icm-Dot system. In this study, we employed a fluorescence-based translocation assay to show that 33 previously identified Legionella eukaryotic-like genes (leg) encode substrates of the Icm-Dot secretion system. To assess which of these proteins may contribute to the disruption of vesicle trafficking, we expressed each gene in yeast and looked for phenotypes related to vacuolar protein sorting. We found that LegC3-GFP and LegC7/YlfA-GFP caused the mis-secretion of CPY-Invertase, a fusion protein normally restricted to the yeast vacuole. We also found that LegC7/YlfA-GFP and its paralog LegC2/YlfB-GFP formed large structures around the yeast vacuole while LegC3-GFP localized to the plasma membrane and a fragmented vacuole. In mammalian cells, LegC2/YlfB-GFP and LegC7/YlfA-GFP were found within large structures that co-localized with anti-KDEL antibodies but excluded the lysosomal marker LAMP-1, similar to what is observed in Legionella-containing vacuoles. LegC3-GFP, in contrast, was observed as smaller structures which had no obvious co-localization with KDEL or LAMP-1. Finally, LegC3-GFP caused the accumulation of many endosome-like structures containing undigested material when expressed in the protozoan host Dictyostelium discoideum. Our results demonstrate that multiple Leg proteins are Icm/Dot-dependent substrates and that LegC3, LegC7/YlfA, and LegC2/YlfB may contribute to the intracellular trafficking of L. pneumophila by interfering with highly conserved pathways that modulate vesicle maturation.
| Legionella pneumophila is a Gram-negative bacterial species that causes a severe pneumonia known as Legionnaires' disease. Inhalation of L. pneumophila–contaminated aerosols results in the infection of lung macrophages. Following infection, the bacteria use a Type IVB secretion system to deliver multiple effector proteins into the macrophages to create a membrane-bound replicative compartment called the Legionella-containing vacuole, or LCV. The LCV is defined by its recruitment of early secretory vesicles and avoidance of the bactericidal lysosomes. We identified several effector proteins that contain eukaryotic domains and share significant homology with eukaryotic organelle trafficking proteins. We demonstrate that 33 Legionella eukaryotic-like genes (leg) encode proteins that are translocated into host cells. When artificially expressed in yeast, three Leg proteins (LegC2, LegC3, and LegC7) were able to disrupt normal vesicle trafficking and vacuole morphology. In addition, the Leg proteins induced the formation of, and were localized within, distinct structures when expressed in mammalian cells. In the protozoan host Dictyostelium discoideum, expression of LegC3 resulted in the accumulation of membranous compartments containing partially digested material. Thus, LegC3, LegC2, and LegC7 represent novel effector proteins that may contribute to the intracellular lifestyle of L. pneumophila by disrupting normal vacuolar trafficking pathways in host cells.
| When inhaled by humans, the γ-proteobacterial species Legionella pneumophila can cause a severe, acute and often fatal form of pneumonia known as Legionnaires' disease [1]–[3]. Legionella sp. remain one of the leading causes of community-acquired pneumonia, with poor diagnosis and inadequate treatment accounting for many of the reported fatalities [4],[5]. L. pneumophila is an environmental organism often found replicating inside phylogenetically diverse species of amoeba and its ability to cause disease in humans is believed to be largely accidental [6]–[11]. Even though the specific requirements for replication inside protozoans and macrophages may differ [12],[13], L. pneumophila's lifestyle in each host is quite similar.
Wild type L. pneumophila first promotes its own phagocytosis [14]–[16] and then rapidly avoids phago-lysosome fusion [17]–[19]. The pathogen then forms a replicative vacuole rich in early secretory vesicles and ER-derived vesicles [20]–[24]. Effector proteins specifically delivered into host cells by the Icm-Dot Type IVB secretion system are believed to supplant or modify normal organelle trafficking to generate and sustain the Legionella-containing vacuole (LCV) [13],[25]. Several effector proteins have been identified which may be responsible for recruiting vesicles from the ER to the LCV. For example, the effector protein RalF recruits ADP-ribosylation factor -1, a critical regulator of ER and Golgi vesicle formation, to the LCV [26]. Another regulator of ER and Golgi vesicle traffic, Rab1, is both recruited to the LCV and activated by the effector proteins DrrA (SidM) and LidA [27]–[29]. Additional signaling molecules may be targets of L. pneumophila effectors as well, including phosphoinositides [30] and ubiquitinylated proteins [31].
Recently, multiple putative effector proteins have been identified via genetic, biochemical and bioinformatic screens [26], [32]–[44]. Interestingly, many L. pneumophila effector proteins have been found to contain eukaryotic domains or have overall similarity to eukaryotic proteins [26],[33],[43],[45],[46]. Interdomain horizontal gene transfer has been proposed as a mechanism through which these Legionella eukaryotic-like genes (leg) may have been acquired [26],[43],[45],[47]. Although the functions of some of these effectors have been elucidated, the great majority remain uncharacterized. With only a few exceptions, virulence phenotypes associated with genetic deletions of individual effector proteins have not been observed, presumably due to functional redundancy within the pool of translocated effectors [33],[34],[43] and/or host specificity [33],[39],[48]. Thus, because more than one effector protein may affect a single host protein or pathway, determining the significance of individual effectors has proven challenging.
Although non-phagocytic, Saccharomyces cerevisiae shares many of the same trafficking pathways of higher eukaryotes [49],[50], and thus offers an attractive model for the study of bacterial virulence factors. Effectors from many bacterial pathogens, including Chlamydia, Shigella, Pseudomonas, Yersinia, and Salmonella have been studied in yeast [51]–[54]. Furthermore, yeast models have already been used to identify and analyze L. pneumophila effectors based on their lethal effects [36] and their ability to disrupt normal vacuolar protein sorting (VPS) and early secretory machinery [22],[38].
In this work, we report the identification of three L. pneumophila effector proteins, LegC3, LegC2/YlfA, and LegC7/YlfB, which are sufficient to cause VPS defects and/or altered vacuolar morphology in yeast. Importantly, these proteins induce the formation of and are located within similar structures when expressed in mammalian cells. Co-localization evidence is provided supporting the hypothesis that LegC3-GFP, LegC2/YlfB-GFP and LegC7/YlfA-GFP are effector proteins of L. pneumophila that can modify the normal endolysosomal pathway.
In a previous report, we used a bio-informatic approach to identify L. pneumophila eukaryotic-like genes (leg genes) based on the occurrence of eukaryotic motifs [43]. In order to determine which Leg proteins are substrates of the Icm-Dot system, we utilized a novel reporter system previously used for Type-III effector proteins [55]. In this system, TEM1 (ß-lactamase) is fused to a putative effector protein and a strain containing this fusion protein is used to infect host cells. Host cells are then loaded with CCF4/AM which, when excited at 409 nm emits green fluorescence (520 nm) due to fluorescence resonance energy transfer (FRET) between the coumarin and fluorescein fluorophores. If the fusion protein was translocated into host cells, it cleaves the ß-lactam ring of CCF4/AM, releasing the two fluorophores and changing the fluorescence emission from green to blue (447 nm) when excited at the same wavelength. The ratio of blue to green fluorescence can then be quantified using a spectrofluorimeter with the appropriate excitation and emission filters.
We previously found that 8 Leg-CyaA hybrid proteins out of 20 tested are substrates of the Icm-Dot system when CyaA was fused to the C-terminal end of the proteins [43]. In this study, we constructed TEM1 translational fusions to all 45 leg genes in which the Leg protein is fused to the C-terminus of TEM1 (Table 1). Using the fluorescence-based assay, we then tested if these TEM1-Leg hybrid proteins are translocated into J774 mouse macrophages. We consider a protein to be translocated if the ratio of blue to green fluorescence in host cells is greater than 1 after 60 minutes of contact with L. pneumophila expressing the hybrid constructs. If a ratio of less than one over background is observed in mutants that lack essential components of the translocon (dotA or icmT), we consider translocation to be dependent on the Icm-Dot system. We used a TEM1-FabI hybrid protein as a negative control, to show that the overexpression of a housekeeping protein (Enoyl-acyl CoA Reductase) does not result in non-specific translocation through the Icm-Dot secretion system. We also performed immunoblots using anti-ß-lactamase antibodies on L. pneumophila strains carrying all hybrid constructs, including those that were found not to be translocated. With the exception of TEM-LegA7, all hybrids are expressed at similar levels (data not shown). We found that 33 of the 45 Leg hybrid proteins tested are translocated in an Icm-Dot dependent manner (Figure 1). Importantly, 23 of these Leg proteins were not previously known to be translocated and 4 proteins (LegP, LegS2, LegC4 and LegA3) that had given no translocation signal using the C-terminal CyaA fusion system are in fact translocated. The 15 N-terminal TEM1 fusions found not to be translocated in this assay were also not translocated as C-terminal TEM1 fusions (data not shown). These results demonstrate that a large fraction (33/45) of the previously predicted Leg proteins are Icm-Dot substrates.
One successful strategy used to identify proteins that are important for intracellular trafficking has been to search for gene-products that are required for vacuolar protein sorting (VPS) in yeast [56]–[58]. Many genes required for VPS are also important for endosomal trafficking and maturation in higher eukaryotes, including mammals [50]. Since the Legionella-containing phagosome avoids maturation and fusion to lysosomes, we hypothesized that L. pneumophila may translocate effectors that specifically interfere with conserved VPS pathways. In fact, it has been recently shown that 3 L. pneumophila effectors (VipA, VipD, VipF) interfere with VPS in yeast [38] while LidA interferes with steps in the secretory pathway [22].
To find out if any Leg proteins cause a VPS defect in yeast, we screened 29 strains that express the Leg proteins or LepA using an invertase overlay assay (Table 1). This assay detects mis-secreted CPY-Invertase, a protein which normally trafficks to the vacuole [59]. The plasmid used for sub-cloning, based on the low-copy pBM272 vector [60], creates C-terminal GFP translational fusions to the proteins of interest with transcription regulated by a galactose-inducible promoter. We found that the expression of LegC3-GFP and LegC7/YlfA-GFP caused the formation of a brown precipitate used to identify the aberrant presence of invertase activity on the surface of yeast colonies (Figure 2A). Strains expressing GFP alone did not cause a precipitate to form, whereas a strain expressing the Class E dominant negative protein VPS4E233Q [61] induced a strong VPS defect. The qualitative assay was further verified using a quantitative assay (Figure 2B). We conclude from these results that the L. pneumophila effector proteins LegC3-GFP and LegC7/YlfA-GFP induce a VPS defect when expressed in yeast cells.
Galactose-induced expression of LegC5-GFP, LegC8-GFP, and the previously described LepB-GFP each caused a severe growth impairment in yeast cells (Figure 3) and conclusions regarding their effect on vacuolar sorting could not be drawn. This growth phenotype was not dependent on the presence of the GFP tag, since strains that expressed these Leg proteins or LepB in the absence of additional sequence exhibited similar growth defects (data not shown). Some Leg-GFP fusion proteins in this screen (LegC1, LegC4, LegD2, LegK1, LegN, and LegY) did not produce a fluorescent product as determined by epifluorescence microscopy and others (LegK2, LegK3, LegL2, LegLC4, LegL7, LegLC8, LegT, and LegU1) were not detectable by Western immunoblotting with anti-GFP antibody (Figure S1). These proteins may be poorly translated or be highly unstable when expressed in yeast cells. Of the fusion proteins where no VPS defect was detected, a considerable amount of heterogeneity with respect to protein levels was also observed. Similar to the toxic gene products LegC5, LegC8 and LepB, it remains possible that some of the poorly expressed Leg proteins may disrupt vacuolar sorting or cause toxicity if different expression systems are used.
The styryl dye FM4-64 is commonly used to monitor the trafficking of endocytic intermediates to the vacuole in yeast [62]. The dye first labels the plasma membrane, followed by labeling of internalized membrane, endocytic intermediates and finally the vacuole. Cells that have a VPS defect may present vacuolar abnormalities that are organized into 6 distinct classes [63]. These abnormalities can vary substantially and include defects in acidification, vacuolar fragmentation, and accumulation of endocytic vesicles in the pre-vacuole compartment.
Fluorescence microscopy was used to visualize the intracellular localization of 30 Leg-GFP protein hybrids with the aid of FM4-64 to examine the vacuolar membrane. Most Leg-GFP proteins appeared to be cytosolic or were not easily detectable and had no apparent effect on the labeling of the vacuole by FM4-64 (data not shown). However, we observed that three proteins, LegC2/YlfB-GFP, LegC7/YlfA-GFP and LegC3-GFP exhibited specific localization patterns. As shown in Figure 4A, LegC3-GFP is found associated with the yeast plasma membrane as well as punctate vacuolar structures within the cell. Compared to cells expressing GFP alone, LegC3-GFP expression also caused a partial fragmentation and tubulation of the vacuole membrane as seen by the FM4-64 staining pattern. The punctate LegC3-GFP containing structures also co-localized with parts of the fragmented vacuole. In contrast, LegC2/YlfB-GFP and LegC7/YlfA-GFP formed nodular structures on the vacuole and co-localized heavily with the FM4-64 stain. The accumulation of FM4-64 in the pre-vacuole is similar to what is observed in Class E VPS mutants, indicating that these proteins block endosome maturation and trafficking to the yeast vacuole via the multivesicular body (MVB) pathway. Western blots on whole cell lysates using a polyclonal rabbit anti-GFP antibody showed stable expression of all three proteins, although LegC7/YlfA is expressed to a lower level than LegC3 and LegC2/YlfB (Figure 4B).
To determine if the localization of the YlfA/LegC7 and LegC3 proteins depends on other components of the Vps/ESCRT complex, we expressed these constructs in a collection of 25 VPS mutants containing null mutations in genes that encode components of the Vps/ESCRT complexes; (Δvps1, Δvps2, Δvps4, Δvps16, Δvps17, Δvps18, Δvps20, Δvps22, Δvps23, Δvps24, Δvps25, Δvps27, Δvps28, Δvps35, Δvps36, Δvps37, Δvps44, Δvps46, Δvps60, Δbro1, Δdoa4, Δhse1, Δsec28, Δsnf7, and Δvta1). We observed that in some cases YlfA/LegC7 and LegC3 exhibited more intense vacuolar staining, but not mis-localization (data not shown). This result indicates that these proteins do not require other known scaffolding components or docking proteins of the ESCRT complex for their association with the yeast vacuole and are likely tethered to endosomes via putative transmembrane domains, and not through interactions with components of the multivesicular body pathway.
In order to determine which regions of LegC3 are required for producing a VPS defect and localization in yeast, we analyzed a library of LegC3 mutations. Using a transposon-based delivery system, we obtained 60 insertions of 15 base pairs each: 33 mutations were found to contain in-frame codon insertions and 27 contained out-of-frame (nonsense) insertions (Figure 5A). We screened the resulting yeast library for LegC3-GFP localization and vacuolar morphology (as determined by FM4-64 staining). As shown in Figure 5B, we found that two insertion mutants at amino acid positions 390 (LegC3-1-GFP) and 395 (LegC3-2-GFP) altered the protein localization and vacuolar morphology. These two mutant proteins had transposon insertions within the hydrophobic domain, exhibited a diffuse GFP signal, and failed to localize to the plasma membrane. Additionally, the vacuole of cells expressing these mutant versions of LegC3-GFP appeared to be more similar to that of strains that express GFP alone. Western blots were performed on lysates of yeast cells grown in fructose and galactose to confirm that the mutant LegC3-GFP proteins were present at the predicted sizes. As shown in Figure 5C, full length LegC3-1-GFP and LegC3-2-GFP were expressed following galactose induction, although the level of each protein was slightly reduced compared to LegC3-GFP. In addition to the hydrophobic domain, we also observed that the first 511 amino acids of the protein (out of 558) are sufficient for causing the VPS defect, as the first permissive truncation that still caused CPY-Invertase mis-secretion was at this position (data not shown).
Because the amino acid insertions in LegC3-GFP that altered the protein localization were within a hydrophobic region predicted to be a an α-helical transmembrane domain, we decided to test if the hydrophobic regions of LegC2/YlfB and LegC7/YlfA were required for protein localization. As shown in Figure 6A and 6B, deleting the hydrophobic domain of LegC2/YlfB-GFP caused the protein to become cytosolic. LegC7/YlfAΔTM-GFP, in contrast, formed a single bright punctate structure distinct from its original localization pattern on the vacuolar membrane. Furthermore, this protein did not co-localize with the vacuole membrane, as was the case with full-length LegC7/YlfA-GFP. As predicted based on the transposon-mutant insertion data for LegC3-GFP, genetic deletion of the two predicted transmembrane domains of this protein resulted in cytosolic localization and it no longer caused structural changes in the yeast vacuole membrane. Western immunoblots confirmed that the deletion mutants fused to GFP were expressed as full-length proteins of the predicted molecular weight (Figure 6C). These results demonstrate that the hydrophobic regions of these Leg proteins are required for their proper localization in yeast, perhaps by anchoring them to endocytic vesicles.
The LCV is known to recruit vesicles derived from the ER (endoplasmic reticulum) and Golgi complex while excluding markers of early endosomes and lysosomes [21],[64]. The results we obtained in the yeast assays are consistent with LegC3, LegC2/YlfB, and LegC7/YlfA localizing to trafficking organelles and possibly playing a direct role in perturbing normal vacuolar maturation. In order to determine if similar phenotypes are observed in higher eukaryotes, each gene was transiently expressed in mammalian CHO cells [65] as a C-terminal GFP fusion protein. Unfused GFP was found throughout the cytoplasm. In contrast, LegC3-GFP formed multiple punctate structures, although some degree of cytoplasmic expression was also evident (Figure 7). Concordant with the results obtained in yeast cells, the formation of punctate structures was dependent upon the hydrophobic domain of LegC3, as deletion of this domain resulted in a diffuse cytoplasmic distribution. To determine if LegC3-GFP expression was associated with a specific compartment of the endocytic pathway, immuno-fluorescent staining was performed using antibodies against ER (α-KDEL) and lysosomal (α-LAMP-1) markers. As shown in Figure 7, the LegC3-GFP structures had no obvious co-localization with KDEL-containing proteins and excluded the lysosome-associated marker LAMP-1.
In contrast to LegC3-GFP, LegC2/YlfB-GFP expression resulted in the formation of large vacuole-like compartments. Interestingly, these vacuole-like compartments resembled the LCV in their strong co-localization with KDEL-containing proteins and exclusion of LAMP-1 (Figure 8). Similarly, expression of LegC7/YlfA-GFP localized within vacuole-like compartments, although the relative size of these compartments was smaller than that seen with LegC2/YlfB-GFP. Consistent with the idea that LegC2 and LegC7 may have similar functions, the LegC7/YlfA-GFP-containing compartments also co-localized with KDEL proteins and did not co-localize with LAMP-1.
Due to the dramatic and novel effects of LegC3-GFP expression on the yeast vacuole, we wanted to determine the effect of LegC3 expression in a protozoan host of L. pneumophila, D. discoideum. We took advantage of both epifluorescence microscopy and transmission electron microscopy (TEM) to examine cells expressing GFP-LegC3-His7X and GFP-His7X, each under the control of a tetracycline-regulated promoter. When examined by epifluorescence microscopy, cells expressing GFP-LegC3-His7X exhibited both vesicular and lamellar staining patterns, whereas a diffuse cytosolic staining pattern was seen with the GFP-His7X control (Figure 9A). We next isolated a population of cells that were GFP positive using a fluorescence activated cell sorter and processed them for TEM using standard techniques for osmium tetroxide staining. As shown in Figure 9B, cells expressing GFP-His7X alone had normal fine structures and vesicles. Endosomes containing undigested (labeled 1), partially digested (labeled 2) and fully digested material (labeled 3) are visible. In contrast, cells expressing GFP-LegC3-His7X presented an abnormally large number of vesicles with undigested material (labeled 4). Furthermore, structures that appeared to be un-fused pro-lysosomes can also be seen in these cells (labeled 5). This phenotype was observed in at least 50% of the 100 cells analyzed and in less than 4% of the cells expressing GFP-His7X alone (Figure 9C). These results suggest that expression of GFP-LegC3-His7X may affect the steady state formation and degradation of endosomal contents in D. discoideum.
Professional phagocytes have developed pathways dedicated to killing internalized microorganisms by rapidly fusing phagosomes with acidic and hydrolytic vesicles known as lysosomes. While the fate of most bacteria that encounter phagocytic cells is death and degradation, some have evolved mechanisms to avoid this pathway. Many pathogens simply prevent phagocytosis, some puncture the phagosome and escape into the cytosol, yet others, such as L. pneumophila, survive by blocking phagosome maturation altogether. However, the formation of a replication-competent compartment is quite involved; L. pneumophila modifies its phagosome in many different ways to transform it into a nutrient-sufficient, non-acidic and non-hydrolytic environment. This process is believed to require the delivery of specific effector proteins into host cells. It is therefore important to identify and characterize the mechanism of action of these effectors to understand how Legionella pneumophila replicates inside host cells. Previously, we reported the identification of 45 genes in L. pneumophila predicted to encode proteins with distinct eukaryotic motifs (leg genes) that may have been acquired via interdomain horizontal gene transfer [43].
In this study, we set out to determine which of the leg genes actually encode translocated effector proteins. We found that 25 Leg proteins are substrates of the L. pneumophila Icm-Dot translocation system, in addition to 8 previously identified Leg substrates. While several L. pneumophila effector proteins have been found to interact with the early secretory machinery (such as RalF, LidA and DrrA/SidM), there is considerable lack of information on what effector proteins may promote a block in phagosome maturation. Here we show that three effector proteins, LegC2/YlfB, LegC7/YlfA and LegC3, may contribute to this process. When expressed in yeast, these predicted transmembrane proteins seem to block the pre-vacuole compartment, preventing endosomes from maturing and fusing to the vacuole. This can be seen by the accumulation of the styryl dye FM4-64 in a compartment that contains these effector proteins and that are associated with the vacuole. A block in endocytosis is further supported by the observation that LegC7/YlfA and LegC3 cause mis-sorting of CPY-Invertase, a protein that normally trafficks to the vacuole. Interestingly, LegC2/YlfB, a paralog of LegC7/YlfA, did not cause a detectable CPY-Invertase defect despite its vacuole-associated localization within cells. It is possible that LegC2/YlfB is located within membranous structures but lacks an appropriate binding partner required for disrupting vacuolar maturation. Moreover, CPY is only one of many cargos that traffic to the vacuole, and certain cargos are delivered with varying genetic requirements [66]. It will be interesting to determine the exact nature of the effect of these proteins in the endocytic network and if they perform similar roles during infection of natural host cells.
Although we did not observe a growth defect in cells that express either LegC7/YlfA or LegC2/YlfB, these genes were previously identified based their lethality in a similar yeast-based assay [36]. This discrepancy may be explained by the different vectors utilized; the previously utilized vector, pJG4-5, is a high-copy plasmid that creates N-terminal fusions to a nuclear localization signal and a B42 transcriptional activation domain, which may cause effector proteins to be expressed at different levels and localize to different organelles. The level of heterologous gene expression can thus have important consequences on the effects of recombinant L. pneumophila proteins in yeast-based screens. Our finding that LegC5, LegC8 and LepB caused growth defects in yeast may thus still prove to be interesting and useful in finding out which cellular pathways they affect. Mutations in some VPS genes, for instance, cause varying ranges of growth defects in yeast (personal communications, Scott Emr). It is intriguing that LepB, a protein that may have a role in the release of the bacteria from protozoan (but not mammalian) hosts, strongly inhibits growth in yeast. LepB has recently been shown to also act as a GTPase-activating protein functioning to remove Rab proteins from the LCV [27]. The lethal effect of LepB expression may open the way for the study of the effects of this effector protein on exocytosis pathways in yeast.
One caveat to yeast-based screens is that overexpressed proteins, particularly those carrying predicted transmembrane or hydrophobic domains, may produce artificial results. However, a recent study has shown that the overproduction of integral-membrane proteins in yeast seems to cause effects that are mainly protein-specific [67]. In addition, we only observed phenotypes for a small number of effector proteins, even though many contained hydrophobic domains. One notable conclusion from our studies is that only effector proteins that contain coiled-coil domains presented phenotypes in the assays we utilized. This protein-binding domain is often found in eukaryotic trafficking proteins and in many VPS proteins in yeast. This raises the possibility that effector proteins with other domains either affect pathways we did not test, or do not function in yeast.
TEM analysis of D. discoideum cells expressing GFP-LegC3-His7X revealed fine structure modifications that are difficult to observe using other types of microscopy. Here we show that expressing this protein in D. discoideum causes the accumulation of many vesicles with undigested matter. This is of direct relevance to L. pneumophila's lifestyle, since this organism is known to alter membrane trafficking in such a way that endocytic maturation is locally halted. One complicating factor, nonetheless, is the finding that a certain portion of the overexpressed LegC3 seems to remain localized to lamellar structures that may be ER. Others have demonstrated that the overexpression of certain proteins in D. discoideum may lead to the accumulation of aggregates in the ER cisternae [68]. However, the TEM pictures of these aggregated ER cisternae are quite distinct from the TEM images of cells expressing GFP-LegC3-His7X. Furthermore, the accumulation of proteins in the ER does not interfere with endocytic vesicles in the previous studies. The dramatic effects that LegC3 causes in both yeast vacuolar membranes and D. discoideum endosomes suggest that this effector plays an important role in L. pneumophila's ability to arrest phago-lysosome fusion.
Following ectopic expression of LegC3, LegC2/YlfB, and LegC7/YlfA as GFP fusion proteins in mammalian CHO cells, specific intracellular localization patterns were observed. LegC3-GFP was concentrated within multiple distinct, punctate structures and this localization was dependent on the presence of the hydrophobic, putative transmembrane domain. The LegC3-containing structures did not appear to co-localize with either the ER or early secretory vesicles (based on α-KDEL staining) or late endosomes/lysosomes. As such, we cannot currently determine whether LegC3 is specifically targeted to a specific organelle or pathway in mammalian cells. It is interesting to note, however, that the punctate localization of LegC3 within mammalian cells is similar to that observed in the yeast screen, suggesting that LegC3 association with membranes plays an important role in its function in each cell type.
In contrast to what was observed with LegC3-GFP, ectopic expression of LegC2/YlfB-GFP and LegC7/YlfA-GFP in mammalian cells induced the formation of and appeared within large vesicle-like structures. In each case, these structures co-localized with KDEL-containing proteins and excluded the lysosomal marker LAMP-1, concurring with previous results reported for LegC7/YlfA [36]. Notably, these vacuole-like structures are similar to the LCV in their association with ER-resident proteins and exclusion of LAMP-1, suggesting that each protein is sufficient to remodel vacuolar trafficking, perhaps to benefit the intracellular survival of L. pneumophila.
Drawing conclusions about individual effector proteins based on their ectopic expression in the absence of the infectious organism may be regarded with skepticism. However, we believe this is a useful strategy to assist in determining the function and significance of effector proteins in the infectious process. Single or even multiple deletions of putative effector proteins rarely have any observed effect on the ability of L. pneumophila to infect and replicate in host cells. In fact, although our results might suggest a significant role for LegC2/YlfB and LegC7/YlfA in modulating the LCV, strains deleted for both genes replicated in protozoan and mammalian host cells with similar kinetics to wild type [36]. The lack of an intracellular growth phenotype in single or double knockout strains is presumed to be due to a redundancy within the pool of translocated effectors, where multiple proteins may target the same pathway(s). The extensive set of homologous effectors present in L. pneumophila may be explained by their origin. If effectors were originally acquired via horizontal gene transfer and then duplicated over time, this set of “foreign” genes could serve as a pool for the selection of effectors whose functions are fine-tuned for specific interactions and/or specific host species. What may first look like “redundancy” may in fact be a source of untapped resources that allows this pathogen to evolve and survive in several distinct niches. The genome sequence of L. pneumophila may thus represent just a snap-shot of effector genes during the process of evolution.
At present, it is difficult to determine which host pathways are appropriated by specific effector proteins. Nevertheless, the importance of the Icm-Dot TFSS for the virulence of L. pneumophila strongly supports the belief that effector proteins translocated into host cells are required to prevent normal bactericidal responses. Our results demonstrate that LegC3, LegC2/YlfB, and LegC7/YlfA can affect normal vacuolar trafficking in yeast, protozoan, and mammalian cells. The data presented here suggest that endosomal trafficking and particularly the Vps/ESCRT pathway may play a role in the construction or maintenance of the Legionella-containing vacuole. Additional work should provide information about the involvement of this or other conserved host cell pathways in the intracellular lifestyle of L. pneumophila.
The bacterial and yeast strains used in this study are listed in Table 2. With the exception of KS79, bacterial strains, media and antibiotics are described elsewhere [43]. KS79 is an isogenic ΔcomR derivative of JR32. This strain was generated by the isolation of a Tn903dIIlacZ insertion in the comR gene followed by the removal of the entire ORF, including the transposon, to generate an unmarked comR mutant. Generation of the yeast strain NSY01 is described elsewhere [38].
Descriptions of all cloning strategies are described below, and a complete list of plasmids used in this study is provided in Table 1.
The blaM gene encoding the mature form of TEM-1 beta-lactamase (residues 24–286) was amplified by PCR from pUC18 with forward and reverse primers introducing a ribosome-binding site and a KpnI-SmaI-BamHI-XbaI polylinker at the 3′ end. The PCR product was cloned into EcoRI/HindIII digested pMMB207C [33] to generate plasmid pXDC61. The leg genes were PCR-amplified and cloned in frame with the beta-lactamase at the KpnI-XbaI sites. The leg genes that contain one of these restriction sites were cloned into 76XbaI (legA10-S2-T), KpnI (legC7-C8-U1) or BamHI site (LegK1). The resulting plasmids were then introduced by natural transformation into KS79 (JR32 ΔcomR), KS79 dotA::Tn903dIIlacZ or KS79 icmT::Tn903dIIlacZ.
The low-copy, galactose inducible plasmid pBM272 is described elsewhere [60]. The gfp S65T variant was amplified by PCR with primers containing the restrictions sites HindIII/SalI. The PCR product was double-digested with these enzymes and cloned into the HindIII/SalI sites of pBM272 to generate pKS84. 30 leg genes and LepA and LepB were amplified by PCR with primers containing the appropriate restriction sites and inserted either into the BamHI site or HindIII site of pKS84 to generate leg-GFP fusions. To create un-fused version of these proteins, legC2, legC3, legC5, legC7, lepA and lepB were also cloned into pBM272 using identical restriction sites. These plasmids were introduced into NSY01 by the Lithium Acetate transformation method [69].
pKS138 was constructed by inserting a DNA fragment encoding a His7x tag into the Sph1 and Mlu1 sites of pMB38-GFP (a generous gift from Barbara Weissenmayer at the Berlin University). legC3 was amplified by PCR with primers containing BamHI restriction sites subsequently digested with the appropriate enzymes. The product was inserted into the BamHI site of pKS138. Plasmids pKS138 and pKS140 were transformed into D. discoideum AX2 MB35 by electroporation.
For eukaryotic expression experiments, legC3, legC3ΔTM, legC2, and legC7 were amplified by PCR using oligonucleotides that introduced an NheI site and a Kozak consensus sequence (gccgccaccATGgtg) immediately before each gene at the 5′ end and a KpnI site at the 3′. PCR products were sub-cloned into pEGFP-N1 (Clontech) at the NheI and KpnI sites using unidirectional cloning to generate translational Leg-GFP fusion proteins under the transcriptional control of the CMV I/E promoter.
J774 cells grown in RPMI containing 10% fetal calf serum (FCS) were seeded in black clear-bottom 96-well plate at 1x106 cells/well 24 hours prior to infection. L. pneumophila strains carrying the various blaM fusions were grown on CYE plates containing choramphenicol and single colonies were then streaked on CYE plates containing chloramphenicol and 0.5 mM Isopropyl β-D-1-thiogalactopyranoside (IPTG) and grown for 24 hours to induce expression of the hydrid proteins. 10 µL of bacteria re-suspended in RPMI at 5×108 cells/mL were used to infect J774 cells (MOI = 50). After centrifugation (600g, 10 minutes) to initiate bacterial-cell contact the plate was shifted to 37°C and incubated for one hour with CO2 exchange. Cell monolayers were loaded with the fluorescent substrates by adding 20 µl of 6x CCF4/AM solution (LiveBLAzer-FRET B/G Loading Kit, Invitrogen) containing 15 mM Probenecid (Sigma). The cells were incubated for an additional 2 hours at room temperature. Fluorescence was quantified on a Victor microplate reader (Perkin-Elmer) with excitation at 405 nm (10-nm band-pass), and emission was detected via 460-nm (40-nm band-pass, blue fluorescence) and 530-nm (30-nm band-pass, green fluorescence) filters. Translocation was expressed as the emission ratio at 460/530 nm to normalize the ß-lactamase activity to cell loading and the number of cells present in each well. The presented data are mean values of the results from triplicate wells from two to three experiments. The cells were also visualized by fluorescence microscopy using an inverted microscope equipped with the Beta-lactamase ratiometric filter set (Chroma).
S. cerevisiae cells expressing effector proteins were grown overnight in 3 mL cultures at 30°C in SC-Ura/fructose. The cultures were then back-diluted to A600 of 0.4–0.6 in 3 mL of SC-Ura/galactose and grown to A600 of 0.8–1.2. For vacuolar staining, the induced cultures were centrifuged and re-suspended in YP medium containing 2% galactose. N-(3-triethylammoniumpropyl)-4-(p-diethylaminophenylhexatrienyl)-pyridinium dibromide, FM4-64 (Molecular Probes), was added to a final concentration of 40 µM. After 15 minutes of incubation at 30°C, the cells were washed twice and re-suspended in 3 mL of fresh YP medium containing 2% galactose. We allowed cells to incubate for an additional 45–60 minutes before visualization. We visualized cells under a Nikon Eclipse TE200 at 100x using an oil immersion phase-contrast objective. For fluorescence microscopy, we utilized filter sets for fluorescein isothiocyanate and Texas red. For Z-stack image acquisition, we used a Hamamatsu digital camera with a computer-controlled Z-axis drive. We acquired approximately 20 Z-sections every 0.3–0.4 µm (spanning 6–8 µm) and performed volume deconvolution using Improvision's Open Lab software.
The quantitative and qualitative invertase assays were based on a previous published methodology [59]. For the qualitative assay, cells streaked on SC-Ura/galactose or SC-Ura/fructose plates were incubated for 4 days at 30°C. The plates were overlaid with 0.75% agar solution containing 125 mM sucrose, 100 mM sodium acetate buffer (pH 5.5), 0.5 mM N-ethylmaleimide (NEM), 10 µg/mL horseradish peroxidase, 8 units/mL glucose oxidase, and 2 mM O-dianisidine. After 5–15 minutes, pictures were taken with a Sony Cybershot camera. Image contrast was adjusted using the open source software GIMP. For the quantitative assays, cells were grown in liquid SC-Ura/galactose to stationary phase and split into two samples for measurement of total invertase activity and exogenous invertase activity. The assay was scaled down and performed in 96-well plates. To measure total invertase activity, one of the samples was first lysed by the addition of Triton X to a final concentration of 2% followed by 4 cycles of freeze-thaw. 20 µL of each sample was then mixed with 20 µL of a 0.1 M Sodium Acetate (pH 4.9) solution and 5 µL of a 0.5 M sucrose solution followed by incubation at 30°C for 30 minutes. To inactivate the invertase enzyme, 30 µL of 0.2 M K2HPO4 (pH 10.0) was added followed by heating to 95°C for 10 minutes. The samples were then allowed to cool down to room temperature and mixed with 150 µL of the glucostat reagent [59]. The reactions were stopped by the addition of 200 µL of 6M HCl. The A540nm of all samples was measured using a microplate reader (Molecular Device). The units of invertase enzyme were calculated using a standard glucose curve and the following definition: one unit of invertase is the amount of enzyme that hydrolyzes sucrose to produce 1 µM of glucose per minute at 30°C.
The LegC3 mutant library was generated using the GPS-LS Linker Scanning kit (New England Biolabs). We followed the manufacturer's protocol to generate a collection of 92 transposon-generated mutations in the LegC3 gene. On average, 2/3 of the mutants generated have 15 bp insertions in the gene of interest while 1/3 of them create truncated products. We sequenced the whole library (Genewiz, inc.) to determine the exact nature and location of each mutation. The 92 plasmids were then transformed into NSY01 for further analysis. The LegC3-1-GFP mutant protein has an insertion at amino acid position 390 of VFKQS and LegC3-2 an insertion at amino acid position 395 of CLNTF.
Yeast strains were grown on SC Ura/Fructose plates for two days at 30°C. Several colonies were picked and spread on SC Ura/Galactose plates and incubated for three days at 30°C. Yeast cells were scraped with 10 mL of 50 mM Tris-HCl pH 7.5 and then centrifuged at 10,000 rpm for 5 minutes. The cells were then re-suspended in 2 mL of 50 mM Tris-HCl pH 7.5, divided into 0.5 mL aliquots and frozen at −80°C. The frozen pellet was then re-suspended in an additional 0.5 mL of 50 mM Tris-HCl pH 7.5 and the absorbance at 600 nm measured. An amount equivalent to OD 5 in a total volume of 300 µL of Tris-SB was boiled for 5 minutes. The sample LegC7-GFP was concentrated ten times relative to this measurement. 15 µL was loaded onto 4–15% polyacrylamide gradient gels. Following electrophoeresis, the proteins were transferred to a nitrocellulose membrane and processed for immunoblots using an affinity-purified rabbit polyclonal antibody to GFP at a 1:2,000 dilution [70].
D. discoideum cells containing pKS138 and pKS140 were grown for three days with slow shaking in a 22°C water bath. The cells were then washed 3 times with an equal volume of SorC buffer. The cells were then re-suspended in HL5 medium and allowed to grow for an additional 6–10 hours at 22°C [71]. We visualized cells under a Nikon Eclipse TE200 at 100× using an oil immersion phase-contrast objective. For fluorescence microscopy, we utilized filter sets for fluorescein isothiocyanate.
D. discoideum cells containing pKS138 and pKS140, grown for three days with slow shaking in a 22°C water bath, were washed 3 times with an equal volume of SorC buffer, resuspended in HL5 medium and grown for an additional 10 hours at 22°C. Cells were spun down at low speed, re-suspended in SorC buffer, prepared for sorting at the Columbia University Cell Sorting facility, and one million GFP+ cells of each strain were sorted. The cells were plunged into ice and prepared for TEM according to the O R Anderson method [72]. An equal volume of suspended cells was added to a 6% TEM grade glutaraldehyde [Ladd 20215] solution containing HL5 and SorC buffer, gently placed on ice for 30 minutes, resedimented by gentle centrifugation, and the pellet was post-fixed with 1 mL of 2% osmium tetroxide solution [Ladd 55090] in cacodylate buffer. The post-fixed cells were sedimented by gentle centrifugation. The pellet was gently resuspended in 0.8% ionagar sol at 42°C by flicking the tube, rapidly pelleted in the agar sol by centrifugation and the enrobed pellet solidified by placing the tube in an ice bath. The enrobed pellet was removed from the tube with a spatula, washed in water, and dissected into 1 mm3 segments. The segments were washed in water, dehydrated in a graded aqueous acetone series, infiltrated with and embedded in low viscosity epon, placed in BEEM capsules containing the epon, and polymerized for 24 hours at 72°C. Ultrathin sections, obtained with a Porter-Blum MT-2 Ultramicrotome fitted with a diamond knife, were collected on uncoated copper grids, post-stained with Reynold's lead citrate, and observed at 60 kV with a Philips 201 TEM.
CHO-FcγRII cells [65] were maintained in α-MEM media containing 10% FBS. Cells were seeded at 2×104 cells/well on 12 mm coverslips and incubated overnight. The following day, cells were transfected with plasmid DNA using Fugene HD (Roche) according to the manufacturers instructions. 16 hours later, cells were fixed with 3.7% PBS-buffered formalin for 20 minutes, washed with PBS, and blocked/permeabilized with 2% BSA in PBS containing 0.1% saponin. All subsequent buffers, including wash buffers, contained 0.1% saponin. Cells were incubated with blocking buffer containing primary antibodies against KDEL (1:200) (Santa Cruz Biotechnology) or UH1, a monoclonal antibody against hamster Lgp-A/LAMP-1 (1:200) (DSHB, University of Iowa) for 1 hour, washed, and incubated with goat anti-mouse Alexa Fluor 564 (1:500) (Invitrogen) for 30 minutes. After washing, coverslips were fixed to glass slides using Vectashield HardSet mounting media (Vector Labs) for confocal microscopy. Confocal images were acquired with a Zeiss LSM510 Meta laser scanning microscope. Captured images were processed and merged using the open-source NIH software, ImageJ (http://rsb.info.nih.gov/ij/index.html).
|
10.1371/journal.pntd.0006118 | Identifying cholera "hotspots" in Uganda: An analysis of cholera surveillance data from 2011 to 2016 | Despite advance in science and technology for prevention, detection and treatment of cholera, this infectious disease remains a major public health problem in many countries in sub-Saharan Africa, Uganda inclusive. The aim of this study was to identify cholera hotspots in Uganda to guide the development of a roadmap for prevention, control and elimination of cholera in the country.
We obtained district level confirmed cholera outbreak data from 2011 to 2016 from the Ministry of Health, Uganda. Population and rainfall data were obtained from the Uganda Bureau of Statistics, and water, sanitation and hygiene data from the Ministry of Water and Environment. A spatial scan test was performed to identify the significantly high risk clusters. Cholera hotspots were defined as districts whose center fell within a significantly high risk cluster or where a significantly high risk cluster was completely superimposed onto a district. A zero-inflated negative binomial regression model was employed to identify the district level risk factors for cholera. In total 11,030 cases of cholera were reported during the 6-year period. 37(33%) of 112 districts reported cholera outbreaks in one of the six years, and 20 (18%) districts experienced cholera at least twice in those years. We identified 22 districts as high risk for cholera, of which 13 were near a border of Democratic Republic of Congo (DRC), while 9 districts were near a border of Kenya. The relative risk of having cholera inside the high-risk districts (hotspots) were 2 to 22 times higher than elsewhere in the country. In total, 7 million people were within cholera hotspots. The negative binomial component of the ZINB model shows people living near a lake or the Nile river were at increased risk for cholera (incidence rate ratio, IRR = 0.98, 95% CI: 0.97 to 0.99, p < .01); people living near the border of DRC/Kenya or higher incidence rate in the neighboring districts were increased risk for cholera in a district (IRR = 0.99, 95% CI: 0.98 to 1.00, p = .02 and IRR = 1.02, 95% CI: 1.01 to 1.03, p < .01, respectively). The zero inflated component of the ZINB model yielded shorter distance to Kenya or DRC border, higher incidence rate in the neighboring districts, and higher annual rainfall in the district were associated with the risk of having cholera in the district.
The study identified cholera hotspots during the period 2011–2016. The people located near the international borders, internationally shared lakes and river Nile were at higher risk for cholera outbreaks than elsewhere in the country. Targeting cholera interventions to these locations could prevent and ultimately eliminate cholera in Uganda.
| Uganda has regularly reported cholera since its first appearance in 1971. Although the Government of Uganda implements cholera prevention and control interventions such as provision of safe water, promotion of sanitation and hygiene, health education and healthcare, the disease continues to threaten many districts in the country. The population with access to improved water supply in the urban and rural areas were 71% and 67% respectively, and with access to improved sanitation was 86% in urban areas and 79% in rural areas. Identifying the districts with increased risk is an important step in defining areas where additional preventive interventions are needed. We used district level confirmed cholera outbreak data for a six year period (2011–2016), and identified cholera “hotspot” districts. Rates of cholera in these districts, with a population of about 7 million, are 2 to 22 times higher than elsewhere in the country. These “hotspots” located along the international borders with Democratic Republic of Congo (DRC) and Kenya and the internationally shared lakes and river Nile. Targeted cholera prevention and control interventions to the hotspots in Uganda could lead to reduction in cholera cases and deaths. The hotspots identified herein provide an affordable way of implementation of the comprehensive cholera prevention and control mechanism as recommended by WHO. Since the hotspots are along the international borders, collaboration with the neighboring countries is the key to eliminate cholera in Uganda and the region as a whole.
| Since 1970, cholera has become endemic in many countries of sub-Saharan Africa and remains a recurring problem [1]. During 2007 and 2011, annual Case-Fatality Ratios (CFRs) for cholera within this region ranged from 2.22% to 2.95%, and exceeded 5% in a country in each year [2–6]. It is evident that in spite of continued scientific advancement in prevention and treatment of cholera, it remains a major public health issue in many parts of sub-Saharan Africa.
Cholera incidence and mortality, and diarrheal diseases, can be reduced through increased access to safe water, sanitation and hygiene (WaSH) and through behavioral change resulting from education and training in communities [7] as was done in Latin America [8, 9] and in industrialized countries [10]. However, improvements in WaSH infrastructure have been slow in many cholera affected African countries [11, 12], resulting in repeated cholera outbreaks.
According to WHO, cholera prevention and control should be a priority in areas at risk for cholera, and vaccines should be included in conjunction with other cholera prevention and control strategies [13]. Safe WHO prequalified cholera vaccines (OCV) [14] provide 65% protection at least for 3 to 5 years [15]. Studies have also demonstrated that large-scale vaccination campaigns are feasible and acceptable, reflected by high acceptance rate and high coverage in both urban and rural settings [16–20]. To ensure access to OCV for cholera affected countries, a global stockpile is now available for both emergency use to control outbreaks and areas with recurrent cholera outbreaks. However, the supply of OCV is limited compared to the large number of people living in countries at risk [21]. Thus, countries with cholera need to identify specific areas at increased risk in order to focus their control efforts using an integrated approach with WaSH improvements and vaccination.
In Uganda, epidemics of cholera have occurred regularly since the disease first appeared in 1971 [22]. The Government of Uganda instituted preventive and control measures that included promotion of access to safe water, sanitation and hygiene; health education and community mobilization; disease surveillance; and case management [23]. However cholera cases continues to be reported annually [24]. It is important for the country to understand the extent of the problem, identify the areas of high risk, and use this information to plan for an effective intervention strategy. In most areas, during cholera outbreaks, cases tend to be clustered in specific areas and among certain population groups [25, 26], which can be demonstrated by the spatial epidemiology of the disease. A study that describes the spatial distribution of disease, its incidence using geographic information system (GIS) and its association to potential risk factors should help guide interventions to control cholera [27]. The aim of this study was to identify cholera high risk districts (hotspots) in Uganda in order to provide insights and guidance for prevention, control and ultimately elimination of cholera in Uganda.
Uganda, is divided into four geographical regions (Central, Eastern, Western, and Northern) with a projected total of 36.6 million people as of June 2016, and annual growth rate of 3.0 percent. There were 116 districts according to June 2016 data. However, in this study we used the 2014 census with 112 districts (S1 Fig). The country has tropical weather conditions moderated by the high altitudes. The Central, Eastern, and Western regions of the country have two rainy seasons per year, with heavy rains from March to early June and light rains between September and December. The level of rainfall decreases toward the north, with just one rainy season from April to October [27]. Uganda’s vegetation varies between tropical rain forest in the south and the savannah woodlands and semi-desert region in the northeast [28]. The population density is higher in the Central and Western regions and declines toward the North [29].
District level confirmed cholera outbreak data from 2011 to 2016 were abstracted from the Uganda Ministry of Health, Health Management Information System disease surveillance database. In order for the districts to confirm cholera outbreaks, 10–20 stool samples are collected from suspected cholera cases and sent for culture to a laboratory in Kampala for isolation of Vibrio cholerae O1 and O139. If a sample is found positive then it is defined as a confirmed outbreak. The district health workers are guided by the following standardized case definition for cholera case detection and confirmation below [30]:
Suspected case:
Confirmed case:
In this study, we included only those cases for which the outbreaks were confirmed as per national standard guidelines for cholera prevention and control.
The population data were obtained from the Uganda Census 2014 posted in the website of the Ministry of Internal Affairs, Uganda Bureau of Statistics (UBOS) 2016 [31]. We used UBOS definition to categorize areas into urban and rural. According to UBOS, the City, Municipality, Town Council or Town Board are the as urban areas. In 2014, Uganda had 197 urban centers (one City, 22 Municipalities and 174 Town Councils) with a total population of 6 million urban areas. The size of the urban centers varied widely, from Kampala City with 1.5 million persons to small Town Councils with less than 5,000 persons.
We obtained annual rainfall data of 2011–2015 from the 2016 Statistical Abstract of the UBOS (http://www.ubos.org/onlinefiles/uploads/ubos/statistical_abstracts/2016%20Statistical%20Abstract.pdf). The data were available from 8 stations throughout the country. We linked the districts to its nearest station, and obtained the rainfall data for the period. We, then, calculated average annual rainfall for each of the districts.
The water, sanitation and hygiene data were obtained from the 2016 Uganda Ministry of Water and Environment Annual Report [32]. In this report, improved water supply sources included the boreholes, protected springs, shallow wells, and rainwater harvesting tanks. Improved piped water supply outlets included public stand posts, yard taps, kiosks, house (domestic) connections and institutional connections. Water for production facilities (dams and valley tanks) were not regarded as improved water supplies for domestic use. The population with access to improved water supply in the urban and rural areas were 71% and 67% respectively. The was achieved due to back-up support for operations and maintenance provided by regionally based Umbrella Organizations (Central Umbrella based in Wakiso; Mid-Western Umbrella based in Kyenjojo; South Western Umbrella based in Kabale; Northern Umbrella based in Lira; Eastern Umbrella based in Mbale; and Karamoja Umbrella based in Moroto) [33].
Improved sanitation (defined as not shared, cleanable, sealable and durable) was reported to be 86% in urban areas and 79% in rural areas [32]. However, most of the latrines did not meet the standards of the WHO/UNICEF Joint Monitoring Program (JMP), which estimated that only 35% of the rural people in Uganda had access to improved sanitation, with an estimated 10% practicing open defecation. The JMP also estimated that only 34% of the urban population has access to improved sanitation as half of the urban population uses improved but shared facilities. One percent of the urban population was estimated to practice open defecation, while 15% use unimproved facilities [33].
We also collected district level data on a sanitation and hygiene benchmark (the higher the benchmark score the better the sanitation and hygiene condition) from the same report [33]. This was calculated as the sum of scores of the following items:
The digital maps of Uganda were obtained from the Energy Sector GIS Working Group Uganda, which is an open data (http://data.energy-gis.opendata.arcgis.com/datasets/f0d63758fb8f4ded85394b51594d294a_0). The digital map of the health facilities in Uganda was abstracted from a free, open, collaborative platform for creation and maintenance of geocoded health facility master list (https://healthsites.io/). The other geographic features such as road, railway, and waterbodies were obtained from DIVA-GIS (http://www.diva-gis.org/Data), an open source platform. These maps were projected in WGS 1984 UTM zone 36S for conducting spatial analysis. We used ArcGIS 10.4.1 (ESRI Inc.) for mapping the cholera risk in the country and obtaining spatial variables such as distances to the nearest lake and the nearest health facility from the centroid of each of the districts. We also calculated population density (in km2) using total population in the district divided by the size of the area of the district.
We used SaTScan version 9.4.4 (http://www.satscan.org/) to identify cholera hotspots in Uganda. In particular, we employed the Poisson-based spatial scan statistic, because the cases in the districts follow a Poisson distribution. The population in the districts were adjusted in detecting the spatial clusters (hotspots). Under the Poisson model, we assumed the expected number of cases in each part of the study area is proportional to its population size. The model detected clusters in a multidimensional point process and allowed variable window sizes to scan for the cases within the study area. The variable window size was chosen, as we did not have prior knowledge about the size of the area covered by a cluster. We selected circular scan window, and the radius of the window was chosen to vary from 0% to 20% of the population at risk. We started with 0% due to use of the center coordinates of the districts as the geographic reference point. The clusters indicate areas with significantly higher rates inside the window compared to that outside the window. Since the location and size of the window were changed during the operation, it created an infinite number of distinct geographical circles. Therefore, computing the number of points at any given time was not possible [34], leading us to calculate the likelihood ratio. Under the Poisson model, the likelihood function for a specific window is:
λ=(nμ)n(N−nN−μ)N−nI(n>μ)
where,
Since we scanned for clusters with only the high rates, I() is 1 when the window had more cases than expected under the null hypothesis, and in all other cases it was 0. The likelihood function was maximized over all windows, identifying the window that constituted the most likely cluster. The most likely cluster (hotspots) is the area that is least likely to have occurred by chance. The likelihood ratio for the window was noted and constituted the maximum likelihood ratio test statistic. Its distribution under the null hypothesis and its corresponding p-value were determined by repeating the same procedure on a large number of random replications of the data set generated under the null hypothesis using a Monte Carlo simulation approach.
We calculated distance (in kilometer) from the centroid of the district to the nearest hospital (in linear distance), nearest DRC or Kenya border, and the nearest bank of lake or river. To calculate weighted incidence rate of the 1st order of neighbors for each district, we first created a spatial weight matrix for each district considering 1st order of neighbor districts using a GIS tool and then calculated weighted incidence rate of the neighbor districts using cholera cases over the six year period and the population in those districts. Note that we chose 1st order of neighbor districts, because we believe that cholera transmission would less likely to go beyond the 1st order of neighbor district.
To evaluate whether the risk for cholera is associated with water and sanitation conditions as well as other spatial characteristics of the districts, we used the zero-inflated negative binomial (ZINB) model. The ZINB model was chosen because our data were a two-component mixture composed of at-risk districts whose responses (i.e. cholera cases) follow Poisson process with overly dispersed (variance exceeds the mean), i.e., negative binomial (NB) and non-risk districts whose responses are constant, i.e., zero [35]. With probability π, the response of the first process is a zero count, and with probability of (1 − π) the response of the second process is governed by a NB with mean λ, which also generates zero counts. The overall probability of zero counts is the combined probability of zeros from the two processes. Thus, a ZINB model for the response Y can be written as:
P(Y=0)=π+(1−π)(1+kλ)−1/k
P(Y=y)=(1−π)Γ(y+1k)(kλ)yΓ(y+1)Γ(1k)(1+kλ)y+1/k,y=1,2,….
District level characteristics along with the number of cholera cases in the districts were used in this analysis. Initially, we created bivariate models taking account of each variable in the model along with the number of cholera cases as the dependent variable. This was followed by multivariate analysis for those variable found to have association with the outcome at p<0.20 in the bivariate model. SAS version 9.4 was used to analyze the risk factors for cholera. We also calculated cumulative incidence rate over the six-year period and the coefficient of variation of the incidence rate based on the year-wise incidence rate of a district.
The study used secondary data aggregated at the district level, thus it did not require any ethical review board approval.
A total of 11,030 cases of cholera were reported during 2011–2016. The highest number of cases were 6,226 in 2012 and the lowest were 229 in 2011 (Fig 1). In 37 of 112 districts (33%) cholera was reported in at least one of the study years. These districts made up 40% of the total population of Uganda. During the study period, the highest burden of cholera was in Nebbi district (2,320 cases) followed by Hoima (1,731 cases), and Buliisa (1,129 cases). The districts that had multiple outbreaks (at least two years) during study period had a population of about 7.8 millions.
The results of the SatScan for cluster detection based on the centers of the districts as the geographic coordinates yielded 16 significantly high-risk clusters of different sizes. Cholera hotspots were defined as districts whose center fell within a significantly high risk cluster or where significantly high risk cluster was completely superimposed onto a district. There were 22 districts whose centroids fell within the cluster detected by SatScan; these were defined as high risk districts. The risk of having cholera in these districts was 1.26 to 21.50 times compared to that elsewhere in the country (Table 1, Fig 2). About 7 million people live in these districts. Of the 22 districts, 13 of them (4.8 million people) are near the border of DRC and 9 of them (2.2 million people) are near the border of Kenya.
We observed a distinct seasonal pattern of cholera between eastern and western region of Uganda. The seasonal pattern also varied by year. There was no cholera outbreak in the eastern region in 2011 and 2013. In 2012, the peak was between March and May in the eastern region, while in the western region the peak was between April and August (Fig 3). Apparently, there was only a little link with rainfall in a specific year and specific region (Fig 3). Fig 4 shows geographic distribution of the rates of cholera which varied in the study area, illustrating that cholera affected the people living near a lake or river. The coefficient of variations for the year-wise rates indicates that border districts experienced cholera more frequently than the inner part of the country (S2 Fig).
The basic statistics of the study variables show a high rate of improved water (82%) and sanitation (75%) coverage at the household level (Table 2). An average of 15% of the population lived in the urban areas. There were 300 hospitals/clinics for Uganda in the world health facility database, and these health facilities were heterogeneously distributed (visual perception) in space (Fig 4). The average distance to the nearest health facility, river of lake, and border of DRC or Kenya from the center of the district was 33, 47, and 91 kilometers, respectively. The average weighted cholera incidence rate (cumulative over the six year period) of the 1st order of neighbors of a district was 28 per 100,000 persons. Average annual rainfall during 2011–2015 was 1324 mm.
The result of the bivariate analysis of the ZINB are presented in Table 3. The NB component of the model shows that higher incidence rate in a district was significantly associated (p < .05) with improved water sources, longer distance to the nearest heath facility, shorter distance to a lake or river, shorter distance to DRC or Kenya border, and a higher cholera incidence rate of the neighbor districts. On the other, although a significant relationship of population density was observed with reporting of cholera case by a district in the zero-inflated model, the point estimate does not indicate an increase or decrease of having cholera by the district. Likewise the NB model, distance to DRC or Kenya border and cholera incidence rate of the neighbor districts yielded a significant relationship (p < .05) with reporting of cholera cases in the zero-inflated model.
We created multivariable model using the variables associated with the outcome at p<0.20. The results of the NB component of the multivariable model show that the distance to the nearest lake or river was significantly associated with the incidence of cholera (incidence rate ratio = 0.98, 95% CI: 0.97–0.99, p<0.01) (Table 4). This indicates that a one kilometer decrease in the distance to lake or river was associated with 2 percentage point of increase of cholera cases. Shorter distance to the DRC or Kenya border and higher cholera incidence of the neighbor districts were significantly associated (p < .05) with higher incidence rate of cholera in the district. On the other hand, the zero-inflated component of the multivariable model shows a one kilometer increase of distance to DRC or Kenya border was associated with 2 percentage point decrease in the odds of having cholera in the district. The zero-inflated model also shows that a one percentage point increase in the cholera incidence rate in the neighbor districts was associated with 3 percentage point increase in the odds of having cholera. Higher annual rainfall in the district was significantly associated (p = .03) with the risk of having cholera in the district in the zero-inflated model.
This results of our study show that most of the high risk districts for cholera were near the border with DRC and Kenya. These high risk districts with a population of about 7 million, make up about 20% of the population of Uganda. Those with a very high risk districts (relative risk was more than 10 compared to that elsewhere in the country) have a population of about 2.4 million (7% of the population of Uganda) from six districts. All of these districts are located near the DRC. Studies [36, 37] indicated that cross-border cholera outbreaks could be a major contributor to the recorded outbreaks in Uganda. Since a majority of the hotspot districts are near DRC or Kenya border, it suggests that a close collaboration with these countries would be an effective strategy for controlling cholera in that part of the world.
Our study also shows that proximity to a lake, specifically Lake Albert and Lake Victoria or to the Nile River, creates increased risk of cholera for the people in Uganda, as found in an earlier study [38]. Interestingly, rates of cholera are low near Lake Kyoga and along the Nile River as the water flows between Lake Kyoga and Lake Albert, but the rates are higher along the Nile River as the water flows north from Lake Albert. The finding in our study is consistent with earlier studies, which indicated that people are continuously being affected by cholera in lakeside areas [39]. Several other studies have also reported that Lake Albert and Lake Victoria could be a source of Vibrio cholerae [40, 41]. However, there is no definitive evidence to support the hypothesis. The link between high incidence of cholera and presence of lakes has also been noted in DRC [40, 42]. Bompangue et al [39] believe that in the absence of lakeside areas, the disease would have disappeared from the country. This may be true as we find cholera was less likely to create a threat to the people living far from the lakes. However, it is not clear if the lake is the major risk factor, or the behaviors or occupations of the people who live along the lake [43]. It should be noted that these lake areas also are borders with the neighboring countries with cholera; thus, it is difficult to separate the risk associated with the lake environment and the risk associated with cross border spread. The association between cholera outbreak and rainfall, as observed in our zero-inflated model, is consistent with earlier findings [39].
The seasonal pattern of cholera varied by year in Uganda suggesting exact timing of the outbreaks in the different regions of the country may not be predicted. A spike of cholera was observed in 2012, which was associated with a large outbreak in the African region that mostly affected six nations including Uganda (http://www.who.int/hac/crises/cholera_afro_22august2012.pdf?ua=1). In 2016, cholera showed a two distinct seasonal pattern between eastern and western region of the country. The peak season for cholera in eastern region started in March and continued until May, and the peak season in the western region started in June and lasted until August. Since Uganda experiences heavy rainfall from March to June, it seems that cholera outbreaks are associated with the rainfall in the country. Association of rainfall with the risk for cholera in Africa as well as in the neighboring country Kenya has already been observed [44, 45].
We did not find an association of water and sanitation with the risk for cholera in Uganda. This could be due to a reported high coverage of improved water sources and improved sanitation in the country. Also, the indicator used for the district which is an average for the district, may not represent the water and sanitation status of those who are vulnerable to cholera. With such high coverage for these WaSH indicators, one may want to reassess whether these indicators are providing an accurate reflection of the true situation for those at risk. The sustainability of water services in the small towns is also reported to be high (92%). This high coverage of access to safe water is largely due to back-up support for operations and maintenance which is being provided by regionally based Umbrella Organizations [33]. On the other hand, access to improved sanitation is reported to be high (85%) in the urban areas, and open defecation is reported to be rare in the both urban and rural areas of the country, except slums [33].
The strength of our study is that the cases of cholera were collected from a surveillance system which has been maintained systematically throughout the country and was done according to the WHO standard. Water and sanitation data were obtained from the report prepared by the Ministry of Water and Environment, Uganda, which are being updated on a yearly basis ensuring greater reliability of the data. The population data were obtained from a recent report prepared by the Uganda Bureau of Statistics [31] providing national population and housing census 2014, which is consistent with the period of our study. Additionally, since the data came from the national surveillance conducted by the Uganda Ministry of Health and was systematically executed throughout the country, we believe the burden of cholera across the districts are comparable.
The main limitation in our study is that cholera data were not population-based, which precluded calculating the absolute risk of the disease. Conducting a national level population-based disease surveillance in this setting is probably unrealistic. However, the data used in this study provided a basis for understanding relative burden of cholera across the districts in Uganda.
Our study identified cholera hotspots in Uganda that should be prioritized to accelerate reduction of cholera by implementation of targeted interventions. The intervention program includes provision of safe water, hygiene and sanitation, provision healthcare service and health education, and should be complemented with OCV and strong cross-border collaboration in outbreak prevention. The findings of our study could be used as a guide to strengthen the cholera control program in Uganda.
|
10.1371/journal.pcbi.1006565 | Atlases of cognition with large-scale human brain mapping | To map the neural substrate of mental function, cognitive neuroimaging relies on controlled psychological manipulations that engage brain systems associated with specific cognitive processes. In order to build comprehensive atlases of cognitive function in the brain, it must assemble maps for many different cognitive processes, which often evoke overlapping patterns of activation. Such data aggregation faces contrasting goals: on the one hand finding correspondences across vastly different cognitive experiments, while on the other hand precisely describing the function of any given brain region. Here we introduce a new analysis framework that tackles these difficulties and thereby enables the generation of brain atlases for cognitive function. The approach leverages ontologies of cognitive concepts and multi-label brain decoding to map the neural substrate of these concepts. We demonstrate the approach by building an atlas of functional brain organization based on 30 diverse functional neuroimaging studies, totaling 196 different experimental conditions. Unlike conventional brain mapping, this functional atlas supports robust reverse inference: predicting the mental processes from brain activity in the regions delineated by the atlas. To establish that this reverse inference is indeed governed by the corresponding concepts, and not idiosyncrasies of experimental designs, we show that it can accurately decode the cognitive concepts recruited in new tasks. These results demonstrate that aggregating independent task-fMRI studies can provide a more precise global atlas of selective associations between brain and cognition.
| Cognitive neuroscience uses neuroimaging to identify brain systems engaged in specific cognitive tasks. However, linking unequivocally brain systems with cognitive functions is difficult: each task probes only a small number of facets of cognition, while brain systems are often engaged in many tasks. We develop a new approach to generate a functional atlas of cognition, demonstrating brain systems selectively associated with specific cognitive functions. This approach relies upon an ontology that defines specific cognitive functions and the relations between them, along with an analysis scheme tailored to this ontology. Using a database of thirty neuroimaging studies, we show that this approach provides a highly-specific atlas of mental functions, and that it can decode the mental processes engaged in new tasks.
| A major challenge to reaching a global understanding of the functional organization of the human brain is that each neuroimaging experiment only probes a small number of cognitive processes. Cognitive neuroscience is faced with a profusion of findings relating specific psychological functions to brain activity. These are like a collection of anecdotes that the field must assemble into a comprehensive description of the neural basis of mental functions, akin to “playing twenty questions with nature” [1]. However, maps from individual studies are not easily assembled into a functional atlas. On the one hand, the brain recruits similar neural territories to solve very different cognitive problems. For instance, the intra-parietal sulcus is often studied in the context of spatial attention; however, it is also activated in response to mathematical processing [2], cognitive control [3], and social cognition and language processing [4]. On the other hand, aggregating brain responses across studies to refine descriptions of the function of brain regions faces two challenges: First, experiments are often quite disparate and each one is crafted to single out a specific psychological mechanism, often suppressing other mechanisms. Second, standard brain-mapping analyses enable conclusions on responses to tasks or stimuli, and not on the function of given brain regions.
Cognitive subtraction, via the opposition of carefully-crafted stimuli or tasks, is used to isolate differential responses to a cognitive effect. However, scaling this approach to many studies and cognitive effects leads to neural activity maps with little functional specificity, hard to assemble in an atlas of cognitive function. Indeed, any particular task recruits many mental processes; while it may sometimes be possible to cancel out all but one process across tasks (e.g. through the use of conjunction analysis [5]), it is not feasible to do this on a large scale. Furthermore, it can be difficult to eliminate all possible confounds between tasks and mental processes. An additional challenge to the selectivity of this approach is that, with sufficient statistical power, nearly all regions in the brain will respond in a statistically significant way to an experimental manipulation [6].
The standard approach to the analysis of functional brain images maps the response of brain regions to a known psychological manipulation [7]. However, this is most often not the question that we actually wish to answer. Rather, we want to understand the mapping between brain regions/networks and psychological functions (i.e. “what function does the fronto-parietal network implement?”). If we understood these mappings, then in theory we could predict the mental state of an individual based solely on patterns of activation; this is often referred to as reverse inference [8], because it reverses the usual pattern of inference from mental state to brain activation. Whereas informal reverse inference (e.g. based on a selective review of the literature) can be highly biased, it is increasingly common to use meta-analytic tools such as Neurosynth [9] to perform formal reverse inference analyses (also know as decoding). However, these inferences remain challenging to interpret due to the trade-off between breadth and specificity that is necessary to create a sufficiently large database (e.g. see discussion in [10, 11]).
The optimal basis for brain decoding would be a large database of task fMRI datasets spanning a broad range of mental functions. Previous work has demonstrated that it is possible to decode the task being performed by an individual, in a way that generalizes across individuals [12], but this does not provide insight into the specific cognitive functions being engaged, which is necessary if we wish to infer mental functions associated with novel tasks. The goal of decoding cognitive functions rather than tasks requires that the data are annotated using an ontology of cognitive functions [13–15], which can then become the target for decoding. Some recent work has used a similar approach in restricted domains, such as pain [16], and was able to isolate brain networks selective to physical pain. Extending this success to the entire scope of cognition requires modeling a broad range of experiments with sufficient annotations to serve as the basis for decoding.
To date, the construction of human functional brain atlases has primarily relied upon the combination of resting-state fMRI and coordinate-based meta-analyses. This approach is attractive because of the widespread availability of resting-state fMRI data (from which brain functional networks can be inferred through statistical approaches [17]), and the ability to link function to structure through the use of annotated coordinate-based data (such as those in the BrainMap [18] and Neurosynth [9] databases). This approach has identified a set of large-scale networks that are consistently related to specific sets of cognitive functions [19, 20], and provides decompositions of specific regions [21, 22]. However, resting-state analysis is limited in the set of functional states that it can identify [23], and meta-analytic databases are limited in the specificity of their annotation of task data, as well as in the quality of the data, given that it is reconstructed merely from activation coordinates reported in published papers.
A comprehensive functional brain atlas should link brain structures and cognitive functions in both forward and reverse inferences [7]. To build such a bilateral mapping, we introduce the concept of “ontology-based decoding,”, in which the targets of decoding are specific cognitive features annotated according to an ontology. This idea was already present in [9, 12, 24]; here we show how an ontology enables scaling it to many cognitive features, to increase breadth. In the present case, we use the Cognitive Paradigm Ontology (CogPO) [15], that provides a common vocabulary of concepts related to psychological tasks and their relationships (see S1 Text Distribution of terms in our database). Forward inference then relies on ontology-defined contrasts across experiments, while reverse inference is performed using an ontology-informed decoder to leverage this specific set of oppositions (see Fig 1 and methodological details). We apply these forward and reverse inferences to the individual activation maps of a large task-fMRI database: 30 studies, 837 subjects, 196 experimental conditions, and almost 7000 activation maps (see S1 Text Distribution of terms in our database). We use studies from different laboratories, that cover various cognitive domains such as language, vision, decision making, and arithmetics. We start from the raw data to produce statistical brain maps, as this enables homogeneous preprocessing and thorough quality control. The results of this approach demonstrate that it is possible to decode specific cognitive functions from brain activity, even if the subject is performing a task not included in the database.
The main challenge to accumulate task fMRI is to account for the disparity in experimental paradigms. One solution is the use of cognitive ontologies that define terms describing the cognitive tasks at hand and enable to relate them. The choice of the ontology must meet two opposite goals: have a good coverage of the cognitive space, and document overlap between studies. In practice, each cognitive term describing mental processes must be expressed in several studies of our database to ensure the generalizability of our inference.
Standard forward inference in functional neuroimaging uses the GLM (general linear model), which models brain responses as linear combinations of multiple effects. We use a one-hot-encoding of the concepts, i.e. we represent their presence in the tasks by a binary design matrix. We test for response induced by each concept with a second-level analysis using cross-studies contrasts.
To disentangle various experimental factors, brain mapping uses contrasts. Individual studies are crafted to isolate cognitive processes with control conditions, e.g. a face-recognition study would rely on a “face versus place” or a “face versus scrambled picture” contrast. To separate cognitive factors without a strong prior on control conditions, the alternative is to contrast a term against all related terms, e.g., “face versus place and scrambled picture”.
We use the categories of our ontology to define such contrasts in a systematic way for the wide array of cognitive concepts touched in our database. This approach yields groups of terms within the task categories, as described in Table 1: the task categories are used to define the conditions and their controls. Inside each group, we perform a GLM analysis with all the “one versus others” contrasts. We denote these ontology contrasts. Compared to a standard group analysis, the benefit of this GLM is that the control conditions for each effect studied span a much wider range of stimuli than typical studies.
For reverse inference, we rely on large-scale decoding [12]. Prior work [12, 24] tackles this question using a multi-class predictive model, the targets of the classification being separate cognitive labels. Our formulation is different as our goal is to predict the presence or absence of a term, effectively inverting the inference of our forward model based on one-hot-encoding. This implies that each image is associated with more than a single label, which corresponds to multi-label classification in a decoding setting.
Using a database of 30 studies, we demonstrate that our approach captures a rich mapping of the brain, identifying networks with a specific link to cognitive concepts. Prediction of cognitive components in new paradigms validates this claim.
We combine forward and reverse inference to construct a one-to-one mapping between brain structures and cognitive concepts. Forward inference across studies requires adapting brain mapping analysis to leverage the ontology. Mapping the brain response to the presence of a concept in tasks selects unspecific regions, as it captures other related effects, e.g. selecting the primary visual cortex for any visual task (Fig 3). To obtain a more focal mapping, we remove these effects by opposing the concept of interest to related concepts in the ontology. Reverse inference narrows down to regions specific to the term. However, as we use a multivariate procedure, some of its variables may model sources of noise [26]. For instance, when using visual n-back tasks with a motor response to map the visual system, the motor response creates confounding signals. A multivariate procedure could use signal from regions that capture these confounds to subtract them from vision-specific activity, leading to better prediction. As such regions are not directly related to the task, they are well filtered with a standard GLM (General Linear Model) used in forward inference. For this reason, our final maps combine statistics from forward and reverse inference: functional regions are composed of voxels that are both recruited by the cognitive process of interest and predictive of this process; see S5 Text Consensus between forward and reverse inference for statistical arguments and [27] for more fundamental motivations regarding causal inference. Fig 3a–3d shows how the neural-activity patterns for the “places” label progressively narrow on the PPA with the different approaches. Thus we link each cognitive concept to a set of focal regions, resulting in a brain-wide functional atlas.
To build functional atlases, it is important to clearly identify the regions associated with different cognitive concepts. Fig 3e shows that reverse-inference meta-analysis with Neurosynth also associates the PPA with the “place” term, but the region is not as well delineated as with our approach. Fig 4 shows functional atlases of auditory and visual regions extracted with various mapping strategies. The relative position and overlap of the various maps is clearly visible. Forward-inference mapping of the effect of each term versus baseline on our database gives regions that strongly overlap (Fig 4a). Indeed, the maps are not functionally specific and are dominated by low-level visual mechanisms in the occipital cortex and language in the temporal cortex. Using contrasts helps decreasing this overlap (Fig 4b), and hence reveals some of the functional segregation of the visual system. However, as the stimuli are not perfectly balanced across experiments, contrasts also capture unspecific regions, such as responses in the lateral occipital cortex (LOC) for faces or places. Reverse inference with a logistic-regression decoder gives well separated regions, albeit small and scattered (Fig 4c). The ontology-informed approach identifies well-separated regions that are consistent with current knowledge of brain functional organization (Fig 4d). Finally, meta analysis with NeuroSynth separates maps related to the various terms better than forward analysis on our database of studies (Fig 4e). Yet some overlap remains, for instance in the LOC for maps related to visual concepts. In addition, the outline of regions is ragged, as the corresponding maps are noisy (Fig 3e), probably because they are reconstructed from peak coordinates. Note that overlaps across term-specific topographies are ultimately expected to remain, especially in associative cortices. In the following, we first discuss quantitative validation of the reverse-inference atlases, and then study in detail the atlas obtained with the ontology-informed approach.
Upon qualitative inspection, the regions extracted by our mapping approach provide a good functional segmentation of the brain. For an objective test of this atlas, we quantify how well these regions support reverse inference. For this, we use the ontology-informed decoder to predict cognitive concepts describing tasks in new paradigms and measure the quality of the prediction. This approach was tested using a cross-validation scheme in which 3 studies were held out of each training fold for subsequent testing. Fig 5 shows the corresponding scores: ontology-informed decoding accurately predicts cognitive concepts in unseen tasks. It predicts these concepts better than other commonly used decoders (logistic regression and naive Bayes, see also S6 Text Evaluating prediction accuracy: cross-validation) and NeuroSynth decoding based on meta-analysis. This confirms that the corresponding atlas captures areas specialized in cognitive functions better than conventional approaches.
Very general labels such a “visual” are found in most studies, and therefore easy to predict. However, higher-level or more specialized cognitive concepts such as viewing digits or moving the left foot are seldom present (see S1 Text Distribution of terms in our database). For these rare labels, the fraction of prediction errors is not a useful measure. Indeed, simply assigning them to zero images would lead to a small fraction of errors. For this reason, Fig 5 reports the area under the receiver operating characteristic (ROC) curve. This is a standard metric that summarizes both false positives and false negatives and is not biased for rare labels. This analysis showed that even for relatively rare concepts, successful decoding was possible.
Our approach links different cognitive terms to functionally-specialized brain regions:
The inference framework introduced here represents a new approach to developing functional atlases of the human brain. It formally characterizes representations for various cognitive processes that evoke overlapping brain responses, and makes it possible to pool many task-fMRI experiments probing different cognitive domains. Existing meta-analysis approaches face the risk of being unspecific, as demonstrated by our standard analysis results on our database (Figs 3 and 4). Databases of coordinates, such as NeuroSynth, can more easily accumulate data on many different cognitive concepts and support formal reverse inference. This data accumulation is promising, but existing reverse-inference approaches do not suffice to fully remove the overlap in functional regions (Fig 6). Our approach gives more differentiated maps for cognitive concepts by analyzing them in a way that leverages the cognitive ontology. They are also sharper, presumably because they are derived from images rather than coordinates. In a multi-modal framework [17], these maps could be combined with resting-state and anatomical data to provide cognitive resolution to brain parcellations. Note that our framework is meant to be used at the population level and does not address individual brain mapping or decoding.
Our analysis framework overcomes the loss in specificity typical of data aggregation. As a result, it enables analyzing jointly more cognitive processes. These richer models can map qualitatively different information. Analyzing more diverse databases of brain functional images can bring together two central brain-mapping questions: where is a given cognitive process implemented, and what cognitive processes are represented by a given brain structure. Answers to the “what” question have traditionally been provided by invasive studies or neurological lesion reports. Indeed, in a given fMRI study, brain activity results from the task. Concluding on what processes are implied by the observed activity risks merely capturing this task. Decoding across studies can answer this question, by demonstrating the ability to perform accurate inference from brain activity to cognitive function [36].
Reverse-inference maps are essential to functional brain mapping. A key insight comes from the analysis in NeuroSynth [9]: some brain structures are activated in many tasks. Hence, a standard analysis –forward inference– showing such a structure as activated does not provide much information about what function is being engaged. Reverse inference puts the observed brain activity in a wider context by characterizing the behavior that it implies. The analysis performed in NeuroSynth accounts for the multiple tasks that activate a given structure, performing a Bayesian inversion with the so-called Naive Bayes model; however, it does not account for other activation foci in the brain that characterize the function. Put differently, our approach departs from the model used by NeuroSynth for reverse inference by what it conditions upon: NeuroSynth’s model asserts functional specialization conditional to other terms, while we condition on other brain locations when predicting concept occurrence. This difference should be kept in mind when interpreting differences between the two types of approaches. The Inferior Temporal Gyrus (ITG), for instance, is more active in object-recognition tasks than in other paradigms. However, observing activity in the ITG does not help deciding whether the subject is recognizing faces or other types of objects: the information is in the Fusiform gyrus. An important difference between reverse-inference maps with a Naive Bayes –as in Neurosynth– and using a linear model –as in our approach– is that the Naive Bayes maps do no capture dependencies across voxels. On the opposite, linear models map how brain activity in a voxel relates to behavior conditionally on other voxels. Technically, this is the reason why Neurosynth reverse-inference maps related to object recognition overlap in the IT cortex (Fig 3e) while maps produced by our approach separate the representations of the various terms in the ventral mosaic (Fig 3d).
Another, more subtle, benefit of the two-layer model over more classical multi-label approaches is that it combines the decisions of classifiers based on subsets of the data, such as the OvO classifiers, which helps learning relevant local discriminative information.
In sum, our mapping approach provides a different type of brain maps: They quantify how much observing activity in a given brain location, as opposed to other brain locations, informs on whether the subject was engaged in a cognitive operation.
Brain functional atlases are hard to falsify: is a functional atlas specific to the experimental paradigms employed to build it, or is it more generally characteristic of human brain organization? The success of statistically-grounded reverse inference, which generalizes to new paradigms from unseen studies, suggests that there must be some degree of generality in the present atlas. In demonstrating this generalization, the present work goes beyond previous work that had shown generalization to new subjects under known task conditions [12], but not to unknown protocols. However, it is worth noting that here too we found that it was easier to predict on held-out subjects (from one of the training studies) than on held-out studies (see S6 Text Evaluating prediction accuracy: cross-validation), consistent with a substantial effect of the specific task (see S2 Text Similarities of activations across the database). Despite this, our ontology-enabled approach was able to successfully predict cognitive processes for new tasks. Interestingly, it opens the possibility to perform prospective decoding analyses on novel data, hence makes it easier to grasp the added information of incoming data.
To enable this generalization across paradigms, we characterize each task by the multiple cognitive concepts that it recruits, that are specified in the ontology. Departing from the subtractions often used in brain mapping, our framework relies on quantifying full descriptions of the tasks. In the context of decoding, this approach leads to multi-label prediction, predicting multiple terms for an activation map, as opposed to multi-class prediction, used in prior works [12, 16], that assigns each new map to a single class. The use of the multi-label approach combined with an ontology capturing the relationships between terms provides a principled way of modeling the multiple components of cognition and thus avoids the need for hand-crafted oppositions that are customarily used in subtraction studies. Defining good ontologies is yet another challenge for the community, but it is not unlikely that brain imaging will become part of that process [36, 37]. Providing a methodological approach founded on an explicit hierarchy of cognitive concepts would allow to test for different cognitive ontologies, and, provided with a comparison metric, select the best ontology according to the available data. Although the present analysis is limited to a relatively small set of cognitive functions, such an approach will be essential as the field attempts to scale such analyses to the breadth of human cognition.
To build brain functional atlases that map many cognitive processes, we have found that reverse inference and an ontology relating these processes were key ingredients. Indeed, because of the experimental devices used in cognitive neuroimaging, some regions –e.g. attentional or sensory regions– tend to be overly represented in forward inferences. An ontology encodes the related cognitive processes that must be studied together to best establish forward or reverse inferences.
Using a relatively small number of independent task fMRI datasets, our brain-mapping approach reconciles the conundrum of multiple cognitive processes/labels mapping to often overlapping brain regions in activation studies. More data will enable even more fine-grained process-region mappings. In particular higher-level cognitive processes elude the present work, limited by the amount and the diversity of the studies in our database. Indeed, high-level terms form very rare classes in the datasets employed here (see S1 Text Distribution of terms in our database). With increased data sharing in the neuroimaging community [38], there is a growing opportunity to perform this kind of analysis on a much larger scale, ultimately providing a comprehensive atlas of neurocognitive organization. A major challenge to such analyses is the need for detailed task annotation; whereas annotation of task features such as the response effector is relatively straightforward, annotation of complex cognitive processes (e.g., whether a task involves attentional selection or working memory maintenance) is challenging and often contentious. The utility of the ontology in the present work suggests that this effort is worthwhile, and that the increased utilization of ontologies in cognitive neuroscience may be an essential component to solving the problem of how cognitive function is organized in the brain.
|
10.1371/journal.pgen.1007034 | Rescuing the aberrant sex development of H3K9 demethylase Jmjd1a-deficient mice by modulating H3K9 methylation balance | Histone H3 lysine 9 (H3K9) methylation is a hallmark of heterochromatin. H3K9 demethylation is crucial in mouse sex determination; The H3K9 demethylase Jmjd1a deficiency leads to increased H3K9 methylation at the Sry locus in embryonic gonads, thereby compromising Sry expression and causing male-to-female sex reversal. We hypothesized that the H3K9 methylation level at the Sry locus is finely tuned by the balance in activities between the H3K9 demethylase Jmjd1a and an unidentified H3K9 methyltransferase to ensure correct Sry expression. Here we identified the GLP/G9a H3K9 methyltransferase complex as the enzyme catalyzing H3K9 methylation at the Sry locus. Based on this finding, we tried to rescue the sex-reversal phenotype of Jmjd1a-deficient mice by modulating GLP/G9a complex activity. A heterozygous GLP mutation rescued the sex-reversal phenotype of Jmjd1a-deficient mice by restoring Sry expression. The administration of a chemical inhibitor of GLP/G9a enzyme into Jmjd1a-deficient embryos also successfully rescued sex reversal. Our study not only reveals the molecular mechanism underlying the tuning of Sry expression but also provides proof on the principle of therapeutic strategies based on the pharmacological modulation of epigenetic balance.
| In eukaryotes, DNA wraps an octamer of the core histones. Covalent modifications on the histones have diverse biological functions including transcriptional regulation. Histone H3 lysine 9 (H3K9) methylation is a hallmark of transcriptionally silenced chromatin. In mammals, the sex-determining gene Sry initiates testis differentiation in embryonic gonads. Sry expression in gonads is fine-tuned in both space and time. Here, we demonstrated that fine-tuning of Sry expression is achieved by the balance in activities between H3K9 demethylase and H3K9 methyltransferase. We found that the GLP/G9a complex is the enzyme catalyzing H3K9 methylation of Sry. Based on this finding, we tried to rescue the sex-reversal phenotype of the mutant mice by modulating the H3K9 methylation balance of Sry. We succeeded by modulating the H3K9 methylation balance not only with a genetic approach but also with a chemical approach using an inhibitor of GLP/G9a enzyme. Aberrant histone methylation levels are associated with diseases, including cancer, and intellectual disability. Our study provides proof for the principle of therapeutic strategies based on the pharmacological modulation of histone methylation balance.
| Covalent modifications of histone tails are epigenetic marks that play roles in many nuclear processes. Among them, methylation of histone H3 lysine 9 (H3K9) is a hallmark of transcriptionally silenced heterochromatin. Various types of H3K9 methyltransferases (“writers”) and demethylases (“erasers”) have been discovered in mammals. Considering that these H3K9 methylation “writers” and “erasers” are expressed in a cell-type-specific and developmental-stage-specific manner, H3K9 methylation levels are regulated not statically but dynamically during development [1]. In this situation, a specific combination of H3K9 methylation “writer” and “eraser” may antagonistically tune the expression levels of their target genes.
We previously demonstrated that H3K9 demethylation plays an indispensable role in mouse sex development [2]. XY mice lacking Jmjd1a (also called Kdm3a), an “eraser” for H3K9 methylation, showed male-to-female sex reversal. Jmjd1a demethylates H3K9 of the sex-determining gene Sry in sexually undifferentiated gonads at embryonic day 11.5 (E11.5), thereby activating Sry transcription. Jmjd1a deficiency induced a decrease, but not a delay of Sry expression in the developing gonads. We found a significant increase of dimethylated H3K9 (H3K9me2) at the Sry locus in embryonic gonads at E11.5 [2], suggesting the existence of an H3K9me2 “writer” that catalyzes H3K9 methylation at the Sry locus.
Several lines of evidence suggest that aberrant histone methylation levels are associated with diseases, including cancer, and intellectual disability [1]. Therefore, normalizing histone modification levels by manipulating the activity of the corresponding modifier is proposed as a potentially powerful therapeutic strategy [3]. Therefore, we speculated that manipulation of the activity of the H3K9me2 “writer” responsible for H3K9 methylation at the Sry locus normalizes Sry expression in the mice lacking the H3K9me2 “eraser” Jmjd1a.
In this study, we identified the H3K9 methyltransferase GLP/G9a complex as the enzyme responsible for H3K9 methylation at the Sry locus. Based on this finding, we aimed to rescue the aberrant sex development of Jmjd1a-deficient mice by modulating the activity of the GLP/G9a complex. The GLP heterozygous mutation rescued not only H3K9 hypermethylation at the Sry locus but also the perturbed Sry expression in Jmjd1a-deficient embryos. Strikingly, the sex-reversal phenotype in Jmjd1a-deficient mice was completely rescued by the GLP heterozygous mutation. We also aimed to rescue the phenotype by artificially manipulating the activity of the GLP/G9a complex. The administration of the GLP/G9a complex inhibitor into Jmjd1a-deficient embryos at a specific developmental time point rescued the aberrant sex differentiation of these mice. Our studies provide a novel strategy by which diseases attributed to the dysfunction of an epigenetic “eraser” can be rescued by blocking the activity of the corresponding epigenetic “writer.”
In mammals, a number of enzymes possess intrinsic H3K9 methyltransferase activities, such as Suv39h1 [4], Suv39h2 [5], Eset [6], G9a [7], and GLP [8]. Among them, G9a (also called Ehmt2/Kmt1c) and GLP (also called Ehmt1/Kmt1d) form a stoichiometric heterodimer complex [9–11]. Jmjd1a deficiency resulted in the increase of H3K9me2, but not trimethylated H3K9me3 in the developing gonads at E11.5, suggesting that the enzyme counteracting Jmjd1a-mediated H3K9 demethylation produces H3K9me2 (Fig 1A and 1B, S1 Fig). We previously demonstrated that the global level of H3K9me2 in developing mouse embryos was dominantly catalyzed by the GLP/G9a complex [10]. Taking these findings together, the GLP/G9a complex was the strongest candidate for an enzyme that counteracts Jmjd1a-mediated H3K9 demethylation in the developing gonads.
Somatic cells of E11.5 embryonic gonads contain subpopulations with high and low expression levels of an orphan nuclear receptor, Nr5a1 (also called Sf-1/Ad4BP) [12]. Because a previous study demonstrated that the Nr5a1-high population contains Sry-expressing pre-Sertoli cells [13], we examined mRNA expression levels of GLP/G9a in this population (Fig 1C). We had established a transgenic mouse line Nr5a1-hCD271-tg, in which the human cell surface marker CD271 (also called LNGFR) is expressed depending on the Nr5a1 promoter [2] [14]. We prepared a single cell suspension from the gonads/mesonephroi of E11.5 Nr5a1-hCD271-tg embryos and then fractionated it according to the expression level of hCD271 by fluorescence-activated cell sorting (FACS) (S2 Fig). The hCD271-negative population contained mesonephric cells and germ cells [2]. As expected, quantitative RT-PCR (RT-qPCR) analysis demonstrated that the endogenous Nr5a1 expression levels were high and low in hCD271-high and -low populations, respectively (Fig 1C, left). Concordant with the previous study [13], Sry transcript was substantially enriched in the hCD271-high population (Fig 1C, right). Jmjd1a transcript was also significantly enriched in the hCD271-high population. GLP and G9a transcripts were detected in all populations at similar levels, suggesting the ubiquitous expression of GLP/G9a complex in the developing gonads (Fig 1C, right). To address whether GLP/G9a complex and Jmjd1a were co-expressed in Sry-expressing pre-Sertoli cells, we performed triple immunostaining analyses of E11.5 embryonic gonads with antibodies against GLP/G9a, Jmjd1a, and Sry. As shown in Fig 1D and 1E, GLP/G9a complex was expressed in Sry-expressing pre-Sertoli cells. Furthermore, Sry- and GLP/G9a complex-positive cells contained robust signals for Jmjd1a (Fig 1D and 1E). Cells containing both signals of GLP (or G9a) and Jmjd1a among the Sry-expressing pre-Sertoli cells ware calculated. As summarized in Fig 1D and 1E (right panels), almost all Sry-positive cells contained the signals of both GLP/G9a complex and Jmjd1a. We therefore concluded that GLP/G9a H3K9 complex and Jmjd1a are actually co-expressed in Sry-expressing pre-Sertoli cells.
We previously demonstrated that GLP is a limiting factor that controls the stability of the GLP/G9a heterodimer complex. In addition, the heterodimer formation of GLP/G9a was shown to be essential for H3K9 methylation in vivo [10] [15]. Thus, we first examined whether a GLP mutation might rescue the increased H3K9me2 levels in Jmjd1a-deficient mice. A homozygous GLP mutation in mice leads to embryonic lethality around E9.5 [10]. We therefore generated mice heterozygous for the GLP mutation (GLPΔ; lacking the coding sequences for the catalytic SET domain) combined with a Jmjd1a-null (Jmjd1aΔ/Δ) background. Embryonic gonads/mesonephroi at E11.5 were stained with antibodies against H3K9me2 and Nr5a1. The H3K9 methylation levels of Nr5a1-positive gonadal somatic cells were compared by FACS analysis (Fig 2A). Jmjd1a deficiency resulted in an increased H3K9me2 level in Nr5a1-positive gonadal somatic cells, indicating the substantial contribution of Jmjd1a to H3K9 demethylation (Fig 2B and 2C). Notably, introduction of the GLP mutation into the Jmjd1aΔ/Δ background significantly reduced the H3K9me2 level in gonadal somatic cells (Fig 2B and 2C). Thus, we concluded that the GLP/G9a complex counteracts Jmjd1a-mediated H3K9 demethylation in the developing gonads in the sex-determining period.
To examine the possible counteracting role of the GLP/G9a complex on Jmjd1a-mediated H3K9 demethylation at single gene level, especially at the Sry locus, chromatin immunoprecipitation (ChIP) analyses were performed. We previously demonstrated that Jmjd1a is enriched at the linear promoter region of Sry in E11.5 gonadal somatic cells [2] (Fig 3A). We purified gonadal somatic cells from E11.5 XY Nr5a1-hCD271-tg embryos and then subjected these cells to ChIP-qPCR analyses. As shown in Fig 3B, we found that G9a is also accumulated in the linear promoter region of Sry. We used Npas4 as a positive control locus, that had been identified as one of the target loci of G9a [16]. We therefore concluded that H3K9 methyltransferase GLP/G9a complex and H3K9 demethylase Jmjd1a both target the Sry locus in embryonic gonadal somatic cells in the sex-determining period.
We next aimed to elucidate the impact of the GLP mutation on the H3K9me2 level of the Sry locus. Gonadal somatic cells were immunomagnetically purified from XY Jmjd1aΔ/Δ, GLPΔ/+, and Nr5a1-CD271-tg embryos for ChIP analysis (the experimental scheme is shown in S3 Fig). Importantly, the numbers of purified cells were similar among XY Jmjd1aΔ/+-, XY Jmjd1aΔ/Δ-, and XY Jmjd1aΔ/Δ;GLPΔ/+ gonads, indicating that these mutations did not affect gonadal somatic cell numbers (S3 Fig). Purified gonadal somatic cells were then subjected to native ChIP-qPCR analyses. Consistent with our previous study, Jmjd1a deficiency resulted in an increase of H3K9me2 at the Sry locus in E11.5 gonadal somatic cells, compared with the level in control cells (Fig 3C). Importantly, the increased level of H3K9me2 at the Sry locus was significantly rescued by the GLP mutation (Fig 3C). We also demonstrated the inverse relationship between H3K9me2 and H3K4me2 at the Sry locus. The latter is an epigenetic mark for transcriptionally activated chromatin (Fig 3C). Taking these findings together, the GLP/G9a complex is the bona fide enzyme responsible for H3K9 methylation at the Sry locus in E11.5 gonadal somatic cells.
We next elucidated whether the Jmjd1a mutation and/or Jmjd1a/GLP compound mutations may also induce transcriptional perturbation of Y chromosome genes other than Sry. Gonadal somatic cells were immunomagnetically purified from E11.5 embryos and then subjected to mRNA expression analysis. As shown in Fig 4A, we could not detect significant differences in the mRNA expression levels of Sry-neighboring genes, Uty, Ddx3y, Usp9y, and Zfy2, between control and mutant gonads. Accordingly, our previous microarray analysis showed that the expression levels of Y chromosome genes other than Sry were not affected by Jmjd1a deficiency [2]. Next, we evaluated the H3K9me2 levels of Uty, Ddx3y, Usp9y, and Zfy2 by ChIP-qPCR analysis using purified E11.5 gonadal somatic cells (Fig 4B). Again, we could not detect significant differences in the H3K9me2 levels of these genes between control and mutant gonadal somatic cells. Taking these results together, Jmjd1a- and GLP/G9a complex-mediated expression tuning is highly confined to the Sry locus and is not extended to other genes on Y chromosome.
It is known that H3K4me3 and H3K9ac are enriched at the linear promoter region of Sry in E11.5 gonadal somatic cells [13]. To address whether Jmjd1a and/or Jmjd1a/GLP compound mutations may influence these modifications, we performed ChIP-qPCR analysis using purified E11.5 gonadal somatic cells. As shown in S4 Fig, neither Jmjd1a nor Jmjd1a/GLP compound mutations induced significant alterations of H3K4me3 and H3K9ac at the Sry locus. It is also known that CpG sequences of the linear promoter region of Sry are hypomethylated in embryonic gonads at the time of Sry expression [17]. To address whether Jmjd1a deficiency may influence DNA methylation of Sry promoter, we fractionated E11.5 gonadal somatic cells carrying the Nr5a1-hCD271 transgene into hCD271-high and -low populations by FACS and introduced them into bisulfite sequence analysis. In control gonads, the DNA methylation level of Sry promoter was significantly low in the hCD271-high population compared to that of the hCD271-low population in E11.5 embryonic gonads (S4 Fig). These results are in accordance with a previous study [13]. We next compared DNA methylation levels of the Sry promoter of the hCD271-high population between Jmjd1aΔ/+ and Jmjd1aΔ/Δ littermates. However, we could not find significant difference levels (S4 Fig). We therefore concluded that Jmjd1a mutation did not induce a significant alteration of DNA methylation at the Sry locus.
Sry activation is the first event in mammalian sex differentiation. In mice, sufficient and temporally accurate expression of Sry in sexually undifferentiated gonads at E11.5 is critical for triggering the testis-determining pathway [18]. To address whether the GLP mutation also rescues the perturbed expression of Sry in Jmjd1a-deficient gonads, we examined Sry expression by co-immunofluorescence analysis for Sry and Gata4, a marker of gonadal somatic cells. As shown in Fig 5A and 5B, the number of Sry-positive cells was reduced to approximately 25% in XY Jmjd1aΔ/Δ gonads at E11.5. The number of Sry-positive cells was significantly, although not completely, rescued in XY Jmjd1aΔ/Δ;GLPΔ/+ littermates, indicating that the GLP/G9a complex and Jmjd1a antagonistically tune Sry expression in E11.5 gonadal somatic cells. We also performed expression analysis of Sox9, a downstream target of Sry, in E11.5 gonads (Fig 5C and 5D). The number of Sox9-positive cells was also increased by the GLP mutation. Interestingly, the GLP mutation had a more profound effect on the increase of Sox9-positive cells compared with that of Sry-positive cells, presumably reflecting the activation of a non-cell-autonomous pathway of Sox9 expression [19].
To examine embryonic gonadal cell differentiation just after sex determination, we investigated the expression of the testicular Sertoli cell marker Sox9 and the ovarian somatic cell marker Foxl2 in XY Jmjd1aΔ/Δ- and XY Jmjd1aΔ/Δ;GLPΔ/+ embryonic gonads at E13.5 (Fig 6A). Control XY gonads had Sox9-positive cells and did not contain Foxl2-positive cells. Furthermore, multiple tubule-like structures were found in control XY gonads. On the other hand, Jmjd1a-deficient XY gonads were ovotestes containing not only Sox9- but also Foxl2-positive cells and had no tubule-like structures. As shown in Fig 6A and 6B, the GLP mutation restored the number of Sox9-positive cells and testicular tubule formation, both of which were perturbed by Jmjd1a deficiency, indicating that the GLP mutation successfully rescued gonadal sex differentiation of Jmjd1a-deficient embryos.
GLP and G9a form a heterodimer, which is essential for H3K9 methylation in vivo. We thus also performed epistatic analyses between G9a and Jmjd1a in mouse sex development. Notably, a G9a heterozygous mutation did not rescue the sex-reversal phenotype of XY Jmjd1a-deficient embryos (S5 Fig). Our previous studies indicated that GLP, but not G9a, is a limiting factor controlling the amount of GLP/G9a holoenzyme complex [10]. Consistent with this, we found that the GLP heterozygous mutation induced a significant reduction of G9a protein in embryonic gonads (S6 Fig). Thus, the decreased level of H3K9me2 associated with the GLP heterozygous mutation may be attributable to the reduction in the amount of the GLP/G9a complex.
We had previously established GLP-tg mice [20]. In this line, cDNA for Flag-tagged GLP was inserted in the Rosa26 locus and was designed to be expressed ubiquitously depending on an artificial CAG promoter [20]. To learn whether the overexpression of GLP affects sex determination, we compared the expression levels of Nr5a1, Sry, and Sox9 in gonads of GLP-tg XY embryos at E11.5. As shown in S7 Fig, although the amount of GLP transcript was actually elevated in XY GLP-tg gonads of E11.5 embryos, those of Nr5a1, Sry, and Sox9 transcripts were indistinguishable between XY control and XY GLP-tg gonads (S7 Fig). A possible explanation is that unidentified limiting factor(s) other than GLP may be required for the GLP/G9a complex to exert its function in the developing gonads.
We finally verified the impact of the GLP mutation on the sex-reversal phenotype of Jmjd1a-deficient mice in adults. Consistent with our previous study [2], XY mice lacking Jmjd1a alone were frequently sex-reversed; for example, analysis of the external genitalia revealed that of 15 XY Jmjd1aΔ/Δ animals, one carried male-type genitalia, two carried intersex-type genitalia with a micropenis and well-developed mammary glands, and another carried typical female-type genitalia (Fig 7A). In addition, analysis of the internal genitalia demonstrated that one exhibited bilateral testes, two exhibited a testis and an ovary, and another had two ovaries. In contrast, all XY Jmjd1aΔ/Δ;GLPΔ/+ littermates exhibited male-type external genitalia with bilateral testes (Fig 7B). Taking these findings together, we conclude that the GLP mutation completely rescued the adult sex-reversal phenotype of Jmjd1a-deficient XY mice.
Our genetic experiments revealed that modulation of H3K9 methyltransferase activity of the GLP/G9a complex might be therapeutically effective for rescuing the aberrant sex development of Jmjd1a-deficient mice. We next aimed to rescue the phenotype in a different way using UNC0642, a chemical inhibitor of the GLP/G9a complex, which was recently developed by Liu et al. [21]. The experimental scheme of UNC0642 administration to Jmjd1a-deficient embryos is shown in Fig 8A. Briefly, 0.5 mg of UNC0642 was intraperitoneally injected once into pregnant females carrying E10.5 Jmjd1a-deficient embryos, and then the gonadal differentiation of the embryos was examined. As shown in Fig 8B and 8C, UNC0642 administration to Jmjd1a-deficient embryos resulted in a significant increase in the number of Sox9-positive male somatic cells at E13.5, while solvent only did not (compare with Fig 6B). This indicates that UNC0642 partially, but significantly, rescued the gonadal sex differentiation of Jmjd1a-deficient embryos after sex determination. We next investigated the impact of the embryonic administration of UNC0642 on the sex development of Jmjd1a-deficient mice by examining the external and internal genitalia of adult mice (Fig 8D). Although partially or completely sex-reversed mice were still found in the UNC0642-administered Jmjd1a-deficient mice, five out of 12 UNC0642-administered animals carried bilateral testes. In contrast, only one out of 11 animals in the solvent-injected control group exhibited bilateral testes (Fig 8D and 8E). Taking these findings together, we conclude that administration of UNC0642 into E10.5 embryos successfully rescued the subsequent sex development of Jmjd1a-deficient mice.
Here, we identified GLP/G9a H3K9 methyltransferase complex as an enzyme counteracting Jmjd1a-mediated H3K9 demethylation at the Sry locus in gonadal somatic cells. To our knowledge, this is the first study to identify the set of histone methyltransferase and demethylase that in combination account for stage- and cell-type-specific gene regulation in mammalian development.
Our data show that the molecular balance of the GLP/G9a complex and Jmjd1a is a critical factor for the tuning of Sry expression (Fig 9). We previously showed that G9a and GLP are expressed in almost all adult tissues in mice [9, 10]. On the other hand, previous studies demonstrated that Jmjd1a is expressed in a tissue- and developmental stage-specific manner [22–24]. Considering that the expression of Sry is suppressed in almost all adult tissues in mice, this Sry silencing might be explained, at least in part, by the robust H3K9 methylation of the GLP/G9a complex and the absence of H3K9 demethylase in these tissues. We previously demonstrated the temporally specific expression of Jmjd1a in embryonic gonadal somatic cells, which reaches a plateau around E11.5 [2]. In addition, we demonstrated the cell-type-specific expression of Jmjd1a in this study, as Jmjd1a transcript was substantially enriched in the Nr5a1-high population of E11.5 embryonic gonads (Fig 1C). High levels of Jmjd1a expression may overcome GLP/G9a complex-mediated H3K9 methylation, thereby inducing Sry expression in the pre-Sertoli cells.
Although the GLP mutation significantly rescued the perturbed Sry expression in Jmjd1a-deficient embryonic gonads, the Sry-positive cell population in Jmjd1aΔ/Δ;GLPΔ/+ gonads remained approximately half of that detected in control gonads (Fig 5B). Therefore, it is a surprising finding that the GLP mutation completely rescued the sex reversal of Jmjd1a-deficient XY mice in adults (Fig 7). The GLP mutation reduces global H3K9me2 levels of Jmjd1a-deficient gonadal somatic cells (Fig 2). Thus, it is also possible that GLP/G9a complex has a role in the sex differentiation pathway, independent of Sry regulation. In this regard, the GLP mutation may inhibit the ovarian development pathway in Jmjd1a-deficient gonads. Although we could not rule out this possibility, it seems likely that the restored Sry expression is a primary cause for the rescue of the sex reversal of Jmjd1a-deficient XY mice. Because there is a certain threshold level for Sry expression to induce the male pathway [18], it is conceivable that Sry expression substantially exceeds the threshold level as a result of the GLP mutation in the Jmjd1a-deficient background, thereby conferring profound effects on the subsequent male pathway.
Eset is known as an enzyme responsible for tri-methylation of H3K9 [6]. A previous study demonstrated that Eset is expressed in embryonic gonads [25]. Accordingly, we also confirmed the expression of Eset in E11.5 XY gonadal somatic cells by RT-qPCR analysis [2]. We introduced the Eset heterozygous mutation into the Jmjd1a-mutant background, and then the sex development of XY Jmjd1aΔ/Δ; EsetΔ/+ embryos was examined (S8 Fig). Consequently, we could not find any profound effect of the Eset mutation on the perturbed sex development of Jmjd1a-deficient mice (S8 Fig). Our previous study indicated that the H3K9me3 level of Sry was unchanged by Jmjd1a deficiency in E11.5 embryonic gonads. Altogether, it seems likely that Jmjd1a-mediated H3K9 demethylation does not counteract Eset-mediated H3K9 tri-methylation, at least in the Sry locus.
Previous studies revealed the enrichment of H3K4me3/H3K9ac and that low levels of CpG methylation are characteristic in the Sry promoter region of gonadal somatic cells during the sex-determining period [17] [13]. In this study, we demonstrated that Jmjd1a deficiency and/or the increased level of H3K9me2 did not affect H3K4me3/H3K9ac and CpG methylation levels of the Sry locus (S4 Fig). On the other hand, it seems likely that H3K9me2 and H3K4me2 marks are exclusively deposited mutually and these marks exert antagonistic functions for transcription, at least in the Sry locus. Jmjd1a deficiency and/or the increased levels of H3K9me2 resulted in the decrease of H3K4me2 of this locus, concomitantly with the suppression of Sry [2], also shown in Fig 3. The fact that Jmjd1a deficiency result in the loss of H3K4me2, but not H3K4me3, in the Sry locus warrants further discussion. It is one possible explanation that Jmjd1a or Jmjd1a-mediated H3K9 hypomethylation may prevent the accession of specific enzyme(s) responsible for H3K4me2 demethylation.
We have demonstrated that just a single administration of GLP/G9a inhibitor to E10.5 embryos significantly rescues the aberrant sex development of Jmjd1a-deficient mice. Mutation, silencing, or downregulation of histone methylation “erasers” was found in several types of cancer [26]. Our experiments suggest a new therapeutic strategy, in which diseases arising from the dysfunction of an epigenetic “eraser” can be rescued by blocking the activity of the counteracting epigenetic “writer.”
Animal experiments were performed under the animal ethical guidelines of Tokushima University and Kyoto University. The Ethics Committee of Tokushima University for Animal Research (Approval number: 14108) and the Animal Experimentation Committee of Kyoto University (Approval number: A12-6-2) approved this study.
Mouse lines of GLPΔ/+, G9aΔ/+, Jmjd1aΔ/+, and Nr5a1-hCD271-tg were sequentially backcrossed to C57BL/6J, and then the F5 or later generation was used. Since the sex reversal frequencies of Jmjd1a-deficient mice were dependent on the origin of the Y chromosome [2], we used mice carrying a Y chromosome of CBA origin in this study. However, we only used mice carrying a Y chromosome of B6 origin in the experiments shown in S5 Fig.
Guinea-pig polyclonal antibodies against Sry were generated by the immunization of bacterially expressed 6xHis-tagged Sry (residues 82–395, NP_035694). Additional antibodies used in this study were as follows: goat anti-Gata4 (Santa Cruz, C-20), rabbit anti-Sox9 (Millipore, AB5535), goat anti-Foxl2 (Abcam, ab-5096), mouse anti-LNGFR (Miltenyi Biotec), rabbit anti-Jmjd1a [2], mouse anti-G9a (Perseus Proteomics, 8620A), mouse anti-GLP (Perseus Proteomics, B0422), rabbit anti-G9a (CST, #3306), mouse anti-H3K9me2 [27], mouse anti-H3K4me2 [27], mouse anti-H3K4me3 [27], mouse anti-H3K9ac [27], and rabbit anti-Nr5a1 (a gift from Dr. K. Morohashi).
Tissues were fixed in either Bouin’s solution or 4% paraformaldehyde, embedded in paraffin, and cut into 4-μm sections. For histological analysis, sections were stained with hematoxylin/eosin or hematoxylin/PAS. For immunohistochemistry, sections were deparaffinized and rehydrated, and then autocleaved at 105°C for 5 min in 10 mM citric acid buffer (pH 6.0). To quench endogenous peroxidase, the sections were treated with 0.3% (v/v) hydrogen peroxide. After blocking with TBS containing 2% skim milk and 0.1% Triton-X100 at room temperature for 1 h, sections were incubated with primary antibodies overnight at 4°C. For fluorescence staining, the sections were incubated with Alexa-conjugated secondary antibodies (Life Technologies) at room temperature for 1 h and counterstained with DAPI. The sections were mounted in Vectashield (Vector) and observed with a confocal laser scanning microscope (LSM700, Carl Zeiss).
Isolated gonads and mesonephroi from E11.5 embryos were digested with Accutase (Nacalai) to obtain a single cell suspension. For flow cytometric analysis, cells were fixed with 2% paraformaldehyde (PFA) in PBS for 10 min, permeabilized with ice-cold ethanol for 20 min, and blocked with 0.5% skim milk in PBS for 1 h. They were then stained with primary antibodies overnight at 4°C and subsequently incubated with Alexa-conjugated secondary antibodies (Life Technologies) for 1 h at room temperature. Data were collected using FACSCanto 2 (BD Bioscience) and analyzed with FlowJo software (TreeStar). For FACS sorting, cells were stained with FITC-labeled anti-LNGFR and sorted based on fluorescence intensity using FACS Aria 2 (BD Bioscience) as shown in S2 Fig.
The experimental scheme for ChIP analysis is shown in S3 Fig. Briefly, two-cell embryos were prepared by in vitro fertilization using sperm derived from Jmjd1aΔ/+;GLPΔ/+;Nr5a1-hCD271-tg males and eggs derived from Jmjd1aΔ/+ females, and then transferred to pseudopregnant recipients. Gonadal somatic cells were purified from embryos that had developed to tail somite stage 17–19, as described previously [2, 14]. For native ChIP analysis of histone modifications, purified cells were pooled (n = 2–4 per genotype) and subjected to ChIP analysis following a protocol described previously [15], with minor modifications. Briefly, cells were suspended in 5 μl of 0.3 M sucrose-containing buffer 1 (60 mM KCl, 15 mM NaCl, 5 mM MgCl2, 0.1 mM EGTA, 0.5 mM dithiothreitol, 0.1 mM PMSF, 3.6 ng/ml aprotinin, 15 mM Tris–HCl, pH 7.5) and lysed by the addition of 5 μl of 0.3 M sucrose-containing buffer 1 with 0.8% NP40 on ice for 10 min. After the addition of 80 μl of 1.2 M sucrose-containing buffer 1, the chromatin fraction was collected as pellets by centrifugation. These pellets were digested with micrococcal nuclease (MNase) (0.02–0.05 U, Takara) in 10 μl of MNase digestion buffer (0.32 M sucrose, 4 mM MgCl2, 1 mM CaCl2, 0.1 mM PMSF, 50 mM Tris–HCl, pH 7.5), using Thermo Mixer (Eppendorf) at 37°C and 1000 rpm for 15 min, and then digestion was stopped with EDTA. Supernatant was obtained by centrifugation and incubated with anti-H3K9me2-, anti-H3K9ac-, anti-H3K4me2-, or anti-H3K4me3-coated magnetic beads (Dynabeads Protein G, Invitrogen) in 50 μl of incubation buffer (50 mM NaCl, 5 mM EDTA, 0.1% NP40, 0.1 mM PMSF, 20 mM Tris–HCl, pH 7.5), at 4°C for 2 h. DNA was extracted from the immune complexes according to the standard protocol and then analyzed by real-time PCR using primers specific for Y chromosome genes (Sry, Uty, Ddx3y, Usp9y, and Zfy2). For cross-link ChIP analysis of G9a, purified gonadal somatic cells from 20 embryos (approximately 8 × 105 cells) were pooled and combined with 5 × 106 cells of female mouse embryonic fibroblasts, cross-linked with 25 mM DSG (Thermo Fisher Scientific) and 1% formaldehyde, and applied to ChIP analysis with rabbit anti-G9a antibody following a protocol described previously [2].
Genomic DNA was isolated using the All DNA/RNA Micro kit (QIAGEN). Genomic DNA was treated with sodium bisulfite using the MethylEasy Xceed Rapid DNA Bisulfite Modification Kit (Human Genetic Signatures) following the manufacturer’s instructions. The bisulfite-treated DNA was PCR-amplified using the primer pair 5′-TTTATATTGGGTTATAGAGTTAGAATAGAT-3′ and 5′-CCAAAATATACTTATAACAAAAATTTTAAT-3′. PCR products were subcloned into the pGEM-T Easy vector (Promega) and sequenced.
The primer sets used in ChIP-qPCR analysis were as follows: Sry linear prom.-f (5′-TGGTCAGTGGCTTTTAGCTCT-3′) and Sry linear prom.-r (5′-AGATGTGATGCAAAGAGAAACA-3′) for Sry, Npas4 ChIP F (5′-CTATGGCCATTTCAGCACCG-3′) and Npas4 ChIP R (5′-AGCTGTTCGACGTCCTGAAG-3′) for Naps4, Gapdh ChIP F (5′-TTGCTTAGGCCTTCCTTCTTC-3′) and Gapdh ChIP R (5′-CATCACCTGGCCTACAGGATA-3′) for Gapdh, ChIP-Uty-F (5′-CCTTTGTGAGGGACTGTTCA-3′) and ChIP-Uty-R (5′-CCACTCAACCACATCAAACC-3′) for Uty, ChIP-Ddx3y-F (5′-ACAATTCCACAACCCAAGGT-3′) and ChIP-Ddx3y-R (5′-AGGTTTCAGCCCACTCATTT-3′) for Ddx3y, ChIP-Usp9y-F (5′-AAGGGACACACAGTTCTCCA-3′) and ChIP-Usp9y-R (5′- CTTGTGAGAAGGGACTGAGG-3′) for Usp9y, ChIP-Zfy2-F (5′- AGGCAGTCTTAGATGCGAAA-3′) and ChIP-Zfy2-R (5′- TCCTGACTCACAACAACAGC-3′) for Zfy2. The primer sets used in RT-qPCR analysis were as follows: Gapdh RT-PCR F (5′-ATGAATACGGCTACAGCAACAGG-3′) and Gapdh RT-PCR R (5′-CTCTTGCTCAGTGTCCTTGCTG-3′) for Gapdh, Ad4BP-e2-F (5′-TTGTCGACTGGGCACGAAGGTGCAT-3′) and Ad4BP-e2-R (5′-GCAGCTCGCTCCAACAGTTCTGCAG-3′) for Nr5a1, Sry-5-SD (5′-TACCTACTTACTAACAGCTGACATCAC-3′) and Sry-3-SD (5′-TGTCATGAGACTGCCAACCACAGGG-3′) for Sry, TSGA-EX 21F (5′-ACTCCAGAGGATCGGAAATATGGGACC-3′) and TSGA-EX 21R (5′-GGGAATTCCCACATAAACCATGACATTGGC-3′) for Jmjd1a, GLP-RT-1B (5′-AACCCAACCTTGTGCCTGTGCGAG-3′) and GLP-RT-2 (5′-CGAGCTGCTCCCCAGCCTGAATCAG-3′) for GLP, G9a-RT-1B (5′-ACCCCAACATCATCCCTGTCCGGG-3′) and G9a-RT-2 (5′-GTCCCAGAATCGGTCACCGTAGTC-3′) for G9a, Sox9-RT-F (5′-AGGAAGCTGGCAGACCAGTA-3′) and Sox9-RT-R (5′-CGTTCTTCACCGACTTCCTC-3′) for Sox9, Uty-RT-F (5′-AAGGCGCTTTGTGGATTAGA-3′) and Uty-RT-R (5′-CTGATTCCACTTTTCCTTCAGC-3′) for Uty, Ddx3y-RT-F (5′- TTGGTCTTGACCTGAAATCATCA-3′) and Ddx3y-RT-R (5′- GCTTCCCTCTGGAATCACGA-3′) for Ddx3y, Usp9y-RT-F (5′- CTTGGTCCCAAATTGCAAGC-3′) and Usp9y-RT-R (5′- TCGGATGGCTTCTTGTCTTG-3′) for Usp9y, Zfy2-Rt-F (5′- GCTTAAGACCTCCAGCAAAAG-3′) and Zfy2-Rt-R (5′- CCGGTCTCTGGCTTTAATGT-3′) for Zfy2. The primer sets used for genotyping were as follows: GLP-6570F (5′-CTGTCCAGTTCCCGATTTTCAAGACTGC-3′) and GLP-5936R (5′-GTCCCACTGGCCACACTGGCAATTC-3′) for detection of the GLPΔ allele; TSGA-G1475R (5′-GAACTGCACCATTAGCTGTCACTTCC-3′), TSGA-1980F (5′-CATGCAGTGAAAGATGCAGTTGCTA-3′), and TSGA-6410F-NheSac (5′-CTAAATATCAAGGCTAGCGAGCTCG-3′) for detection of the Jmjd1aΔ allele; Sf1-1741F (5′-CACAGACCAGGGCAATCCCAAGCCA-3′) and pMACS-LI 2264R (5′-GTCGGAGAACGTCACGCTGTCCAG-3′) for Nr5a1-hCD271-tg; Rbmy1a1-F (5′-AATATGCCAAGAGGAGAGCCGGCGTCTTCC-3′) and Rbmy1a1-R2 (intron) (5′-CCAAGTTGTTGTGGCATTTGGACATC-3′) for detection of the Y chromosome; and GE28R (5′-GCTCCAGGGCGATGGCCTCCGCTGAATGC-3′), GI27-2F (5′-CGGGACAGGGTTTCTCTGTGTAGTCC-3′), and GI-25F (5′-CTGCACGCTGCCTAGATGGAGCATG-3′) for detection of the G9aΔ allele.
Pregnant females at E10.5 were administered 0.5 mg of UNC0642 (Tocris) dissolved in 30 μl of DMSO and mixed with 17.5 μl of ethanol, 52.5 μl of castor oil, and 100 μl of PBS.
Eset-mutant mice were produced by electroporating Cas9 mRNA and gRNA into mouse zygotes according to a protocol published recently [28]. Briefly, 400 ng/μl Cas9 mRNA and 100 ng/μl of each gRNA targeting the genomic sequences of Eset (shown in S8 Fig) were introduced into zygotes (C57BL/6J × C57BL/6J) by electroporation using Genome Editor GEB15 (BEX, Tokyo, Japan). The electroporation conditions were four pulses of 30 V (3 ms ON + 97 ms OFF). The surviving two-cell-stage embryos were transferred to the oviducts of pseudopregnant females. Genotyping of the generated mice was performed using the primer pair 5′-CCCTGGCTGTCCTAGAACTCAC-3′ and 5′-AGGGTTCATTCAGGCTACAAAG-3′.
One-way analysis of variance (one-way ANOVA) and Tukey’s honestly significant difference test were used for statistical analysis.
|
10.1371/journal.ppat.1002818 | Directly Infected Resting CD4+T Cells Can Produce HIV Gag without Spreading Infection in a Model of HIV Latency | Despite the effectiveness of highly active antiretroviral therapy (HAART) in treating individuals infected with HIV, HAART is not a cure. A latent reservoir, composed mainly of resting CD4+T cells, drives viral rebound once therapy is stopped. Understanding the formation and maintenance of latently infected cells could provide clues to eradicating this reservoir. However, there have been discrepancies regarding the susceptibility of resting cells to HIV infection in vitro and in vivo. As we have previously shown that resting CD4+T cells are susceptible to HIV integration, we asked whether these cells were capable of producing viral proteins and if so, why resting cells were incapable of supporting productive infection. To answer this question, we spinoculated resting CD4+T cells with or without prior stimulation, and measured integration, transcription, and translation of viral proteins. We found that resting cells were capable of producing HIV Gag without supporting spreading infection. This block corresponded with low HIV envelope levels both at the level of protein and RNA and was not an artifact of spinoculation. The defect was reversed upon stimulation with IL-7 or CD3/28 beads. Thus, a population of latent cells can produce viral proteins without resulting in spreading infection. These results have implications for therapies targeting the latent reservoir and suggest that some latent cells could be cleared by a robust immune response.
| While HIV is a treatable disease due to effective antiviral therapies, these drugs do not cure HIV. When therapy is stopped, a pool of infected, long-lived, treatment resistant cells re-establishes infection. These latently infected cells, mainly resting CD4+T cells, are barriers to a cure. Studying and understanding the properties of these cells is therefore important to eradicating HIV. It is believed that these latent cells do not produce viral proteins and thus are invisible to the immune system. Here, we show using an in vitro HIV model that a population of latently infected cells can produce HIV Gag. Interestingly, this protein production does not result in the release of detectable infectious virus and so the latent cells are unaffected by antiviral therapy. We therefore examined why some latent cells can produce viral proteins without viral spread. We found that resting cells have the ability to make some of the components required for spreading infection but not all are in sufficient quantity. These results have important implications for treating the latent reservoir, as our work suggests that latent cells might be recognized by a boosted immune response.
| Highly active antiretroviral therapy (HAART) has been successful in suppressing HIV-1 replication and maintaining CD4+T cell counts in patients. However, long-lived, treatment resistant reservoirs are still a barrier to curing HIV. These latently infected cells are predominantly resting CD4+T cells capable ofreleasing infectious virions after stimulation [1], [2]. A major hurdle in studying HIV latency in vivo is the very low frequency ofthese cells. Thus, developing in vitro models with a high frequency of latently infected cells is critical to study the establishment, maintenance, and properties of the latent reservoir. Such models in turn can be used to develop therapies to eliminate these cells.
Several in vitro latent models using primary cells have been described [3]–[13]. Most of these models rely on activation steps for not only expanding CD4+T cells but also for infection, as several reports have shown blocks to HIV infection in resting CD4+T cells [10], [14]–[19]. While these models can generate sufficient numbers of cells for drug screening [8], [9], they have distinct disadvantages. Activation steps are typically vigorous and result in several changes in cell phenotype [20], [21], some of which narrow the types of CD4+T cell subsets that can be studied in vitro [4], [8], [9]. Other models, with less vigorous stimulation steps, avoid these issues but result in low levels of infection [10], [22]. We previously demonstrated that HIV directly integrates into resting CD4+ T cells without requiring any stimulation using a technique called spinoculation [6], [23], [24]. Here, we take advantage of the high level of infection that we can obtain with this method to ask if viral proteins can be expressed in latently infected cells.
Studies in vivo have indicated a population of resting cells can transcribe and translate HIV and SIV proteins [25], [26]. These cells were phenotypically resting but were believed to be productively infected and not truly quiescent due to prior activation or their cytokine milieu [25]. Here, we investigate whether latently infected resting CD4+T cells can transcribe and translate viral proteins without stimulation while in a latent state.
Before we could examine protein expression in latently infected cells, we needed to address concerns regarding the effects of spinoculation on the susceptibility of resting cells to HIV infection. A recent report suggests that spinoculation might induce signaling cascades that artificially allow integration to occur in otherwise resistant resting cells [27]. This evidence is consistent with reports that cytokine stimulation is required for integration to occur in quiescent cells [11], [15]. Therefore, we wanted to determine if spinoculation and/or cytokine treatment enhance integration efficiency in resting CD4+T cells.
To first confirm that integration could occur in resting cells, we spinoculated purified resting CD4+T cells (HLA-DR−, CD25−, CD69−, greater than 98% pure) with HIV (Figure 1A). We have previously shown that this cell population is in the G0/G1a stage of the cell cycle [6], [28]. We found that integration occurred in these cells albeit with slower kinetics than activated cells (Figure 1B), consistent with our prior results [29]. Notably, the level of integration in resting cells approached the levels detected in activated cells after 3 days in culture (Figure 1B).
We next wanted to test if spinoculation enhanced infection primarily at the step of viral binding in resting cells rather than at integration as has recently been suggested [27]. We therefore measured viral binding, total HIV DNA, and integrated HIV DNA when resting cells were infected without spinoculation or when they were spinoculated at 300×g or 1200×g. Binding was 3–4 higher in cells spinoculated at 300×g and 13–16 fold higher in cells spinoculated at 1200×g. This increased binding resulted in a similar increase in reverse transcription and integration (Figure 1C). These data indicate that spinoculation enhances HIV infection at the step of viral binding and does not enhance reverse transcription or integration efficiency. Thus, there was no further enhancement downstream of viral binding.
We next tested whether cytokine stimulation would enhance the efficiency of integration in our system as has recently been described in other models [11], [22]. To do this, we cultured untreated resting CD4+T cells or prestimulated the cells for 3 days with IL-7, CCL19, or CD3/28 beads (Figure 1A) and then spinoculated the cells with HIV, after which we measured the ratio of integrated to total HIV DNA 48 hours post infection. We found that the efficiency of integration was similar in all conditions as calculated by the ratio of reverse transcripts that integrate (Figure 1D, E). The level of integration was approximately 5-fold higher in the CD3/28 treated cells (p = 0.018, Figure 1D), but this increase in integration was largely due to an increase in reverse transcription, leaving the integration efficiency unchanged. Thus, treatment with cytokines does not enhance integration efficiency in our model.
We then investigated whether stimulation enhanced integration efficiency when cells were not spinoculated. To test this we infected resting CD4+T cells or cells pre-stimulated with CCL19, IL-7 or CD3/28 beads with HIV (Figure 1A). We compared cells infected without spinoculation to those spinoculated at 300×g and 1200×g by measuring total and integrated HIV DNA. The efficiency of integration was never higher in the spinoculated samples and was similar in most cases (Figure 1D), although CCL19 treatment resulted in slightly lower integration frequency at 1200×g (∼2 fold, p = 0.01, Figure 1D). However, due to the small effect, we refrain from making any conclusions based on this decrease. We note that CD3/28 treated cells trended towards a higher efficiency of integration (Figure 1D), suggesting that integration efficiency may be higher in artificially activated cells. These results indicate that cytokine stimulation did not enhance integration efficiency even without spinoculation.
A final concern regarding our model was the possibility that spinoculation altered the activation state of the resting cells. We therefore assessed the effect of spinoculation on the resting phenotype of our cells using both activation marker expression and a glucose uptake assay since resting cells are known to consume less glucose than activated cells [30]. We found similar levels of activation marker expression (data not shown and [6]) as well as similar levels of glucose uptake in resting cells infected with and without spinoculation (data not shown). Therefore, spinoculating resting cells is a viable model system that enhances infection levels without affecting the susceptibility of the cells to HIV integration. Overall, our results suggest that the step of integration is not restricted in resting cells.
We next asked whether HIV infected resting CD4+T cells could produce viral proteins. Previous data has shown that phenotypically resting CD4+T cells can express SIV Gag in vivo in certain tissues [25], [26]. However, these results could not be repeated in vitro by directly infecting resting cells unless the cells were cultured in a lymphoid tissue microenvironment [18], [31]. This led to the prevailing belief that protein expression in resting cells was due to prior activation or exposure to a certain cytokine milieu. In fact, it was assumed that these Gag expressing resting cells found in vivo were actually productively infected and not latent [25]. However, Gag production does not necessarily mean that productive infection occurs as other viral proteins are required to make infectious particles.
To determine whether directly infected resting cells could produce Gag, we employed our spinoculation model to achieve a higher level of infection that would allow for easier detection of viral protein expression. We therefore spinoculated bulk unstimulated CD4+T cells, comprised of endogenously activated cells that expressed activation markers (HLA-DR, CD25, or CD69), and resting CD4+T cells that did not express any of these markers. We measured intracellular Gag production over time in both of these populations based on activation marker expression. We found that both resting and endogenously activated cells were capable of producing Gag above background levels (Figure 2A). However, these two populations did have slightly different kinetics (Figure 2A) consistent with our integration data (Figure 1B). We found Gag was expressed more rapidly in endogenously activated compared to resting cells (Figure 2A). However, integration levels were similar between the two populations (data not shown) indicating the cells were similarly susceptible to HIV integration.
We next examined whether these results were an artifact of spinoculation and if they could be repeated using another latency model. We therefore infected purified resting and CCL19 treated cells with or without spinoculation and monitored intracellular Gag expression. We detected Gag expressing resting cells in both latency models, with and without spinoculation (Figure 2B). Differences in Gag expression between cells infected with and without spinoculation reflected differences in integration levels (data not shown). This data suggests viral protein expression was not enhanced by spinoculation and indicates that spinoculation primarily increases viral binding (Figure 1).
We also examined which resting CD4+T cell subsets were capable of expressing Gag. To do this we sorted resting naïve, central memory, and effector memory CD4+T cells and infected them with HIV. Each subset was capable of Gag expression (Figure S1). These results indicate that resting CD4+T cells can produce Gag without stimulation in multiple CD4+T cell subsets.
We next questioned whether the Gag production in resting cells resulted in spreading infection. To test this, we infected purified resting CD4+T cells with HIV, cultured them for 4 days in the presence or absence of a protease inhibitor, which would prevent viral spread, and then stained the cells for intracellular Gag. If the cells were latent, no spreading infection should occur and there would be no difference between cells treated or not treated with the protease inhibitor. We found the amount of Gag was similar in the samples treated with or without a protease inhibitor 96 hours post infection (Figure 3A). Similar results were achieved up to 7 days post infection (data not shown). These data indicate that spreading infection does not occur in resting cells (at least to detectable levels). We confirmed that these resting cells were latently infected by stimulating out infectious virus with phytohemagglutinin (PHA) in the presence or absence of a protease inhibitor. We found more Gag positive cells when cells were activated for 48 hours in the absence of the protease inhibitor (Figure 3A), which suggests that infectious virus was released from these cells and resulted in spreading infection. Thus, the Gag producing resting cells were latently infected.
We wanted to confirm these results were not an artifact of a high viral inoculum so we infected cells with 15-fold less virus. Since protein differences would be harder to detect at this lower MOI, we chose to use a more sensitive PCR based assay to detect spreading infection. In addition, we wanted to test the ability of different cytokines to stimulate virus production. Therefore, we compared total DNA differences between cells treated with or without a protease inhibitor. If spreading infection were to occur, we would expect to see higher levels of total HIV DNA in the cells not treated with the protease inhibitor. First, we spinoculated resting and stimulated CD4+T cells (CCL19, IL-7, or CD3/28) as in Figure 1. We then cultured the cells in the presence or absence of a protease inhibitor and measured total HIV DNA in these fractions after 7 days of culture post infection (except for CD3/28 stimulated cells, which were collected 72 hours post infection). We found that resting CD4+T cells and CCL19 treated cells had the same levels of HIV DNA in the presence or absence of a protease inhibitor (Figure 3B), indicating these cells do not produce detectable infectious virus. On the other hand, cells treated with IL-7 or CD3/28 beads showed a statistically significant increase in HIV DNA without a protease inhibitor (Figure 3B). These data suggest that resting and CCL19 treated CD4+ T cells are latently infected while IL-7 and CD3/28 treated cells are capable of supporting efficient viral spread.
As resting cells are less susceptible to infection than CD3/28 stimulated cells, it was possible that the difference in spreading infection between resting and activated cells was due to this difference in susceptibility. We therefore tested for the presence of infectious virus in the supernatant of these two cell types. We infected resting CD4+T and CD3/28 activated cells with a high inoculum as in Figure 2. We then used the supernatant of both of these cells (collected at 72 hours, 96 hours, and 120 hours post infection) to infect (via spinoculation) the activated T cell line CEMss-GFP, which expresses GFP upon HIV infection [32]. No GFP expression was detected above background when using resting CD4+T cell supernatant from any time point, but GFP was detected when using supernatant from activated cells (Figure 3C and D). Thus, we were unable to detect release of infectious virus by resting cells with three complementary methods. Overall, our results indicate latently infected cells can produce Gag without producing detectable infectious virions. We therefore wanted to determine why these cells were not productively infected.
As Gag, Pol, and Env proteins are absolutely required for spreading infection and we saw Gag production (and thus likely Pol) in resting cells, we next questioned whether HIV envelope could be detected in these cells. To test this, we spinoculated purified resting cells and measured Env expression on the surface of these cells. Interestingly, resting cells produced little to no detectable Env protein above background levels (Figure 4 A and B, Env positive cells were undetectable in 1 of 3 experiments and barely detectable in 2 of 3 experiments), even though infection levels reached ∼70% based on integration levels (data not shown). IL-7 and CD3/28 treated cells, on the other hand, consistently led to detectable and significantly higher levels of Env positive cells (3.4% and 24.6%; p = 0.015 and 0.0018 respectively, Figure 4B). Thus, infected cells capable of spreading infection were more frequently positive for Env. It is possible that resting cells may produce low levels of Env that might be sufficient for low levels of spreading infection. However, if viral spread occurs in resting cells, it occurs at very low levels not detectable with our assays.
We then examined Gag production as a control to see if IL-7 and CD3/28 had similar enhancements on Gag expression, or if enhanced translation was limited to Env. We found on average 4.1% of resting cells were positive for HIV Gag (Figure 4B) while IL-7 and CD3/28 treatment led to statistically higher levels of Gag positive cells (p = 0.032 and 0.0059, respectively), indicating enhanced protein expression was not limited to Env. We found that CD3/28 activated cells had not only a ∼10 fold higher frequency of Gag expression than resting cells but also ∼10 fold higher levels of Gag expression (MFI, Figure 4C, p = 0.0005). These data are consistent with studies in vivo suggesting resting cells release produce approximately 10 fold less Gag than activated cells during SIV infection [26]. Our results suggest that stimulation may globally upregulate HIV protein expression and not just increase one particular viral protein.
We next examined if the differences in protein levels were due to differences in transcription and/or splicing. We first measured the levels of gag RNA to compare with our protein measurements. Because virions contain gag RNA, we needed to compensate for background RNA from incoming virions as described in the Methods. We calculated the amount of unspliced RNA in resting cells to be 300 copies per integrated HIV (Figure 5, S3). There were significantly higher levels of gag in IL-7 and CD3/28 treated cells (1000 and 12000 copies per integrated HIV, p = 0.015 and p = 0.032 respectively), which corresponds with the protein levels in Figure 4 and are consistent with the different RNA levels detected in vivo [25].
We next tested if levels of envelope protein also corresponded with the amount of env transcripts in resting cells. We measured env levels using the primers depicted in Figure 5A. Resting cells contained 0.5 copies per integrated HIV while IL-7 treated cells produced approximately 10 fold more env RNA and CD3/28 treated cells produced roughly 130 fold more env RNA (Figure 5B, p = 0.017 and 0.088 respectively). These results indicate that Env protein levels matched env RNA levels, suggesting a pre-translation block in resting CD4+T cells.
We next tested whether other spliced transcripts were also low in resting cells. We began by measuring levels of vif since it is singly spliced and the only transcript that encodes the Vif protein. Vif RNA levels followed the same pattern as gag and env with both IL-7 and CD3/28 treated cells trending to produce more RNA (approximately 6 and 60 fold) than resting cells (2.1 copies/integrated HIV), (Figure 5B, p = 0.107, p = 0.084 respectively). These results suggest that the effects of stimulation were similar among various splice products and that a specific block to certain spliced forms did not explain the lack of productive infection in resting cells.
We then simultaneously quantified both tat and rev transcripts, designated tat/rev. Tat is the major transcriptional regulator of HIV while Rev promotes RNA export out of the nucleus. Tat/rev RNA was barely detectable in resting cells (0.01 copies per integrated HIV, Figure 5B). IL-7 and CD3/28 treatment resulted in significantly more tat/rev transcripts (20 and 130-fold more, p = 0.043 and p = 0.0005 respectively). Thus, it is possible that these higher tat/rev levels could explain transcriptional differences between resting and stimulated cells.
Since it remained possible that spinoculation artificially enhanced HIV RNA expression, we repeated our RNA measurements in resting cells infected with and without spinoculation (Figure 5C). We quantified env and vif expression as their quantitation is more robust than gag since there is no contribution from incoming virus. We found similar levels of both env and vif per integrated HIV in cells infected with or without spinoculation (Figure 5C). Thus, our viral transcription data is not an artifact of spinoculation.
Overall, our RNA data indicate that while substantial levels of unspliced transcripts exist in resting CD4+T cells, there are significantly less spliced messages formed in these cells. This suggests that resting cells may not have sufficient levels of spliced products required for spreading infection, therefore providing a potential mechanism for the absence of viral spread in these cells. Enhanced transcription in activated cells not only increases the level of gag transcripts, but also increases viral spliced products to a level that allows efficient spread.
Since we detected viral protein production in a population of resting cells, we asked why some cells produced viral protein and others did not. As integration site selection could explain a difference in protein production, we tested whether HIV integration site selection was different in Gag positive and negative cells. To do this, we sorted infected CD4+T cells into 4 populations based on activation status (resting and activated) and Gag expression (Gag positive and Gag negative), confirmed infection in each subset using HIV DNA measurements, and analyzed where in the human genome HIV integrated in each of these populations (Figure 6A). We found statistically significant differences between Gag positive and Gag negative resting cells (Figure 6B), reaffirming that these are two distinct populations. Surprisingly, the differences were similar to those found between Gag positive and negative activated cells (Figure 6B). Our data indicate that there were greater similarities in integration site selection among any Gag producing cells than among cells with the same activation state. This suggests that differences in integration site selection were in fact reflective of viral protein production. However, these differences were small. There was also a modest increase in integration frequency near heterochromatic alphoid repeats in the Gag negative subsets (data not shown), paralleling previous studies [33], [34]. Nevertheless, even in cells not expressing viral proteins, HIV still preferred to integrate in active genes, just to a lesser extent than in Gag producing cells. This pattern was also seen for other genomic features associated with active genes such as GC content and CpG islands (Figure 6, [35]). Overall, the differences in integration site selection explain a small part of the differences between Gag positive and negative cells.
We have previously shown that resting CD4+T cells are susceptible to HIV integration without stimulation [6]. Here, we employed a latency model that achieves high levels of infection without requiring stimulation to identify a population of latently infected resting CD4+T cells that express Gag but do not support viral spread. The block to productive infection corresponded with barely detectable levels of envelope protein, potentially due to low levels of tat in these cells. Thus, our latent model reveals that there is a continuum of latent cells from cells in a pre-integration state of latency (not focused on in this report) to translationally silent cells to cells expressing HIV proteins without spreading infection. The ability of some latent cells to produce viral proteins has important implications for therapies targeting the latent reservoir as these cells could be recognized by a robust immune response.
Previous work has shown that phenotypically resting cells can produce HIV and SIV RNA and protein in vivo in certain tissues [25], [26]. However, it is impossible for in vivo studies to determine at what activation state the resting cells are actually infected, whether the in vivo cytokine milieu is required for the expression of viral proteins, and whether the cells are productively infected. Here we show by RT-PCR and flow cytometry that directly infected resting cells can produce HIV Gag in vitro without additional stimulation while remaining latently infected. Our results also reveal that CCL19 does not enhance integration efficiency (Figure 1), as has been shown recently [10]. This discrepancy is not due to spinoculation or to cytokine concentration as we repeated our results without spinoculation using several doses of CCL19, up to 1 µM, and confirmed the activity of our cytokine preparation in a chemotaxis assay (data not shown). It is possible that differences in sera used, potency of the virus, or the sensitivity of our integration assay could explain this difference.
We show that the majority of resting Gag positive cells are latent in that they are unable to support productive infection based on similar protein and DNA measurements in protease treated and untreated fractions (Figure 3 A,B). Additionally, the supernatant of infected resting cells cannot infect CEMss-GFP cells (Figure 3C, D). However, these cells can release infectious virus upon activation (Figure 3A). The block to productive infection appears to be due, at least in part, to the barely detectable levels of envelope protein in these cells (Figure 4). IL-7 and CD3/28 treatments were able to overcome this block and resulted in both higher levels of envelope protein and spreading infection. Our results are consistent with the extensive evidence that resting cells are not capable of productive infection in the absence of various stimuli (reviewed in [15]). While it is conceivable that a small number of resting cells are able to release infectious virions in our model, these cells are so few that we could not detect viral spread. Additionally, as our cultures can never be 100% pure, it is almost impossible to conclusively prove that any low level spreading infection results from resting cells instead of contaminating activated CD4+T cells. Nonetheless, it is clear that infected resting cells are very inefficient at viral spread and this may be the most important difference between HIV infection in resting and activated CD4+T cells.
Although we found that resting CD4+T cells in our model do not support productive infection, calling these cells latent may still seem controversial. Several authors claim that any cells producing viral proteins are not latent since protein production will result in elimination of the cell due to cytotoxic effects or immune clearance [34], [36]. With regards to cytotoxicity, the Gag positive resting cells described here produce both quantitatively less and fewer types of viral proteins than activated cells (Figure 4) and so may be better able to survive any cytotoxic effects of the viral proteins. Notably, Env is particularly known for its toxicity and so low Env expression should enhance the longevity of resting cells [37], [38]. Additionally, the longer half-lives of resting cells may allow them to survive better in the face of protein production than their activated counterparts [39]. Furthermore, recent work has shown that latently infected cells from patients on HAART produced Gag when treated with SAHA but were not killed by HIV cytopathic effects and had substantially longer half-lives than expected [40]. With regards to immune clearance, since resting cells are producing less protein than activated cells, they may be harder for the immune system to clear. Just because a cell may be an immune target does not mean it will be successfully cleared, especially since the immune system declines over the course of HIV infection [41] and the frequency of HIV-specific CD8+T cells decreases on HAART [42], [43]. A recent study showed that while CD8+T cells from elite suppressors (ES) could clear latent cells stimulated to produce HIV Gag in vitro, the same was not true for patients on HAART [40]. Overall, while some Gag positive resting cells may die due to cytotoxicity or to immune clearance, it is likely that a subset of these cells will survive. This is consistent with the fact that resting cells containing viral RNA are detected on HAART [19], [44]. Since the cells do not produce infectious virions and likely persist on HAART, they should be considered latent.
As we found Gag but little Env production in resting cells, the obvious question was whether this difference was due to blocks at transcription, splicing, and/or other post-transcriptional steps. We found that the regulation of protein expression was mainly at the level of viral transcripts (Figure 5). gag RNA levels were substantially higher than that of any spliced transcript in resting cells, suggesting splicing frequency may be low in resting cells; however, stimulation enhanced the amounts of gag, env, vif, and tat/rev (unspliced and spliced RNA species) to similar extents (Figure 5). Our gag levels were similar to those described in vivo [25]. These data show that activation globally enhances levels of all viral transcripts, suggesting there is no specific block to any particular spliced form in resting cells. Instead, each splice product was likely controlled by the strength of the splice site, allowing higher levels of some transcripts over others. For example, it is known that the tat splice site is relatively weak, which may explain the low levels of tat in resting cells [45].
The higher levels of transcription in activated cells corresponded with an increase in tat/rev RNA levels, which were barely detectable in resting cells (Figure 5). Transcriptional efficiency in resting cells was thus not likely due to Tat but to other transcription factors. NFAT and NF-κB are important regulators of HIV transcription and are known to have low activity in resting cells [15], [46]. But how then do we explain any transcription in these cells much less protein expression? First, it is possible that other transcription factors, such as TCF, may be involved in HIV transcription and translation in quiescent cells [47], [48]. Second, it is possible that transcriptional activity from surrounding genes could explain HIV transcription in certain cells, particularly if the HIV provirus is integrated in a transcription unit in the same orientation as the host gene [33], [46], [49]. This would suggest integration site selection would play a role in Gag expression. Indeed, we found that there were small but statistically significant differences in integration site selection between Gag expressing and non-expressing resting cells (Figure 6). However, the differences were small and unlikely to fully explain differences in viral protein expression.
A related question is how Gag expression occurs without Rev, which plays a critical role in nuclear export of unspliced and singly spliced HIV RNA (reviewed in [50]). First, it is possible that low levels of Rev are expressed in resting cells allowing successful nuclear export of gag; however, the low levels of rev RNA makes this unlikely. Second, it is possible that due to low levels of HIV splicing in resting cells (Figure 5), enough gag RNA accumulates to allow inefficient export into the cytoplasm. As we see substantial gag levels in resting cells (∼300 copies of gag/integrated HIV), there may be sufficient quantities of unspliced RNA to enable some RNA to exit the nucleus. Our data is consistent with published data showing higher levels of unspliced than spliced RNA in resting CD4+T cells and PBMC in patients on HAART [44], [51]–[53]. If HIV splicing were to occur efficiently in resting cells, one would expect spliced forms would be more prevalent than unspliced RNA due to the low levels of rev (Figure 5 and [52]). Additionally, our data is consistent with data from Zack and colleagues [54] showing that while spliced forms are detected earlier in pre-stimulated HIV infected cells, latent cells that are activated produce detectable gag RNA before spliced forms.
Overall, our RNA data agree with prior literature, as described above. Furthermore, our data are likely consistent with a described block to nuclear export of HIV RNA in resting cells [19], [22] as we see the expected pattern of high gag RNA levels but a substantially lower percentage of Gag positive cells (∼4%, Figure 4). Nonetheless, the export block was not absolute as proteins were translated, indicating a fraction of RNA is transported to the cytoplasm. While sufficient unspliced RNA must exist in resting cells to result in nuclear export and translation of Gag, the same could not be said for Env. In fact, a nuclear export block is consistent with our data that we do not see substantial Env expression even though there are approximately 0.5 env copies per integrated HIV DNA (Figure 5). Our model therefore suggests the block to productive resting cell infection is not absolute but is instead a series of less efficient steps that result in significantly different outcomes in resting and activated cells. Thus, resting cell infection mainly results in latent infection while infection in activated cells primarily results in productive infection.
Our spinoculation model generated higher levels of infection than typically obtained by other models (e.g. 70% of cells contained integrated HIV). This high frequency of infection was essential to demonstrate HIV proteins were expressed in a subset of latently infected cells. However, as cells are not spinoculated in vivo, we needed to confirm that spinoculation did not artificially affect our results, as has recently been suggested [27]. While our prior studies indicated that the major effect of spinoculation was at the level of binding [55], we show here that spinoculation did not enhance integration efficiency (Figure 1C,D) or transcriptional efficiency (Figure 5C) in resting cells, consistent with our prior reports showing similar efficiencies of integration per bound virion with or without spinoculation, even at low viral inoculums [24]. As previously mentioned, our RNA and protein data are also consistent with in vivo results (Figure 2, 4 and 5, [19], [25], [26], [44], [51]). Furthermore, our data were not the result of an artificially high viral inoculum as similar outcomes of infection were seen with or without spinoculation at the level of integration (Figure 1), transcription (Figure 5C), translation (Figure 2B) and viral spread (Figure 3B). While signaling may occur due to spinoculation and the physiology of the cells may be altered, our data show that these changes do not affect the course or efficiency of HIV infection in these cells. Overall, our data suggests that spinoculation is a valid and useful model for studying HIV latency as it only affects the frequency of infected cells that progress through the HIV life cycle rather than the course of infection.
The ability of latently infected resting cells to express Gag protein without spreading infection has important implications for treating the latent reservoir. First, as some latently infected resting cells can produce protein, they may be cleared by the immune system, particularly in patients with strong immune responses to HIV such as ES [56], [57]. This could potentially explain the lower reservoir levels in these patients [58]. Our data suggest that developing strategies to boost the immune response in patients could be vital in clearing the latent reservoir. Second, since resting cells can produce Gag protein but much less Env, strategies targeting Gag producing cells may have added benefits compared to therapies targeting Env expressing cells. Currently, several studies have begun targeting Env producing cells through immunotoxin approaches (reviewed in [59]). Since resting cells do not produce substantial Env levels, these types of therapy would be unable to target latent resting CD4+T cells without stimulation; however, therapies targeting Gag may be able to eliminate some latent cells. Finally, our data indicate that resting cells can produce protein without spreading infection. Thus it may be possible to develop therapies that will stimulate cells to produce enough protein to be cleared without causing ongoing replication.
Targeting and eliminating the latent reservoir has become an attractive approach for curing HIV. Here we describe a direct infection in vitro latency model that allows sufficient infection levels for study without requiring artificial stimulation and the related physiological consequences on the cells. Using this model, we show resting cells are capable of producing Gag protein without spreading infection. Our model thus contains an extensive range of latently infected resting CD4+T cells from cells in a pre-integration latent state to cells producing protein without spreading infection. By including all of these populations, our model may better represent HIV latency in vivo. Through the use of spinoculation, we can generate sufficient levels of infected cells to study the differences between latent cells capable of Gag production and cells incapable of protein expression. Characterizing these differences may lead to therapies that could turn translationally silent cells into Gag expressing latent cells that could be cleared by a robust immune response.
Primary human CD4+T cells used in these studies were obtained through anonymous donation to the University of Pennsylvania's Center for AIDS Research Human Immunology Core after written informed consent and approval by the University of Pennsylvania's institutional review board.
Unstimulated CD4+T cells were purified from leukapheresis-enriched PBMC using Rosette Sep (Stemcell Technologies) and were obtained from the University of Pennsylvania's Center for AIDS Research Human Immunology Core. To purify resting CD4+T cells, cells were stained with PE labeled antibodies against CD25, CD69, and HLA-DR (BD Biosciences) and anti-PE magnetic beads (Miltenyi Biotec) as recommended by the manufacturer. Cells were then depleted using LD columns as recommended by the manufacturer (Miltenyi Biotec). Resting CD4+T cells were typically greater than 98% pure.
Resting CD4+T cells were either left untreated in RPMI containing 10% heat inactivated FCS (Invitrogen, Qualified) supplemented with 100 µg/mL penicillin-streptomycin and GlutaMax (Invitrogen), or were stimulated with 20 ng/mL IL-7, 100 nM CCL19 (R&D Biosystems) or with CD3/28 beads at a concentration of 3 beads/cell (Invitrogen) with 100 U/mL IL-2 (R&D Biosystems) for three days. Cells were then spinoculated at a concentration of 1×107 cells/mL in viral transfection supernatant (MOI of 3 as assessed by infection of CEMss-GFP cells, unless otherwise noted) for 2 hours at 1,200×g at 25°C (unless otherwise indicated). NL4-3 viral stocks were prepared by 293T transfections by the University of Pennsylvania's Center for AIDS Research Viral/Molecular Core. After spinoculation, cells were washed twice in CO2 independent media (Invitrogen) and treated with 50 µg/mL Dnase I (Roche) and 10 mM MgCL2 to remove plasmid DNA. Cells were then cultured in the presence of 1.25 µM of the protease inhibitor saquinavir (Roche) to prevent viral spread (unless otherwise noted). For experiments measuring Env protein and RNA, cells were infected in the presence of 8 µg/mL polybrene (Millipore), excluding infections of resting cells without spinoculation.
Viral binding was estimated by measuring cell-associated p24 via a p24-specific enzyme-linked immunosorbent assay (Perkin-Elmer) as previously described [23].
DNA was prepared after infection using the QIAamp DNA Micro Kit (Qiagen). Real-time PCR was used to detect total HIV DNA (complete SST), β-globin, and integrated DNA as previously described [24].
RNA was isolated using Tri-Reagent (Sigma Aldritch) as per the manufacturer's recommendations in the presence of 10 µg/mL Glycogen (Roche). Two chloroform extractions were performed and 75% of the RNA fraction was collected both times to ensure RNA purity and to achieve a predictable yield. Cells were counted prior to isolation using the Countess (Invitrogen) automated cell counter. Using the predicted percent RNA yield described above, cell concentrations were calculated. These counts were independently confirmed by 18S RNA measurements based on cell type (resting CD4+T cells, CD3/28 activated cells etc). RNA copies were quantified per cell based on these calculations.
Gag standards were generated as previously described [55]. Standards for vif, env and tat/rev were all generated in the following way. Reverse primers, as explained in Figure 4A, for vif, env or tat/rev were used to generate cDNA from RNA isolated from our CEMss integration standard cell line [60]. To generate cDNA, 200 ng of total RNA was added to a reaction with the reverse primer only in a master mix following the High Capacity cDNA Reverse Transcription kit protocol (Applied Biosystems). The following cycling conditions were used: 25°C for 10 minutes, 37°C for 120 minutes and 85°C for 5 minutes. The cDNA was then diluted 1∶10 in 10 mM Tris-HCl pH 8.0 and 4 µL of this dilution was added to a PCR reaction with the following: 1× buffer (Invitrogen), 3.5 mM MgCl2, 300 µM dNTPs, 100 nM forward (Figure 4A) and reverse primers and H2O to a total volume of 20 µL. The following cycling conditions were used: 95°C for 10 minutes and 40 cycles of 95°C for 30 seconds, 55°C for 30 seconds and 72°C for 1 minute. The resulting products were run on a 2% agarose gel. Appropriate sized bands were excised and purified with the QIAquick gel extraction kit (Qiagen). This purified cDNA was then cloned into the pCR2.1 TOPO vector following the TOPO TA cloning kit (Invitrogen). Chemically competent E. coli cells (Invitrogen) were transformed with the plasmids and grown on LB agar (Becton Dickinson) with ampicillin (Sigma) overnight at 37°C. Colonies were selected, shaken overnight in LB broth (Becton Dickinson) with ampicillin at 37°C. Plasmids were extracted with the Qiaprep Spin Miniprep Kit (Qiagen). A digestion with EcoRI (New England Biolabs) was performed to confirm presence of the appropriate sized band. Clones with appropriate bands were also sequenced to verify we obtained the appropriate RNA splice forms. After sequence confirmation, appropriate clones were grown in large cultures overnight in LB broth with ampicillin at 37°C. Plasmids were isolated following the PureLink HiPure Plasmid Filter Maxiprep Kit (Invitrogen) and eluted into 500 µL of Tris-HCl pH 8.0. Confirmatory digestion and sequencing were again performed. Plasmids were then linearized by digestion with SpeI (New England Biolabs). The SpeI enzyme was heat inactivated after digestion for 20 minutes at 80°C. Plasmids were then in vitro transcribed to generate RNA using the HiScribe T7 in vitro Transcription Kit (New England Biolabs). RNA was purified using the RNeasy MinElute Cleanup kit (Qiagen) including the optional on column DNase treatment (Qiagen) to remove leftover plasmid DNA. Finally, RNA was measured by spectrophotometry and the copy numbers were calculated based on the concentration and number of bases per RNA transcript. We confirmed our primers solely detected their respective products via gel electrophoresis.
RT-PCR reactions were performed using a one-step reaction at 20 µL total volume using the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems). Reactions contained 280 nM dNTP, 1.68 nM of each primer, and 0.56 nM of probe. Platinum Taq polymerase (Invitrogen) was used at 0.75 U/reaction and reverse transcriptase was used at 7.5 U/reaction (Applied Biosystems). The RT-PCR was run on an ABI 7500 Fast Instrument with the following protocol: 1) 95°C for 30 min 2) 95°C for 15 s 3) 60°C for 30 s 4) 72°C for 1 min. Steps 2–4 were repeated for 40 cycles. RT-PCR for env, vif, and tat/rev were all performed with the RU5 forward primer (RF) 5′GCCTCAATAAAGCTTGCCTTGA-3′ and the probe 5′CCAGAGTCACACAACAGACGGGCACA-3′. The reverse primer for env (ER) was 5′-GATTACTATGGACCACACAACTATTG-3′. The reverse primer for vif (VF) was 5′-CCATGTGTTAATCCTCATCCTGTC-3′. The reverse primer for tat/rev was 5′-CTTCTTCCTGCCATAGGAGATGCC -3′. gag was detected using the forward primer (GF) 5′-AGTTGGAGGACATCAAGCAGCCATGCAAAT-3′, the reverse primer (GR) 5′-YGCTATGTCAGTTCCCCTTGGTTCTCT-3′, and the probe 5′-GCGAGCGAGACCATCAATGAGGAAGCTGCAGA-3′.
As has recently been described [22], measuring gag levels in resting cells is difficult since the viral genome, like gag, is unspliced. Therefore, gag levels were calculated by subtracting the gag/integrated DNA signal in cells treated with 1 µM of the integrase inhibitor raltegravir (AIDS Reagent Program) from the gag/integrated DNA signal in uninhibited infected cells.
For Gag staining, cells were fixed and permeabilized using the Fix and Perm Cell Permeabilization Kit (Invitrogen) as recommended by the manufacturer and were intracellularly stained with a KC57-FITC antibody (Beckman Coulter). For HIV Env staining, cells were incubated with a gp120 clone 2G12 antibody (AIDS Reagent Program). Cells were then washed twice with PBS. Next, cells were stained with a mouse anti-human IgG PE conjugated antibody (Southern Biotech). Control cells were treated with 1 µM efavirenz (AIDS Reagent Program) post inoculation to serve as a negative control for flow cytometry gating. To ensure Env signal was a result of actual production by the cell and not release of virus and rebinding to a neighboring cell, cells were cultured in the presence of 25 µg/mL anti-CD4 clone 19 (generously donated by Ron Collman).
For integration site selection analysis, infected unstimulated CD4+T cells were stained with intracellular Gag and activation markers as described above. Cells were sorted using a FACSAria II Cell Sorter (BD Bioscience) into four populations: activated Gag− (92% pure), activated Gag+ (95% pure), resting Gag− (99% pure), resting Gag+ (97% pure). The main contaminant for each population was resting Gag− cells.
Uninfected PBMC were stained with FITC-conjugated lineage markers (CD8, CD11c, CD14, CD16, CD20, CD56, BDCA-2) and PE-conjugated activation markers (CD25, CD69, HLA-DR) to mark resting CD4+T cells (BD Bioscience). Cells were also labeled with CD45RO PE-Texas Red (Beckman Coulter) and CCR7 PerCpCy5.5 (BD Bioscience) to distinguish naïve (CD45RO−, CCR7+), central memory (CD45RO+, CCR7+) and effector memory cells (CD45RO+, CCR7−). These resting subsets were sorted using a FACSAria II Cell Sorter (BD Bioscience). Naïve cells were typically 99% pure while central memory and effector memory cells were typically >95% pure. Sorted subsets were then infected with NL4-3 and intracellular p24 was quantified 72 hours post infection as above.
Genomic DNA was digested overnight using MseI and Tsp509I. Fragments from each sample were then ligated overnight at 16°C to their own unique PCR adapter. To isolate integration site junctions, two rounds of PCR were performed as previously described [61] with a set of nested primers specific for each linker and the viral LTR. Amplification products were sequenced using 454 sequencing. DNA barcodes were included in the nested LTR primers to allow sample pooling prior to sequencing [61], [62].
DNA sequences that contained an exact match to the terminal LTR sequence (TCTAGCA; lies between the LTR primer and the site of integration), aligned within three base pairs of the beginning of the sequence read, and had a single best alignment with ≥98% identity to the human genome (hg18, version 36.1) were counted as true integration events. Random genomic sites were computationally selected for comparison (matched random controls, [63]). Gene expression analysis was based on microarray data from T cells [64]. All sequences will be deposited in the SRA repository upon acceptance of this manuscript for publication.
A one-tailed Student's t-test with the Holmes correction was used to compare statistical differences between experimental conditions (Graphpad Prism). For the spinoculation comparisons in Figure 2E, a one-way ANOVA was performed with a Bonferoni post test. Logistic regression and other statistical methods as described in [35], [63] were used to compare distributions of integration sites to those of genomic features.
|
10.1371/journal.ppat.1000856 | Antagonism of Tetherin Restriction of HIV-1 Release by Vpu Involves Binding and Sequestration of the Restriction Factor in a Perinuclear Compartment | The Vpu accessory protein promotes HIV-1 release by counteracting Tetherin/BST-2, an interferon-regulated restriction factor, which retains virions at the cell-surface. Recent reports proposed β-TrCP-dependent proteasomal and/or endo-lysosomal degradation of Tetherin as potential mechanisms by which Vpu could down-regulate Tetherin cell-surface expression and antagonize this restriction. In all of these studies, Tetherin degradation did not, however, entirely account for Vpu anti-Tetherin activity. Here, we show that Vpu can promote HIV-1 release without detectably affecting Tetherin steady-state levels or turnover, suggesting that Tetherin degradation may not be necessary and/or sufficient for Vpu anti-Tetherin activity. Even though Vpu did not enhance Tetherin internalization from the plasma membrane (PM), it did significantly slow-down the overall transport of the protein towards the cell-surface. Accordingly, Vpu expression caused a specific removal of cell-surface Tetherin and a re-localization of the residual pool of Tetherin in a perinuclear compartment that co-stained with the TGN marker TGN46 and Vpu itself. This re-localization of Tetherin was also observed with a Vpu mutant unable to recruit β-TrCP, suggesting that this activity is taking place independently from β-TrCP-mediated trafficking and/or degradation processes. We also show that Vpu co-immunoprecipitates with Tetherin and that this interaction involves the transmembrane domains of the two proteins. Importantly, this association was found to be critical for reducing cell-surface Tetherin expression, re-localizing the restriction factor in the TGN and promoting HIV-1 release. Overall, our results suggest that association of Vpu to Tetherin affects the outward trafficking and/or recycling of the restriction factor from the TGN and as a result promotes its sequestration away from the PM where productive HIV-1 assembly takes place. This mechanism of antagonism that results in TGN trapping is likely to be augmented by β-TrCP-dependent degradation, underlining the need for complementary and perhaps synergistic strategies to effectively counteract the powerful restrictive effects of human Tetherin.
| Restriction factors are cellular proteins that interfere with the multiplication and transmission of viruses and are therefore important components of natural immunity. Tetherin (also known as BST-2) is a recently identified restriction factor that traps viruses at the cell-surface, preventing their release and thus infection of other cells. Viruses have, however, developed means to counteract this restriction factor. Viral protein U (Vpu) is an accessory protein encoded by HIV-1, the causative agent of AIDS. Vpu antagonizes Tetherin and consequently promotes the release of HIV-1 particles. A series of recent reports proposed that Vpu would induce the degradation of this restriction factor in order to overcome its anti-viral activity. Here, we report that Vpu is able to enhance HIV-1 release in absence of Tetherin degradation. Instead, we found that Vpu interacts with Tetherin and interferes with the transport of the restriction factor towards the cell-surface. This would lead to re-localization of Tetherin in an intracellular organelle called the trans-Golgi network, resulting in insufficient levels of Tetherin at the cell-surface to trap progeny viruses. This mechanism of antagonism that results in TGN trapping could be augmented by the induction of degradation to effectively counteract the powerful restrictive effects of human Tetherin. Further characterization of this mechanism will improve our understanding of host antiviral defenses as well as provide new targets for the development of novel anti-HIV drugs.
| Recent advances in retrovirology have revealed that mammalian cells do not always provide a hospitable environment for the replication of viruses that parasitize them. It is indeed becoming increasingly clear that mammalian cells express a variety of molecules and activities that interfere with specific steps of the replication cycle of retroviruses and other viruses [1]. Among these so-called restriction factors, the cellular protein CD317/BST-2/HM1.24, also designated as Tetherin in reference to its ability to tether HIV-1 virions to infected cells, was recently identified as a potent inhibitor of the release step of retroviruses [2],[3]. Tetherin is a heavily glycosylated type II integral membrane protein with an unusual topology in that it harbors two completely different types of membrane anchor at the N- and C-terminus; it is composed of a short N-terminal cytoplasmic tail linked to a transmembrane anchor (TM), an extracellular domain that include three cysteine residues important for dimerization, a predicted coiled-coil and a putative C-terminal glycophosphatidyl-inositol (GPI)-linked lipid anchor that is believed to ensure incorporation of Tetherin into cholesterol-rich lipid rafts [4],[5]. Tetherin inhibits the release of widely divergent enveloped viruses, including members of the lentivirus (primate immunodeficiency viruses), gammaretroviruses (murine leukemia virus), spumaretrovirus (foamy virus), arenavirus (Lassa virus), filovirus (Ebola and Marburg virus) families as well as Kaposi's sarcoma herpesvirus (KSHV) [2],[3],[6],[7],[8],[9],[10]. This broad-spectrum inhibition of enveloped virus particle release by Tetherin indicates that this restriction is unlikely to require specific interactions with viral proteins. In that regard, recent evidence indicates that Tetherin configuration rather than primary sequences is critical for antiviral activity since an entirely artificial Tetherin-like protein consisting solely of domains from three proteins that were analogous to Tetherin in terms of size and topology but lacking sequence homology with native Tetherin, inhibited particle release in a manner strikingly similar to Tetherin [11]. Tetherin-mediated restriction of virus particle release is believed to occur at sites of virus particle assembly at the plasma membrane since a strong co-localization between Tetherin and nascent particles generated from retroviral or filoviral structural proteins was observed at the cell-surface [7],[12]. In fact, recent findings using the artificial Tetherin-like protein support a model of restriction in which Tetherin directly cross-links virions to the plasma membrane [11]. Under basal conditions, Tetherin is expressed in B and T cells, plasmacytoid dendritic cells and myeloid cells and many transformed cell lines [2],[13],[14],[15],[16]. In addition, Tetherin expression is induced in many cell-types by type I and type II interferon (IFN), which suggests that it might be an important component of a broader antiviral innate immune defense [2],[13],[17]. In response to this restriction, many viruses express Tetherin antagonists such as KSHV K5, Ebola virus envelope glycoprotein (GP), simian immunodeficiency virus (SIVmac/smm) Nef, HIV-2 Env, SIVtan Env and HIV-1 viral protein U (Vpu), which was the first anti-Tetherin factor identified [2],[3],[9],[10],[18],[19],[20],[21],[22].
Vpu is an oligomeric type 1 integral membrane protein with two major activities during HIV-1 infection [23]. It contributes to the down-regulation of the CD4 receptor by targeting newly synthesized CD4 molecules that are bound to envelope glycoproteins (Env) in the endoplasmic reticulum (ER) for degradation by the ubiquitin-proteasome system [24],[25]. This degradation process relies on Vpu ability to associate with CD4 and to recruit β-TrCP, a component of the SCFβ-TrCP E3 ubiquitin (Ub) ligase, via phosphorylation of serines 52 and 56 within its DSGΦXS β-TrCP recognition motif [26],[27]. In addition, Vpu promotes HIV-1 particle release by suppressing human Tetherin activity in restrictive Tetherin-expressing cells such as epithelial cell lines (HeLa), T cell lines (Jurkat, CEM) and primary T lymphocytes and macrophages [2],[3],[17]. In contrast, no effect of Vpu is observed in permissive human cell lines devoid of Tetherin expression such as HEK 293T and HT1080. Interestingly, Vpu does not exert its anti-Tetherin activity in non-human cell lines regardless of their Tetherin expression levels [6],[28]. Indeed, although Tetherin variants found in rhesus macaques, African green monkeys (agm) and mouse cells are able to inhibit HIV-1 particle release, they are resistant to antagonism by HIV-1 Vpu [29],[30]. Analysis of Tetherin variants encoded by different species highlighted positively selected determinants in the Tetherin TM domain responsible for conferring sensitivity to Vpu antagonism [29],[30],[31],[32]. The mechanism by which Vpu counteracts Tetherin antiviral activity on HIV-1 particle release is still a matter of debate. Vpu was found to decrease the expression of Tetherin at the cell-surface [3] and to prevent Tetherin and Gag co-localization at sites of particle assembly [7],[12], suggesting that removal of Tetherin from its site of tethering action could underlie the mechanism by which Vpu counteracts this cellular restriction, although this model has lately been challenged [33]. Recently, a series of reports proposed proteasomal and/or endo-lysosomal degradation of Tetherin through a β-TrCP-dependent process as potential mechanisms by which Vpu antagonizes Tetherin antiviral activity [12],[20],[30],[31],[34],[35]. However, in all of these studies, Vpu-induced Tetherin degradation did not entirely account for the anti-Tetherin activity of Vpu. Thus, the precise mechanism(s) through which Vpu antagonizes Tetherin is yet to be elucidated.
In this study, we investigated the effect of Vpu on Tetherin expression and trafficking to obtain a better insight into the mechanism through which Vpu antagonizes Tetherin-mediated restriction of HIV-1 particle release. Here, we provide evidence that Vpu can promote HIV-1 particle release without affecting the total steady-state levels or the turnover rate of Tetherin. We further show that even though Vpu did not enhance Tetherin internalization from the plasma membrane, it did significantly slow-down the transport of the restriction factor towards the cell-surface. Notably, expression of Vpu led to a specific removal of cell-surface Tetherin and a re-localization of the residual pool of Tetherin to a perinuclear compartment that extensively overlapped with the TGN. Finally, we show that Vpu and Tetherin associate most probably via their TM domains and provide evidence that this association is necessary to relocate Tetherin from the cell-surface to the TGN and to counteract its restrictive activity on HIV-1 release. Overall, our results are consistent with a model whereby antagonism of Tetherin by Vpu involves sequestration of the restriction factor in a perinuclear compartment, away from virus assembly sites on the plasma membrane, a process that could be augmented by the induction of Tetherin degradation.
To assess whether the reduction of Tetherin levels by Vpu was necessary and sufficient to promote efficient HIV-1 particle release, we analyzed the steady-state levels of exogenously-expressed HA-Tetherin in permissive HEK 293T cells in conditions where varying levels of virally-encoded Vpu was co-expressed (Fig. 1A). This cellular system was previously used to evaluate the effect of Vpu on Tetherin steady-state levels [30],[31],[34]. At low Vpu expression levels (1 µg of Vpu+ proviral construct), the levels of Tetherin were essentially similar to those detected in absence of Vpu (1 µg of Vpu-defective proviral construct) (compare lane 5 with lane 6), while at higher levels of Vpu expression (2 µg of Vpu+ proviral construct), they were significantly reduced (compare lane 3 with lane 4). As expected, ectopic expression of HA-Tetherin strongly inhibited the release of Vpu-defective HIV-1 particle relative to the Tetherin-negative control as demonstrated by the drastic reduction of virion-associated p24 in the supernatant (Fig. 1A, compare lanes 3 and 5 with lane 1; quantified in Fig. 1B). Interestingly, although Vpu did not affect the total levels of exogenous HA-Tetherin at low concentration, it still promoted efficient release of HIV-1 particle (Fig. 1A, compare lane 5 with lane 6; quantified in Fig. 1B). These results suggest that Vpu can reduce the total levels of Tetherin, yet this process does not appear to be absolutely necessary to promote HIV-1 particle release.
To further confirm these observations, we analyzed the turnover of exogenously-expressed native Tetherin in condition of efficient Vpu-mediated virus particle release by pulse-chase labeling analysis (Fig. 2A-C). HEK 293T cells were co-transfected with the proviral constructs HxBH10-vpu- or HxBH10-vpu+ and with a plasmid encoding native Tetherin. Forty-eight hours post-transfection, cells were pulse-labeled, chased for different intervals of time and analyzed for Tetherin and Vpu expression levels by sequential immunoprecipitation using specific antibodies (Abs). In parallel, transfected cells as well as virus-containing supernatants were collected prior to radio-labeling to monitor HIV-1 particle release by western blot. Tetherin-specific bands ranging from ∼20 kDa to ∼29 kDa and likely representing putative glycosylated forms of monomeric Tetherin were immunoprecipitated (Fig. 2A). Ectopic Tetherin turnover was not altered by Vpu since none of the Tetherin-specific bands showed any significant accelerated reduction over time in the presence of the viral protein (Fig. 2A; compare lanes 7–10 with lanes 3–6). Quantitative analysis of Tetherin turnover revealed that exogenous Tetherin has a half-life of approximately 3.5 h regardless of the presence of Vpu (Fig. 2B). Importantly, this lack of effect of Vpu on Tetherin turnover was observed in conditions of efficient Vpu-mediated HIV-1 particle release (Fig. 2C).
We further evaluated the half-life of endogenous Tetherin in infected HeLa cells in the presence or absence of Vpu. Vesicular stomatitis virus glycoprotein G (VSV-G)-pseudotyped HxBH10-vpu- or HxBH10-vpu+ virus-infected HeLa cells were pulse-labeled, chased for different intervals of time and analyzed for Tetherin expression levels as described above. In this system, endogenous mature Tetherin was detected as a ∼30–37 kDa smear (Fig. 2D). A lower ∼27 kDa band, distinct from the predicted Mr of 20 kDa for unglycosylated Tetherin, was also detected at time 0 and most probably corresponds to immature glycosylated forms of newly synthesized Tetherin still residing in the ER. Exogenously- and endogenously-expressed Tetherin were recently reported to display distinct mobilities (∼20–29 kDa (Fig. 2A) vs ∼27 and 30–37 kDa (Fig. 2D)) because they undergo different types of carbohydrate modifications [36]. Indeed, as demonstrated by Andrew and colleagues, we found that treatment of exogenous and endogenous Tetherin with Peptide: N-Glycosidase F (PNGase), an enzyme that cleaves all N-linked oligosaccharides, resulted in both cases in deglycosylated Tetherin proteins with a Mr of 19–20 kDa that were recognized by our anti-Tetherin serum (data not shown).
Figure 2D reveals that mature endogenous Tetherin has a half-life (t½) of approximatively 8h (Fig. 2D, lanes 1–5; quantified in Fig. 2E), which is indeed longer than exogenously-expressed Tetherin (t½: 3.5 h). Tetherin turnover was accelerated in presence of Vpu (t½: 3.5 h) (compare lanes 6–10 to lanes 1–5; quantified in Fig. 2E), consistent with recent results reported by Douglas and colleagues using HeLa cells transduced with Vpu-expressing adenoviral vectors [20]. Since Vpu-mediated Tetherin degradation was reported to rely on its capacity to recruit β-TrCP [12],[20],[34],[35], we evaluated the turnover of the restriction factor in presence of the β-TrCP-binding defective Vpu S52D,S56D mutant. This mutant harbors mutations at the key amino-acids required for interaction with β-TrCP (Ser52, and Ser56 for Asp) and displays a phenotype very similar to the well-characterized Vpu S52N,S56N mutant [3],[12],[20],[27],[33],[35],[37],[38]. Notably, Vpu S52D,S56D is unable to mediate CD4 degradation [25] and to recruit β-TrCP (Fig. S1A), but is however able to partially down-regulate Tetherin from the cell-surface (Fig. S1B) and to promote HIV-1 particle release, albeit to a different extent than WT Vpu (Fig. S1C-D). Interestingly, even though Vpu S52D,S56D was still capable of promoting HIV-1 particle release, its expression did not affect Tetherin turnover (Fig. 2D, compare lanes 11–15 and lanes 1–5; quantified in Fig. 2E), consistent with the reported role of β-TrCP in Vpu-mediated Tetherin degradation [12],[20],[34],[35]. Taken together, these results provide evidence that Vpu can promote HIV-1 particle release without a detectable reduction of Tetherin intracellular levels or a notable modification of its turnover, suggesting that reduction of total levels of Tetherin by a degradative process may not be necessary and/or sufficient to fully explain the anti-Tetherin activity of Vpu.
Rodent Tetherin is internalized from the plasma membrane and delivered back to the TGN through a clathrin-dependent pathway that requires the sequential action of AP2 and AP1 adaptor complexes [5]. Importantly, internalization was found to be dependent upon a dual tyrosine (Tyr)-based motif (YXXΦ, where Y corresponds to Tyr, the Xs are residues that are highly variable, and Φ corresponds to residues with bulky side chains) in the N-terminal cytoplasmic tail (amino acids at position 6 and 8) of the protein (Fig. 3A). One alternative mechanism to explain how Vpu down-regulates Tetherin cell-surface expression and counteracts its antiviral activity is by enhancing the rate of Tetherin endocytosis. To evaluate whether the natural pathway of Tetherin endocytosis was necessary for the anti-Tetherin activity of Vpu, we generated a mutant of Tetherin that harbored alanine substitutions at the two key Tyr residues within the dual Tyr-based internalization motif of the protein (HA-Tetherin Y6Y8) (Fig. 3A). Consistent with the previously reported role of this Tyr-based motif in rodent Tetherin endocytosis, substitution mutation of Tyr6 and Tyr8 prevented HA-Tetherin from being efficiently internalized from the cell-surface (Fig. 3B). To assess whether mutation of the Tyr-based motif affected Tetherin sensitivity to Vpu, HEK 293T cells were co-transfected with HxBH10-vpu- or HxBH10-vpu+ proviral constructs and plasmids encoding for HA-Tetherin or HA-Tetherin Y6Y8. Even though HA-Tetherin Y6Y8 was expressed at higher levels at the cell-surface (mean fluorescence intensity (MFI) = 240) as compared to HA-Tetherin wt (MFI = 90), both proteins were down-regulated from the cell-surface by Vpu and indeed appeared to reach similar cell-surface steady state levels (MFI of 51 and 41, respectively) (Fig. 3C). HIV-1 particle release was also monitored by western blot, 48h post-transfection. Consistent with its higher cell-surface expression levels, the restriction of virus particle release was more pronounced in presence of HA-Tetherin Y6Y8 than with HA-Tetherin wt (Fig. 3D; compare lanes 5 and 3 with lane 1; quantified in Fig. 3D). It is interesting to note that the mutant protein was overall expressed at higher levels than the WT protein (Fig. 3D, compare lanes 5 and 3). This is likely the result of the inefficient clearance of HA-Tetherin Y6Y8, which is not efficiently internalized from the cell-surface. Nevertheless, Vpu was still proficient at overcoming the restricting activity of HA-Tetherin Y6Y8 on HIV-1 particle release as demonstrated by the increased levels of virion-associated p24 released in the supernatant (Fig. 3D, compare lane 6 with lane 5; quantified in Fig. 3E). Similarly, a 2h treatment with 10 µM chlorpromazine, a drug that blocks clathrin-coated pit assembly at the plasma membrane [39], did not affect the ability of Vpu to overcome Tetherin-mediated restriction of HIV-1 particle release in HeLa cells (data not shown). Altogether, these results suggest that Vpu does not manipulate clathrin-mediated endocytosis, the natural pathway of Tetherin endocytosis, as a mean to deplete the restriction factor from the cell-surface or to antagonize its antiviral activity.
Since Vpu could accelerate Tetherin endocytosis by a clathrin-independent process, we next asked whether Vpu affected Tetherin internalization kinetics from the cell-surface using an assay that measures the contribution of all endocytosis pathways. To this end, we compared the rate of endocytosis of endogenous Tetherin in HeLa cells that were producing Vpu-positive HIV-1 (HxBH10-vpu+) with those producing a Vpu-defective virus (HxBH10-vpu-) (Fig. 4). Consistent with previous studies of rodent Tetherin [5], human Tetherin was internalized constitutively. Interestingly, even though Vpu down-regulated Tetherin expression on the cell-surface (data not shown), the internalization kinetics of cell-surface Tetherin was unaffected (Fig. 4). These results indicate that Vpu does not counteract Tetherin restriction by promoting Tetherin endocytosis.
We have recently reported that regulation of HIV-1 release correlates with co-localization of Vpu and Tetherin in the TGN, thus raising the possibility that Vpu could act intracellularly by affecting Tetherin trafficking [40]. A cell-surface Tetherin re-expression assay was developed to determine whether Vpu affects Tetherin trafficking to the cell-surface. Conceptually, this assay implies the loss of Tetherin epitopes at the cell-surface and their subsequent recovery over time as demonstrated previously for analysis of the Mtv-1 Superantigen protein trafficking [41]. HeLa cells were co-transfected with the HxBH10-vpu- or HxBH10-vpu+ proviral constructs as well as with a GFP-encoding plasmid to allow gating of transfected cells. Forty-eight hours post-transfection, cells were harvested and treated with pronase (0.05%) to proteolytically remove cell-surface protein epitopes. After quenching the proteolytic reaction, stripped cells were incubated at 37°C for different time intervals to allow protein intracellular trafficking and re-expression at the cell-surface and then stained at 4°C with anti-Tetherin Abs. Expression of Tetherin in transfected (GFP-positive) or untransfected (GFP-negative) subpopulations was analyzed by flow cytometry (Fig. 5). As expected, dot plots revealed that Tetherin levels were down-regulated by Vpu in the untreated cells (compare the GFP-positive/HxBH10-vpu+ subpopulation MFI (MFI = 36.8) with those of the GFP-negative/HxBH10-vpu+ (MFI = 56.7) or GFP-positive/HxBH10-vpu- (MFI = 61.0) subpopulations) (Fig. 5A, untreated). Pronase treatment markedly reduced the levels of Tetherin at the cell-surface, indicating that Tetherin epitopes were efficiently removed from the cell-surface. Interestingly, the levels of Tetherin detected at time 0 in the GFP-positive/HxBH10-vpu+ subpopulation was still lower (MFI = 6.7) than those detected in the GFP-negative/HxBH10-vpu+ (MFI = 9.7) or GFP-positive/HxBH10-vpu- (MFI = 12.7) subpopulations (Fig. 5A, time 0 min). Similar levels of Tetherin re-expression was detected at the cell-surface of GFP-positive/HxBH10-vpu- cells and GFP-negative cells after 180 min of incubation as demonstrated by the comparable MFI detected in the two subpopulations. In contrast, Vpu caused a substantial reduction in Tetherin re-expression after 180 min, as shown by the lower MFI value in the GFP-positive/HxBH10-vpu+ population (MFI = 19) compared to GFP-negative/HxBH10-vpu+ cells (MFI = 38.1) (Fig. 5A; 180 min). Treatment of cells with 10 µM Brefeldin A (BFA), a fungal metabolite that blocks protein sorting from the ER to the Golgi, prevented efficient Tetherin re-expression at the cell-surface both in transfected (GFP-positive) and untransfected (GFP-negative) cells, demonstrating the specificity of the re-expression assay (Fig. 5A, BFA). It is interesting to note that the absolute difference (Δ) in MFI detected between the Vpu-expressing cells and control cells (GFP-positive or negative/HxBH10-vpu- and GFP-negative/HxBH10-vpu+) was amplified after 180 min (Δ = ∼19) as compared to time 0 (Δ = ∼3–6), suggesting an effect of Vpu on cell-surface Tetherin re-expression kinetics. Indeed, the kinetics of Tetherin re-expression at the cell-surface, as measured by evaluating Tetherin levels (MFI) at the surface of pronase-treated cells relative to the corresponding untreated GFP-negative control over 180 min, increased linearly in the GFP-positive/HxBH10-vpu- and was indistinguishable from that of the GFP-negative/HxBH10-vpu+ control (slope of 0.27–0.29; Fig. 5B). After 180 min, approximately 75% of cell-surface Tetherin was recovered in both cases. In contrast, in presence of Vpu only ∼35% of Tetherin expression was recovered at the cell-surface. Indeed, the kinetics of Tetherin re-expression in these Vpu-expressing cells was much slower than the controls (slope of ∼0.14; Fig. 5B). Interestingly, the kinetics of Tetherin re-expression was similarly delayed in presence of the Vpu S52D,S56D mutant (slope of ∼0.11), indicating that the observed effect of Vpu on Tetherin re-expression at the cell-surface is not the consequence of Tetherin degradation. Similar analysis based on the proportion of Tetherin-positive cells as a read-out (cut-off was arbitrarily set on the pronase-treated HxBH10-vpu+/GFP+ time 0 sample) revealed analogous results (data not shown). Thus, our results suggest that Vpu expression interferes with Tetherin trafficking along the secretory and/or recycling pathways.
To further support these observations and identify the intracellular compartment where Tetherin might accumulate in presence of Vpu, we analyzed the intracellular distribution of endogenous Tetherin in VSV-G-pseudotyped HxBH10-vpu- and HxBH10-vpu+ virus-infected HeLa cells by immunostaining and confocal microscopy. To this end, we developed a staining protocol, which allowed simultaneous detection of Tetherin at the cell-surface and in intracellular compartments as described in the Materials and Methods. In the absence of Vpu, Tetherin was detected primarily at the plasma membrane but also to a lower extent on internal membranes that overlapped partially with TGN46, a cellular marker of the TGN (Fig. 6, Vpu- panels), consistent with previous intracellular localization studies of rodent Tetherin [4],[13]. This localization pattern was drastically altered by the presence of Vpu, which caused an effective removal of Tetherin from the cell-surface without, however, significantly affecting the pool of proteins localized in the perinuclear compartment that co-stained with TGN46 (Fig. 6A, Vpu+ panels and Fig. 6B). Notably, as reported previously by our laboratory [40], Vpu and Tetherin co-localized extensively in the TGN. This altered localization pattern of Tetherin was not observed in neighbouring untransfected cells, which indeed displayed a strong Tetherin staining at the plasma membrane. Since we found that Vpu was accelerating Tetherin turnover in HeLa cells in a β-TrCP-dependent manner (Fig. 2D), we next analyzed the distribution of Tetherin in presence of the Vpu S52D,S56D mutant. In contrast to infected HeLa cells expressing WT Vpu, residual Tetherin was still readily detected at the plasma membrane of infected cells expressing the Vpu S52D,S56D mutant (Fig. 6A, Vpu S52D,S56D panels), a finding that most probably reflects the fact that this mutant is less efficient at down-regulating Tetherin from the cell-surface than WT Vpu (Fig. S1B). Interestingly, expression of this mutant resulted in a re-localization of the cellular pool of Tetherin in the TGN (Fig. 6A). Importantly, Vpu S52D,S56D caused an ∼4-fold increase of the absolute staining signal of Tetherin in the TGN relative to WT Vpu or the Vpu- control (Fig. 6B), suggesting that in absence of degradation, Vpu traps Tetherin in the TGN. Taken together, these microscopy studies suggest that HIV-1 Vpu promotes the sequestration of endogenous Tetherin in the TGN, most probably before triggering β-TrCP-dependent Tetherin degradation, thus preventing Tetherin's trafficking to the plasma membrane. Since these localization studies were performed in HIV-1-infected cells, these results further indicate that Tetherin sequestration occurs at physiological levels of Vpu expression.
Having shown that Vpu expression affects the intracellular trafficking of Tetherin to the cell-surface and promotes a sequestration of Tetherin in the TGN, we next assessed whether Vpu can associate with Tetherin. HEK 293T cells were co-transfected with the HxBH10-vpu- or HxBH10-vpu+ proviral constructs as well as with a plasmid encoding HA-tagged Tetherin. Forty-eight hours post-transfection, cells were lysed in stringent detergent conditions with RIPA-DOC lysis buffer to avoid unspecific association resulting from membrane bridging. Tetherin was then immunoprecipitated from cell lysates with anti-Tetherin antibodies and the immunocomplexes were analyzed for the presence of Vpu by western blot. Immunoprecipitation of HA-Tetherin led to a selective pull-down of the Vpu protein, suggesting that Vpu can directly or indirectly interact with Tetherin in cells where it can antagonize Tetherin antiviral activity (Fig. 7B, lane 8). Early studies aimed at mapping the regions of Vpu necessary for enhancing HIV-1 release identified the TM domain of the protein as an important functional determinant since Vpu mutants that contained a randomized TM region (Vpu RD) or harbored point mutations within the TM spanning domain, such as Vpu KSL, failed to enhance virus particle release, yet were still stable, properly localized in a perinuclear compartment and able to induce efficient CD4 degradation [42],[43]. Vpu KSL contains a three amino-acid substitution in which Ile6, Ile8, and Val9 were replaced by Lys, Ser, and Leu, respectively, and contains a positively charged amino-acid in an area that is devoid of charge. To evaluate whether Vpu interacted with Tetherin through the TM anchor domain, we performed similar co-immunoprecipitations in HEK293T cells co-transfected with a plasmid encoding HA-tagged Tetherin and proviral constructs encoding Vpu RD (HxBH10-vpu RD) or Vpu KSL (HxBH10-vpu KSL) (Fig. 7A). In contrast to WT Vpu, both the Vpu RD and Vpu KSL mutants failed to co-immunoprecipitate efficiently with HA-Tetherin (Fig. 7B, compare lanes 9 and 10 with lane 8), suggesting that the association between the two proteins involves the TM domain of Vpu.
Since the TM of Vpu represents an important determinant for the association with Tetherin, we next determined whether the TM domain of Tetherin was also involved. Tetherin TM was recently proposed to contain the determinants responsible for the species-specific sensitivity to Vpu [29],[31],[32]. Notably, reciprocal exchange of TM domains between human (h) and rhesus monkey Tetherin (THN) proteins conferred sensitivity and resistance to Vpu and alterations in the human Tetherin TM domain that correspond to differences found in rhesus and agm Tetherin proteins were sufficient to render human Tetherin completely resistant to HIV-1 Vpu [29],[31],[32]. We constructed a set of expression plasmids encoding HA-tagged Tetherin chimeras with reciprocal exchanges of the TM between the human and agm proteins (Fig. S2A). In addition, we generated an expression plasmid encoding a HA-tagged human Tetherin that harbors double mutations in the TM domain (HA-human Tetherin ΔGI,T45I) comprising a deletion of Gly25 and Ile26 residues and a substitution of Thr45 for an Ile (Fig. S2A). This mutant was previously reported to strongly inhibit HIV-1 particle release but was completely resistant to antagonism by Vpu [29]. HEK 293T cells were co-transfected with the HxBH10-vpu- or HxBH10-vpu+ proviral constructs as well as with the indicated HA-Tetherin-encoding plasmids (Fig. 7C-E). Forty-eight hours post-transfection, Tetherin was immunoprecipitated from cell lysates and immunocomplexes were further analyzed for the presence of Vpu by western blot. Since HA-agm Tetherin and the HA-agmTHN(hTM) chimeras were less expressed and/or detected using our anti-human Tetherin Abs (these chimeras were not efficiently precipitated with anti-HA Abs), Vpu association was compared between Tetherin variants displaying similar expression profiles. As shown in Figure 7C, immunoprecipitation of HA-human Tetherin co-precipitated Vpu (Fig. 7C, lane 12). As expected, HA-agm Tetherin, which was reported to be resistant to Vpu antagonism [29],[32], did not show a strong and specific association with Vpu. In contrast, replacement of the agm Tetherin TM domain with that of human Tetherin (HA-agmTHN(hTM)) restored the association with Vpu (Fig. 7C, compare lane 14 with lanes 16 and 12). Exchange of the human Tetherin TM with that of agm Tetherin (HA-hTHN(agmTM)) led to a marked reduction of Vpu association despite comparable levels of Tetherin expression (Fig. 7D, compare lane 12 with lane 10). Finally, introduction of the ΔGI,T45I mutations in the TM domain of HA-human Tetherin drastically reduced the association with Vpu (Fig. 7E, compare lane 12 with lane 10). These results suggest that the integrity of the TM domain of human Tetherin is necessary for the association with Vpu. Taken together, our data suggests that association of Vpu and human Tetherin involves their TM anchor domains.
We next assessed whether association of Vpu to Tetherin was required to counteract Tetherin antiviral activity (Fig. 8). To do so, we transfected a proviral construct encoding WT Vpu or the Vpu KSL mutant in HeLa cells and first assessed their ability to associate with endogenous Tetherin. Wild type Vpu co-precipitated with endogenous Tetherin while Vpu KSL did not, confirming the results obtained when exogenous HA-tagged Tetherin was overexpressed in HEK293T cells expressing HxBH10-vpu+ or HxBH10-vpu KSL proviruses (Fig. 8A, compare lane 6 with lane 5). Association of Vpu with endogenous Tetherin was specific since HIV-1 envelope glycoprotein precursor gp160, another type 1 integral protein, did not co-precipitate with Tetherin (Fig. 8A). Interestingly, the Vpu KSL mutant was strongly attenuated in its ability to down-regulate the steady-state levels of Tetherin at the cell-surface relative to WT Vpu since Tetherin cell-surface expression in presence of Vpu KSL was significantly higher than in presence of WT Vpu (MFI of 191 vs 138), yet still slightly lower than in the Vpu-negative control (MFI = 235) (Fig. 8B). Importantly, as previously reported [43], this mutant was drastically attenuated in its ability to promote efficient HIV-1 particle release (Fig. 8C, compare lanes 3 and 2 with lane 1; quantified in Fig. 8D). Similarly to Vpu KSL, the Vpu RD mutant, which is also defective for Tetherin binding, was also markedly attenuated in its ability to down-regulate Tetherin cell-surface expression and to promote efficient HIV-1 release in HEK 293T cells ectopically-expressing Tetherin (Fig. S3A and B), consistent with results reported by previous studies [3],[42]. Human Tetherin containing the agm TM domain (HA-hTHN(agmTM)) or the ΔGI,T45I mutations (HA-human Tetherin ΔGI,T45I) lost the ability to bind Vpu (Fig. 7D and E) and as expected still restricted HIV-1 particle release even in the presence of Vpu (Fig. S2B and C for HA-hTHN(agmTM) and Fig. S2D and E for HA-human Tetherin ΔGI,T45I). Unexpectedly, introduction of the human Tetherin TM domain in agm Tetherin (HA-agmTHN(hTM), did not reinstate a significant sensitivity to Vpu (Fig. S2 B and C), despite a detectable restoration of the Vpu binding (Fig. 7C). This functional phenotype (absence of Vpu sensitivity) is different from that obtained by McNatt and colleagues [29] using agm or rhesus Tetherins containing the human Tetherin TM domain but is similar to that reported by Goffinet and colleagues [30] using rodent Tetherin proteins containing the human Tetherin TM domain. Difference in the configuration of the TM domain in these Tetherin chimeric constructs may explain this discrepancy. The result obtained with the HA-agmTHN(hTM) chimeric construct suggests that association of Tetherin with Vpu is necessary but not sufficient to overcome Tetherin-mediated restriction of HIV-1 particle release. However, we cannot rule-out the possibility that the binding of Vpu to the HA-agmTHN(hTM) chimeric protein may indeed be less efficient than with HA-human Tetherin since our antibody may underestimate the levels of expressed HA-agmTHN(hTM) proteins (Fig. 7C, compare lane 14 and lane 12). Thus, alternatively, a threshold level of Vpu association to Tetherin may be necessary to antagonize the antiviral activity of the restriction factor. Nonetheless, overall, these results suggest that association of Vpu to Tetherin is required to antagonize the antiviral function of the restriction factor.
Since we showed that expression of Vpu could cause a re-localization of Tetherin from the plasma membrane to the TGN, we next tested whether the Vpu TM domain mutants, Vpu RD and Vpu KSL, which are unable to associate with Tetherin, could still sequester endogenous Tetherin in the TGN. Hela cells were infected with VSV-G-pseudotyped viruses expressing WT Vpu or Vpu RD or Vpu KSL and infected cells were immunostained and analyzed by confocal microscopy. Figure 9 reveals that both the Vpu RD and Vpu KSL lost the ability to efficiently remove Tetherin from the cell-surface and to relocate the restriction factor to the TGN. Indeed, quantitative analysis of the Tetherin signal localized in the TGN relative to the total cellular Tetherin signal showed that cells expressing the Tetherin degradation-defective Vpu S52D,S56D mutant displayed an increased proportion of Tetherin signal in the TGN compared to the Vpu- control. (∼57% relative to ∼18%; p<0.001) (Fig. 6). In contrast, cells expressing the Vpu RD or Vpu KSL mutants showed ∼25% of the Tetherin signal in the TGN, a proportion that was in fact just slightly higher than that found in cells infected with a Vpu-defective virus (Fig. 9). Accordingly, in contrast to Vpu S52D,S56D, these mutants did not show any increase in the absolute levels of Tetherin signal in the TGN (Fig. 9). These results indicate that Vpu promotes the sequestration of Tetherin in the TGN by a process that is dependent on the association of the two proteins. Taken together with the functional analysis and the co-precipitation experiments, these results further suggest that the antagonism of Tetherin function by Vpu involves binding and sequestration of the restriction factor in the TGN.
The Vpu accessory protein stimulates the release of HIV-1 virions by antagonizing a restriction on virus particle release mediated by Tetherin at the cell-surface [2],[3]. This antagonism appears to closely correlate with the ability of Vpu to mediate down-regulation of Tetherin expression from the cell-surface [3],[12],[20]. While this down-regulation is accompanied by enhanced degradation of Tetherin in several infected cell types, such as Jurkat [20] and CEM-G37 T cells [21] or macrophages [33], several lines of evidence also suggest that degradation of Tetherin per se cannot entirely account for Vpu-mediated counteraction of the restriction factor since: 1) Vpu decreased total cellular Tetherin to a lesser extent than cell-surface Tetherin in HeLa cells [12]; 2) Vpu expression did not result in a reduction of intracellular Tetherin in infected CEMx174 and H9 cells, yet virus replication in these cells was Vpu-responsive [33]; 3) Vpu mutants that contained substitution mutation in the DSGΦXS β-TrCP recognition motif that rendered them deficient for directing β-TrCP-dependent degradation of Tetherin were still able to partially [3],[12],[20],[34],[37] (Fig. S1C and D) or in some instances totally [33],[44] overcome the particle release restriction; 4) binding of Vpu to Tetherin was recently shown to be sufficient for a partial relief of the restriction [20],[34].
We show here that Vpu can promote efficient HIV-1 particle release without a detectable reduction of the total steady-state levels of Tetherin (Fig. 1) nor a notable modification of the restriction factor turnover rate in transfected HEK 293T cells (Fig. 2A -C), suggesting that degradation of the antiviral factor per se is not necessary and/or sufficient to account for Tetherin antagonism at least in this experimental system. Interestingly, Vpu expression in HeLa cells increased endogenous Tetherin turnover (Fig. 2D and E), as reported previously by Douglas and colleagues [20]. This Vpu-mediated degradation process was however still relatively slow (half-life of Tetherin decreases from ∼8 h to 3.5 h in presence of Vpu) as compared to the very efficient CD4 receptor degradation induced by Vpu (half-life of CD4 decreases from ∼6 h to ∼12 min in presence of Vpu) [45], and as such is unlikely to explain the powerful antagonism of Tetherin by Vpu. Thus, Vpu-mediated counteraction of Tetherin restriction must involve other mechanisms. Our results further indicate that Vpu does not promote endocytosis of Tetherin as a mechanism to antagonize the restriction factor. Indeed, mutation of the two critical tyrosine residues located within a dual tyrosine-based sorting motif in the cytoplasmic domain of the protein, did not abolish the sensitivity of Tetherin to Vpu, as was indeed recently reported by Iwabu and colleagues [35], yet led to an increased accumulation of the restriction factor at the cell-surface and to a more potent restriction of HIV-1 particle release (Fig. 3). Moreover, although analysis of Tetherin endocytosis kinetics showed that the protein is constitutively internalized, it did not reveal any increase in the rate of Tetherin endocytosis in presence of Vpu (Fig. 4), and as such confirmed the results recently reported by Mitchell and colleagues [12]. These results are consistent with previous findings showing that Vpu-mediated enhancement of virus particle release is not significantly affected by expression of dominant negative (DN) mutants of Dynamin or Rab5 nor by depletion of clathrin heavy chain or by treatment with inhibitors of endocytosis such as chlorpromazine (data not shown) [6],[40],[46]. Collectively, they indicate that Vpu does not antagonize Tetherin by enhancing its endocytosis from the cell-surface. Instead, these findings suggest that Vpu affects Tetherin anterograde trafficking events whose net effect is depletion of the restriction factor from the cell-surface. In support of this hypothesis, a previous report suggested that Vpu may acts as a regulator of protein transport along the secretory pathway [47]. In fact, upon analyzing the kinetics of Tetherin expression at the cell-surface of protease-treated HeLa cells, we noticed that Tetherin re-expression was significantly reduced in presence of Vpu (Fig. 5). It is important to note, here, that this effect is occurring regardless of whether Vpu is inducing Tetherin degradation in these experimental conditions since: 1) it is the rate of re-expression, as defined by the slope of the graph of Figure 5B that is affected in presence of Vpu; 2) re-expression of cell-surface Tetherin was also delayed in presence of the Tetherin degradation-defective Vpu S52D,S56D mutant (Fig. 5B). Consistent with this observation, examination of infected HeLa cells expressing Vpu revealed that the cellular pool of Tetherin was re-localized from the cell-surface to a perinuclear compartment that stained positive with the TGN marker TGN46 and Vpu itself (Fig. 6). Although some of the loss of cell-surface Tetherin in Vpu-expressing HeLa cells could be attributed to Vpu-mediated degradation of the restriction factor, similar experiments performed with the Vpu S52D,S56D mutant demonstrated that Vpu can cause a redistribution of Tetherin from the plasma membrane to the TGN. These results suggest that reduction of Tetherin levels at the plasma membrane involves a step whereby Vpu sequesters Tetherin in the TGN. They also provide a possible mechanism to explain the partial relief of Tetherin restriction observed by several groups upon expression of β-TrCP-binding defective Vpu mutants [12],[20],[34]. However, since residual Tetherin was still readily detected at the plasma membrane of HeLa cells expressing the Vpu S52D,S56D mutant (Fig. 6A and Fig. S1B), this sequestration may not be sufficient to effectively prevent Tetherin to reach the cell-surface at least in this model. These observations underline, perhaps, the need for β-TrCP-mediated trafficking and degradation processes as a complementary mechanism to efficiently remove Tetherin from the cell-surface, particularly in certain cell types or under specific conditions, such as IFN exposure, where cellular Tetherin expression levels are high. Thus, the different particle release phenotypes observed with Vpu mutants that are unable to recruit β-TrCP, such as Vpu S52D,S56D, (these range from as efficient than WT Vpu to a ∼50% attenuation) [3],[12],[20],[33],[34],[37],[44] may in fact relate to differing levels of Tetherin expression in the target cells. Although, it is conceivable that Vpu-mediated sequestration of Tetherin in the TGN may explain how Vpu antagonizes Tetherin in the absence of a decrease in total Tetherin expression, it still remains unclear how Vpu could antagonize Tetherin in the absence of down-regulation from the cell-surface and a decrease in total cellular expression in certain cell types [33].
We also confirmed by co-immunoprecipitation studies that Vpu interacts with Tetherin [20],[32],[34],[35] and demonstrated using well-characterized Vpu and Tetherin mutants as well as chimeric proteins between human and African green monkey Tetherin molecules that this physical interaction involves the transmembrane domains of the two proteins (Fig. 7 and Fig. 8). These findings are consistent with previous results showing that the transmembrane domain of Vpu is required for both the down-regulation of surface Tetherin and the enhancement of HIV-1 particle release [3],[42],[43]. They are also supported by recent data indicating that the inability of Vpu to antagonize the restrictive effect of African green monkey and rhesus Tetherin proteins is a consequence of amino-acid changes in the transmembrane domain of the rhesus and African green monkey protein relative to the human form [29],[32]. Importantly, our data establish a direct functional link between association of Vpu to Tetherin and Tetherin antagonism since we showed that Vpu's ability to interact with Tetherin was necessary: 1) to counteract the restriction on HIV-1 particle release and down-regulate Tetherin from the cell-surface (Fig. 8, Fig. S2 and S3); and, 2) to induce a re-localization of the cellular pool of Tetherin from the plasma membrane to the TGN (Fig. 9). Unexpectedly, our observation showing that introduction of the human Tetherin TM domain in African green monkey Tetherin did not reinstate a significant sensitivity to Vpu (Fig. S2B and C) despite a detectable restoration of the Vpu binding raises the possibility that this interaction may not be sufficient to explain Tetherin antagonism. The requirement for additional cellular factor(s) in the Vpu-mediated Tetherin antagonism is therefore a possibility that warrants further investigations.
Since our co-immunoprecipitation and functional data indicate that a physical interaction between Vpu and Tetherin is required for Tetherin antagonism and sequestration, this raise the possibility that Vpu may simply interact with Tetherin and inhibit its outward trafficking from the TGN since all membrane proteins are transported to the plasma membrane through the TGN (Fig. 10). Since mutation of the dual Tyr signal in the Tetherin cytoplasmic tail does not abolish the sensitivity to Vpu, it appears that Tetherin sequestration in the TGN may occur before its endocytosis from the cell-surface and as such may involve newly synthesized Tetherin en route to the plasma membrane. However, we cannot completely rule-out that Vpu could interact with endocytosed Tetherin in the TGN and prevent its recycling back to the cell-surface, given that previous data from the Spearman group demonstrated a requirement for the recycling endosomes in Vpu function [48]. Furthermore, recent studies reported that AP-2 depletion [12] or over-expression of DN mutant of Dynamin (Dyn2-K44A) [35] could partially interfere with Vpu-mediated down-regulation of Tetherin expression from the cell-surface. Indeed, since Vpu is expressed from a Rev-dependent bicistronic mRNA encoding Env and consequently is made late during the virus life cycle [49], the direct removal of Tetherin from the plasma membrane via endosomal trafficking may be critical to ensure a rapid and efficient neutralization of the restriction on HIV-1 release. It is nevertheless surprising that mutation of the Tetherin dual Tyrosine-based endocytosis motif did not affect the ability of Vpu to counteract Tetherin-mediated restriction of HIV-1 particle release. Perhaps co-expressing transiently the two proteins simultaneously may not adequately reflect physiological conditions given that Tetherin is normally already present at the plasma membrane when Vpu is expressed. Alternatively, it is also possible that sequestration of newly synthesized Tetherin in the TGN may rapidly clear the restriction factor from its site of virion-tethering action at the plasma membrane depending on the physiologic rate of Tetherin turnover at the plasma membrane. More studies will be required to assess whether Vpu affects Tetherin trafficking at a pre- or/and post-endocytic step.
Vpu-mediated sequestration of Tetherin in the TGN could be complemented by β-TrCP-dependent degradation processes, thus enhancing Tetherin antagonism. It is conceivable that Vpu by forming a complex with β-TrCP and Tetherin in the membrane of the TGN could potentially induce ubiquitination via interaction with the SCFβ-TrCP E3 ubiquitin ligase. This ubiquitination event could either enhance the sequestration by preventing Tetherin recycling to the cell-surface or/and targets it to lysosomes for degradation (Fig. 10). Our data indicating that Vpu causes a re-localization of the cellular pool of Tetherin in the TGN, where indeed the two proteins strongly co-localize, suggests that Vpu affects Tetherin trafficking and potentially its degradation from a post-ER compartment(s), and as such is difficult to reconcile with a mechanism of proteasomal degradation of the antiviral factor through the cellular ER-associated degradation (ERAD) pathway [30],[31],[34]. However, given the fact that Vpu induces degradation of the CD4 receptor in the ER by an ERAD-like mechanism [25], it is still possible that some proteasomal degradation in the ER may contribute to Tetherin depletion. Whether Vpu and Tetherin trafficking cross each other in the TGN, thus permitting a physical interaction between the two proteins, remains to be determined. In this regard, it is interesting to note that recent data from our laboratory have shown that the ability of Vpu to suppress Tetherin-mediated restriction of HIV-1 particle release was linked to its localization in the TGN [40]. Further studies aimed at identifying the determinants regulating Vpu trafficking and localization in the TGN will likely shed light on this mechanism.
HIV-1 Vpu appears to share the ability to sequester Tetherin in the TGN with other Tetherin antagonists, namely HIV-2 Rod and SIVtan Env. Like HIV-1 Vpu, HIV-2 Rod Env was recently shown to interact with Tetherin and to cause a redistribution of Tetherin in the TGN, although in the case of HIV-2 Env, no evidence of Tetherin degradation was observed [21]. Similarly, SIVtan Env was also shown to downregulate cell-surface Tetherin by sequestering the restriction factor in intracellular compartment(s) [22]. Collectively, these findings suggest a common mechanism of antagonism that results in TGN trapping, which can be augmented by the induction of degradation in the case of Vpu. Given the powerful restrictive effects of human Tetherin on HIV production, this dual mechanism of antagonism mediated by Vpu may have provided HIV-1 with stronger countermeasures to antagonize Tetherin. This genetic and functional divergence between HIV-1 and HIV-2 may perhaps account for the different virulence properties displayed by these two closely related viruses [21],[50]. Future studies on the role of Tetherin antagonism in the pathogenesis of primate immunodeficiency virus will likely shed light on the contribution of this innate antiviral factor in the control of viral infection and spread in vivo and will reveal whether enhancing the antiviral activity of Tetherin is a valid option to thwart HIV-1 replication.
Polyclonal rabbit antibodies against human Tetherin were produced according to animal experimentation research protocols approved by the Animal Care Committee of the Institut de recherches cliniques de Montreal and in accordance with Canadian Council on Animal Care (CCAC) guidelines and policies.
Anti-Vpu rabbit polyclonal serum was described previously [40]. Anti-Tetherin rabbit polyclonal serum was generated by immunization of rabbits with a bacterially-produced Glutathione-S-transferase (GST) fusion protein containing a polypeptide corresponding to amino-acids 40 to 180 of human Tetherin. Rabbit pre-immune serum was collected prior to rabbit immunization. Monoclonal anti-HA (clone 12CA5), anti-p24 (catalog no. HB9725) and anti-myc (clone 9E10) Abs were isolated from the supernatants of cultured hybridoma cells obtained from the American Type Culture Collection (ATCC). The monoclonal anti-gp120 antibody [51],[52] was obtained from NIH AIDS research and Reference Reagent Program. All secondary Alexa-conjugated IgG Abs were obtained from Invitrogen. Sheep anti-TGN46 (Serotec), mouse anti-BST2 (Abnova), anti-actin Abs (Sigma), pronase (Calbiochem), PNGase (New England Biolabs), BFA, paraformaldehyde (PFA) and chlorpromazine (Sigma) were all obtained from commercial sources. All reagents were stored according to the manufacturer's instructions.
HEK 293T, HeLa and Cos-7 cells were obtained from ATCC. All cells were maintained as described previously [40]. HEK 293T and HeLa cells were transfected using the calcium-phosphate method and lipofectamine 2000 (Invitrogen), respectively. Functional and biochemical analyses were performed 48h post-transfection. Empty plasmid DNA was added to each transfection to keep the amount of transfected DNA constant.
HxBH10-vpu+ and HxBH10-vpu- are two infectious molecular clones of HIV-1 that are isogenic except for the expression of Vpu [53]. HxBH10-vpu S52D,S56D, SVCMV-vpu- and SVCMV-vpu+ were previously described, [25]. HxBH10-vpu KSL, HxB10-vpu RD and SVCMV-vpu S52D,S56D were generated by PCR-based site-directed mutagenesis [40]. The pcDNA/Myc-His-β-TrCP plasmid was kindly provided by Dr. Richard Benarous [27]. To generate pcDNA-Tetherin and pCMV-HA-Tetherin, the human Tetherin open reading frame was amplified by PCR from pCMV-SPORT6-hBST2 plasmid (Open Biosystems) and cloned into pcDNA3.1/Hygro (+) (Invitrogen) and pCMV-HA (Clontech). The agmTetherin open reading frame was amplified from Cos-7 cells mRNA using RT-PCR as described [29]. Tetherin chimeras were designed according to the structure prediction from Kupzig and Banting [4]. All Tetherin genes were inserted into pcDNA3.1/Hygro (+) using NheI and Asp718 restriction sites or into the pCMV-HA plasmid using BglII and Asp718 restriction sites to generate pcDNA-Tetherin and pCMV-HA-Tetherin constructs, respectively. pCMV-HA-Tetherin Y6Y8 and pCMV-HA-Tetherin ΔGI,T45I were generated by PCR-based site-directed mutagenesis. The vesicular stomatitis virus (VSV) glycoprotein G-expressing plasmid, pSVCMVin-VSV-G, was previously described [54]. The pQBI25 GFP-expressing plasmid was obtained from Qbiogene.
Transfected HeLa or HEK 293T cells were lysed in CHAPS lysis buffer (50 mM Tris, 5 mM EDTA, 100 mM NaCl, 0.5% CHAPS, pH 7.2) or in radio-immunoprecipitation assay (RIPA-DOC) buffer (10 mM Tris pH 7.2, 140 mM NaCl, 8 mM Na2HPO4, 2 mM NaH2PO4, 1% Nonidet-P40, 0.5% sodium dodecyl sulfate, 1.2 mM deoxycholate), respectively. Proteins from lysates were resolved on 12.5% SDS-PAGE, electro-blotted and analyzed by western blot as described previously [40].
Pulse-chase experiments were perfomed as described previously [40]. Briefly, transfected HEK 293T or HeLa cells infected with VSV-G-pseudotyped HxBH10-derived virus at a MOI of 1 were pulse-labeled for 30 min and 2 h respectively, with 800 µCi/ml of [35S]methionine and [35S]cysteine (Perkin Elmer) and chased for different interval of times. Following lysis of radio-labeled cells in RIPA-DOC, lysates were first pre-cleared with protein A sepharose beads coated with pre-immune rabbit serum for 1 h. Pre-cleared cell lysates were then incubated with anti-Tetherin Abs for 2 h at 4°C prior to immunoprecipitation using protein A sepharose beads. In HEK 293T cells, Vpu proteins were sequentially immunoprecipitated using the same method. Labeled proteins were analyzed by SDS-PAGE and autoradiography.
Cells were lysed in RIPA-DOC. Tetherin-containing lysates were incubated with anti-Tetherin Abs for 2 h at 4°C prior to addition of protein A sepharose beads. Following an incubation of 2 h at 4°C, beads were isolated by centrifugation, washed with denaturing buffer (New England Biolabs), resuspended in denaturing buffer and then boiled at 95°C for 10 min. The supplied reaction buffer was added along with NP-40 (0.1%) according to the manufacturer's suggestion. Samples were then digested with 1500 units of PNGase (New England Biolabs) at 37°C for 3 h. Control samples were incubated without the enzyme. Proteins were eluted from beads by boiling in an equal volume of sample buffer for 10 min and analyzed by immunoblotting.
The virus release assay was described previously [40]. Briefly, supernatants of transfected cells were clarified by centrifugation and filtered through a 45 µm filter. Viral particles were pelleted by ultracentrifugation onto a 20% sucrose cushion in PBS for 2 h at 130000 g at 4°C and lysed in RIPA-DOC. Gag products were analyzed by western blot. Viral release efficiency was evaluated by determining the ratio between the virion-associated Gag (p24) band signal and all intracellular Gag-related band signal using laser scanning densitometry.
HEK 293T cells were transfected with HxBH10 proviral constructs and pSVCMVin-VSV-G as described previously [54]. Supernatants of transfected cells were clarified, filtered and pelleted by ultracentrifugation as described above and resuspended in DMEM supplemented with 10% bovine serum (FBS). Viruses were titrated using a standard MAGI assay [54].
Cells were washed in PBS, resuspended at a concentration of 1×106cells/ml and stained with the specific anti-Tetherin serum for 45 min at 4°C. After incubation, cells were washed and stained using appropriate fluorochrome-coupled secondary Abs for 30 min at 4°C. Cells were then washed and fixed with 2% PFA. Transfected GFP-expressing cells were analyzed for cell-surface Tetherin expression by flow cytometry. Rabbit pre-immune serum served as a staining control. Fluorescence intensities were acquired using a FACScalibur flow cytometer (BD Biosciences) and data was analyzed using FlowJo software v. 7.25 (Treestar). MFI values presented in the histograms correspond to the specific signal obtained after substraction of the MFI value from the pre-immune control.
Cells were washed in PBS, re-suspended in PBS containing the anti-Tetherin serum at a concentration of 1×106cells/ml and incubated for 45 minutes at 4°C. Following washes in cold PBS, cells were incubated at 37°C in DMEM medium supplemented with 5% FBS for different time intervals. At each time point, cells were harvested, washed in cold PBS and stained with the appropriate fluorochrome-coupled secondary Abs for 30 min at 4°C. Transfected GFP-expressing cells were analyzed for cell-surface Tetherin expression by flow cytometry.
Cells were harvested, washed in PBS, re-suspended at a concentration of 1×106 cells/ml in PBS-pronase 0.05% and incubated for 30 minutes at 37°C. Cold DMEM containing 10% FBS was added to block surface protein proteolysis. Cells were then washed, incubated at 37°C for different time intervals and stained for cell-surface Tetherin as described above. Expression of Tetherin at the cell-surface of transfected GFP-positive and untransfected GFP-negative cells was analyzed by flow cytometry.
HeLa cells were infected with VSV-G-pseudotyped HxBH10-derived viruses at a MOI of 0.125. Forty-eight hours post-infection, cells were immunostained with anti-Tetherin Abs (Abnova) for 30 min at 4°C, washed in cold PBS, fixed with 4% PFA and permeabilized with 0.2% Triton X-100. Next, cells were incubated with anti-Tetherin (Abnova), anti-Vpu and anti-TGN46 Abs for 2 h at 37°C, washed and incubated with the appropriate secondary Abs for 30 min at room temperature. Analyses were performed with a LSM710 laser scanning confocal microscope (Zeiss). Quantitation of Tetherin signal was performed using the Zeiss LSM510 software. The absolute Tetherin signal in the TGN was determined by measuring the specific Tetherin signal in the region delineated by the TGN46 marker on digital pictures produced using similar acquisition time. Percentage of Tetherin accumulating in the TGN was calculated by evaluating Tetherin signal intensity in the TGN relative to the total Tetherin signal intensity detected in the cell as described previously [40]. Analysis was performed on at least 25 distinct cells.
For co-immnuoprecipitation of Vpu and Tetherin, transfected HEK 293T and HeLa cells were harvested 48 h post-transfection and lysed in RIPA-DOC or CHAPS buffer, respectively. Five percent of each lysates were preserved to control for protein expression. Cell lysates were first pre-cleared with protein A sepharose beads coated with pre-immune rabbit serum for 1 h at 4°C and then, incubated with anti-Tetherin Abs for 2 h at 4°C, prior to precipitation with protein A sepharose beads. Immunoprecipitates were analyzed for the presence of Vpu and Tetherin by western blot.
For co-immunoprecipitation of Vpu and β-TrCP, HEK 293T cells were transfected with SVCMV-vpu- or SVCMV-vpu+ or SVCMV-vpu S52D,S56D and pcDNA/Myc-His-β-TrCP. Transfected cells were then radiolabelled with 800 µCi/ml of [35S]methionine and [35S]cysteine (Perkin Elmer) and lysed in CHAPS buffer. Lysates were first pre-cleared with protein A sepharose beads coated with pre-immune rabbit serum for 1 h. Pre-cleared cell lysates were then incubated with anti-myc Abs for 2 h at 4°C prior precipitation using protein A sepharose beads. Vpu was then sequentially immunoprecipitated using anti-Vpu Abs using the same method. Labeled proteins were analyzed by SDS-PAGE and autoradiography.
Scans were performed on a Duoscan T1200 scanner (AGFA) followed by densitometric quantitation using the Image Quant 5.0 software (Molecular Dynamics). Statistical analysis was performed using a paired Student's t test, and statistical significance was considered at p<0.001.
NCBI reference number for HxBH10 Vpu and human Tetherin proteins are P69699 and AAH33873, respectively. Genebank accession number for agm Tetherin is FJ943430.
|
10.1371/journal.pcbi.1002004 | Conformational Sampling and Nucleotide-Dependent Transitions of the GroEL Subunit Probed by Unbiased Molecular Dynamics Simulations | GroEL is an ATP dependent molecular chaperone that promotes the folding of a large number of substrate proteins in E. coli. Large-scale conformational transitions occurring during the reaction cycle have been characterized from extensive crystallographic studies. However, the link between the observed conformations and the mechanisms involved in the allosteric response to ATP and the nucleotide-driven reaction cycle are not completely established. Here we describe extensive (in total long) unbiased molecular dynamics (MD) simulations that probe the response of GroEL subunits to ATP binding. We observe nucleotide dependent conformational transitions, and show with multiple 100 ns long simulations that the ligand-induced shift in the conformational populations are intrinsically coded in the structure-dynamics relationship of the protein subunit. Thus, these simulations reveal a stabilization of the equatorial domain upon nucleotide binding and a concomitant “opening” of the subunit, which reaches a conformation close to that observed in the crystal structure of the subunits within the ADP-bound oligomer. Moreover, we identify changes in a set of unique intrasubunit interactions potentially important for the conformational transition.
| Molecular machines convert chemical energy to mechanical work in the process of carrying out their specific tasks. Often these proteins are fueled by ATP binding and hydrolysis, enabling switching between different conformations. The ATP-dependent chaperone GroEL is a molecular machine that opens and closes its barrel-like structure in order to provide a folding cage for unfolded proteins. The quest to fully understand and control GroEL and other molecular machines is enhanced by complementing experimental work with computational approaches. Here, we provide a description of the molecular basis for the conformational changes in the GroEL subunit by performing extensive molecular dynamics simulations. The simulations sample the conformational population for the different nucleotide-free and bound states in the isolated subunit. The results reveal that the conformations of the subunit when isolated resemble those of the subunit integrated in the GroEL complex. Moreover, the molecular dynamics simulations allow following detailed changes in individual interatomic interactions brought about by ATP-binding.
| GroEL participates in the folding of 5-10% of cellular Escherichia coli proteins by providing an isolated chamber for non-native substrate proteins together with a heptameric ring shaped co-chaperonin, denoted GroES [1]–[5]. It has also been suggested that GroEL might forcefully unfold kinetically trapped misfolded intermediates [6], [7]. GroEL is composed of two heptameric rings made up of identical subunits, each of 57 kDa. The two separate rings, denoted cis (active) and trans (inactive), are stacked back-to-back to form two folding environments (or cages) working off-phase, analogous to a two-stroke engine. Each subunit is divided into three domains; equatorial (residues 1-133, 409-548), intermediate (134-190, 377-408), and apical (191-376) domain, separated by two hinges that facilitate large conformational transitions in the complex [8]–[10].
Substantial mechanistic insight has been obtained through comparison of the large amount of structural data available for the GroEL complex. The first high-resolution X-ray structure was released in 1994 by Braig et al. [8], and since then, numerous structural studies, including X-ray crystallography, cryo-EM, and NMR have been published of different functional states of GroEL (for reviews see [11], [12]). On this structural background it has been possible to make predictions and educated hypothesis about the transition pathways during the protein functional cycle. This includes the ATP dependent opening of the cis cavity with the concomitant increase of its volume from to [10]. The conformational transitions occurring on the subunit level are substantial, and its trajectory is generally explained in a sequential manner.
Binding of ATP to each of the seven equatorial domains in the cis ring, together with a non-native polypeptide produces a counterclockwise twist of the apical domain, and a downward rotational movement of the intermediate domain [13], [14]. This structural state of the ring is denoted R (where each of the 7 subunits are in the r state) while the initial closed state is denoted T (each of the 7 subunits are in the t state). Reaching the R state facilitates GroES association to the apical domains of the cis ring promoting much larger conformational changes and resulting in the fully open R′ conformation (all seven subunits in the r′ state). This r′ conformer is characterized by a elevation and clockwise twist of the apical domains (opposite direction to that seen upon ATP binding) [10]. Non-native polypeptide folding takes place within the cis ring of the R′ state of the reaction cycle, which is the longest lived (about 8–10 s) [15], and continues until ATP hydrolysis induce the R″ conformation permiting ATP binding to the opposite trans ring [16], [17]. This final rearrangement result in a conformer very similar to the r′ form (RMSd of 1.46 Å).
Despite this extensive structural insight, the mechanisms involved in allosteric signaling are not yet fully understood at atomic and residue level. In general, X-ray crystallography provides invaluable snapshots of different states of the protein reaction cycle, but not of the transitions between them. In this context, computational approaches have the potential to nicely complement the experimental techniques. In particular, molecular dynamics (MD) simulations provide important insight into protein dynamics at the atomic level, and allow following subsequent individual atomic interactions and fluctuations as a function of time [18]. Additionally, normal mode analysis (NMA) has proven to be efficient and accurate in the task of predicting and describing large scale conformational transitions in proteins [19]–[22]. NMA analytically characterizes all possible deformations of a protein around a stable equilibria with respect to their energetic cost. Although the utilization of NMA involves a loss of time-dependent fluctuations and resolution, i.e. by the use of coarse-grained models, it has been shown that it provides functionally relevant motions, and information on allosteric mechanisms [23]–[28].
Various computational methods have indeed provided essential information on the GroEL subunit dynamics. Notably, the transitions between the main functional states (T, R, and R″) have been further refined by utilizing NMA [27]–[29], targeted MD (TMD) [30], brownian dynamics [31], principal component analysis (PCA) [32], and MD simulations [33], [34]. Moreover, a number of computational studies have been dedicated to find pathways for both intra- and inter-ring communication in GroEL [35]–[40]. From these, several residues have been pointed out as important for the allosteric signaling thus increasing the understanding of the involved mechanisms.
The concept of the two-stage transition has been strengthened by the TMD simulation of the GroEL subunit which pulls the t form to the fully open r″ form along a 500 ps long MD simulation [30]. This computational study suggested that the transition begins with a downward tilt of helix M, and the subsequent counterclockwise twist of the apical domain. Moreover, biasing the MD simulation by employing temperature acceleration to increase the conformational sampling recently showed the ability of the isolated GroEL subunit to undergo the t to the r″ state transition [34]. However, this simulation was not able to sample the closing of the binding pocket as seen in the X-ray structures of the r″ conformers. Finally, unbiased MD simulation of the GroEL subunit has been performed in order to investigate the ATP-driven conformational changes [33]. This simulation samples the transition from the t to the semi-relaxed r conformation during 20 ns, nicely illustrating formation and rupture of hydrogen bonds.
A full scale monitoring of individual particle motions to probe the mechanistic basis for the transitions requires extensive unbiased conformational sampling at atomistic resolution. The 20 ns long MD simulations of Sliozberg and Abrams are too short to observe the full scale transitions of the subunit [33]. Moreover, significant variations between individual MD simulations have been detected [23], [41], [42], thus highlighting the importance of multiple simulations to extract statistically relevant information on the conformational changes. Investigating the positive (intra-ring) and negative cooperativity (inter-ring) of ATP-binding would require MD simulations of the entire GroEL oligomer. Such simulations on relevant timescales for conformational transitions are beyond the capabilities of present computational resources. It has also been reported the surprising facility to obtain stable folded monomers of GroEL by a variety of means [43]–[47], which can be explained by the small area buried by the monomers at subunit interfaces in the X-ray structure of the GroEL complex [8]. Furthermore, under some particular experimental conditions it seems that monomeric GroEL exhibits a weak chaperone activity [46]. Despite the fact that the chaperone activity of monomeric GroEL is a controversial issue [48], the ability of the monomer to fold into a native-like conformation that can bind nucleotide, points to the subunit as a relevant structural unit to be investigated by MD simulations.
In the current work we present extensive (in total long) unbiased MD simulations of the GroEL subunit starting from the closed (t) and open (r″) conformations, with and without bound nucleotide. We use PCA of 287 experimentally obtained GroEL subunits to interpret the conformational sampling of our simulations. We observe nucleotide dependent shifts in the conformational ensembles, and show that the subunit response, as observed in the oligomeric structure, is intrinsically coded in the structure-dynamics of the isolated subunit. These simulations provide so far unexplored sampling from t all the way to a structure close to r″. A nearly complete transition in the opposite direction is also sampled by removing ADP from the r″ form. Another interesting outcome of our simulations is that the inherent motion of unliganded GroEL subunit is biased along the transition pathway towards both the r and r″ states. Furthermore, the MD simulations reveal a weak stabilization of the equatorial domain upon ATP binding, resulting in a modest decrease of the configurational entropy of this domain. Conversely, a larger increase of entropy is found for the whole GroEL subunit. Finally, we decipher the underlying mechanisms for the conformational transitions by investigating the atomic interactions unique to the unliganded and nucleotide bound structures obtained from the X-ray structures and MD simulations. Several of the interactions that characterize the conformational intra-subunit effects brought about by ATP binding were not revealed in previous studies.
27 crystal structures of the GroEL complex, with a total of 364 subunits, were collected from the RCSB protein databank. Six of the crystal structures were not considered due to missing coordinates, which leaves 21 GroEL complexes with 287 subunits for the following analysis. These include ATP and ADP bound conformers as well as apo forms from wild type and mutant structures. The 287 GroEL subunits were superimposed onto the invariant ‘core’ defined as the area with least structural variation [49], and PCA was performed to investigate the major conformational differences between the collected structures. As much as 91.5% of the total variance of the atomic fluctuations was captured along the first principal component (PC), while 2 and 3 dimensions were necessary to capture 95.3% and 97.5%, respectively (see inset in Figure 1A).
The GroEL subunits can be divided into three major groups along the two first PCs; the closed cis/trans t forms, open cis r″ forms, and the semi-relaxed r forms (Figure 1A). The first PC is shown in Figure 1B and describes the main differences between the r″ and t conformers. The motions described by PC1 consist of (1) an upward movement of the apical domain away from the equatorial domain; (2) a small rotation of the apical domain; (3) a downward tilt of helix M (residues 386-409) and (4) a translation of helices F and G (residues 141–152 and 155–169, respectively) along their axis of inertia. This region of the intermediate domain moves down to cover the nucleotide binding site in the equatorial domain while the apical domain twists upwards. In addition, modest internal fluctuations are observed in the equatorial domain. The stem loop spanning from Arg36 to Lys51 has a flapping motion opposite to the intermediate domain, and helices A (residues 8-27) and C (residues 65-85) show a rotation relative to helices D (residues 88-107), E (residue 113-134), and R (residues 496-514). Similar internal motions in the equatorial domain are also observed in PC2. In particular, helix A (residues 8-27) undergoes a translational motion along its longitudinal axis, while both helices D (residues 88-107) and R (residues 496-514) rotate along their axis. PC2 represents the largest variation between the t and r conformers, consisting of a counterclockwise twist and a modest elevation of the apical domain. Similarly, PC3 describes a rotational motion of the apical domain together with a translational motion of helix A along its longitudinal axis.
The internal fluctuations within the equatorial domain reveal the presence of two sub-domains (helices A+C, and helices D+E+R) consistent with a previous study utilizing 35 subunit structures [32]. Moreover, it is interesting to notice that the invariant core is a relatively small part of the equatorial domain (helices E, O, and R), thus pointing out the structural variability within this domain.
To investigate the intrinsic response of nucleotide binding to GroEL we conducted 14 MD simulations of the isolated GroEL subunit starting from both ends of the reaction cycle; the closed (t) and open (r″) states, with (holo) and without (apo) bound nucleotide (ATP or ADP). Of these, we performed four 300 ns long simulations: (A) closed t unbound, (B) open r″ unbound, (C) closed t ATP-bound, and (D) open r″ ADP-bound. An additional ten 100 ns long simulations of the closed t form were also carried out: 5 (E–J) unliganded, and 5 (K–P) ATP-bound. Simulations (D) and (K) consist of the 0–100 ns interval of simulation (A) and (C), respectively.
Each conformer obtained from the MD simulations was projected onto the first two PCs determined from the X-ray structures (Figure 2). These projections display the relationship between the MD conformers in terms of the conformational differences described by the two first PCs, thus enabling interpretation of the conformational space sampled in each of the simulations.
Remarkably, the simulations of the closed subunit all sample the space along PC2, which describes the transition between the t and r form, independently of whether ATP is bound or not. The main difference between the unliganded and ATP-bound simulations lay in the sampling along PC1; the ATP bound simulations shift the ensemble of conformations along PC1, which thus comes closer to the fully open r″ form (Figure 2A+C). This difference is most apparent in the 300 ns simulations where the ensemble samples closest to the fully open structure (see Video S1). The difference in sampling along PC1 is also significant for the 100 ns simulations (Figure S10), however they are too short to fully observe this effect. Admittedly, 100 ns is not long enough simulation time to bring the ATP bound simulation (N) away from the t form (Figure 2N).
A similar shift of the ensemble along PC1 and PC2 is also observed for the open simulations. Interestingly, the ADP-bound simulation samples the space in the proximity of the r″ form (Figure 2D). Conversely, removing ADP yields a sampling of the conformational space which is significantly shifted, in particular along PC1, towards the t conformers (Figure 2B).
Clustering analysis on each of the four 300 ns long trajectories was performed to identify the predominant conformations throughout the simulations. Each frame in the trajectories is attributed to one particular cluster depicted as color bars in Figure 3 along with the resulting RMSd values. The conformations sampled in the closed apo simulation (Figure 3A) resembles its initial starting structure (t). Conversely, the closed holo simulation, shown in Figure 3C, comes much closer to the open r″ form than its apo counterpart. The RMSd values for the average conformations in each of the clusters with respect to both the closed and open X-ray structures are shown in Table 1. The predominant clusters of the closed apo simulation are clusters A1+A2 which have a relatively close similarity to the closed X-ray structure, with an RMSd value of 2.56 and 3.45 Å, respectively. These low RMSd values of the closed apo simulation stand in contrast to higher RMSd values of the closed holo simulation, i.e. 5.97 and 4.09 Å for the two dominating clusters C2+C3. Cluster C4 of the closed holo simulation, which consists of 92 member conformations out of a total of 1500 MD conformers, has a higher average similarity to the open (RMSd 6.86 Å) than the closed X-ray structure (9.21 Å). We thus sample relatively close to the experimental r″ conformer (minimal deviation of 5.8 Å) with the major difference being an open nucleotide pocket, similar to what a recent temperature biased MD simulation by Abrams and Vanden-Eijnden showed [34]. Interestingly, both the unbound and the ATP bound simulations also show a tendency to mimic the r form (minimal deviation of 2.6 Å) characterized by the counterclockwise rotation of the apical domain (i.e. clusters A3 and C1).
The RMSd values and clustering of the open (r″) apo and holo simulations are shown in Figure 3B+D. The apo simulation initially samples conformations around its starting point; close to the open X-ray structure. The predominant clusters of this simulation are B2 and B4 with a total of 933 member conformations, and the average conformations of these clusters have RMSd values much closer to the open than to the closed X-ray structure as shown in Table 1. After about 240 ns of simulation the structure undergoes a remarkable conformational change resulting in higher similarities towards the closed (t) form. These structures are assigned to cluster B6, and the member conformations have an average RMSd of 7.23 Å towards the closed X-ray, and 8.11 Å towards the open X-ray structure. We thus sample relatively close to the experimental t conformer starting from the r″ conformer (minimal deviation of 5.1 Å) with the major difference being a small twist of the apical domain. Conversely, the open holo simulation consistently shows an open structure with RMSd values closer to the open r″ than to the closed X-ray structure (Figure 3D and Table 1).
To probe the differences between the unliganded and ATP-bound simulations we investigated the residue fluctuations based on the 10–60 ns interval of the subunit simulations in the closed form. As expected, most residues show a similar fluctuation pattern in the holo and apo simulations. In the equatorial domain, residues adjacent to ATP show the most evident differences between the apo and holo simulations (see Figure 4A–B). In particular, residues Lys28, Lys34, Asp87, Asn457, Glu461, Tyr478-Glu483, and Tyr485, and its immediate neighbors are significantly () stabilized in the presence of ATP. Conversely, Glu61, Glu63, Arg421, and Asn475, show increased fluctuations in the holo simulations indicating rearrangement of these residues upon ligand binding. Modest differences in the fluctuation profiles are also observed within the apical (Figure S11A) and intermediate (Figure S11B–C) domains. Only two residues are shown to be significantly altered (); Arg350 (apical) and Lys390 (intermediate).
Since the equatorial domain possesses the ATP binding site it is thus the epicenter for the initiation of the conformational changes. In order to relate the differences in fluctuations to binding of ATP we calculated the potential hydrogen bonds between ATP and the protein along the simulations. Figure 5 shows the occupancy of each of the ATP hydrogen bonds in the closed holo simulations. Three hydrogen bonds are shown to be particularly strong with an occupancy of more than 90% of the simulation; Thr89 and Thr91 bind to the beta and gamma phosphate, respectively. The fifth single strongest bond is Asn479 which binds to the group of the adenine ring, with a occupancy of about 80% of the simulation. The negatively charged carboxyl group of Asp495 forms 4 alternating hydrogen bonds with the ribose OH, which have occupancies of approximately 40% each. Weaker bonds also exist, such as Gly32, Lys51, Thr90, and Ala480. Moreover, Asp87 is held in close contact with ATP through the tight binding to , resulting in a mean distance of 3.58 Å () between the oxygens of Asp87 and the three adjacent phosphate oxygens of ATP.
The identified hydrogen bonds between ATP and specific protein residues can be directly linked to the change in the fluctuation pattern observed in Figure 4. In particular hydrogen bonds involving Gly32, Thr89-Thr91, and Asn479-Ala480 might cause the stabilization of these areas. Moreover, the tight binding between Asp87 and might be a stabilizing factor for Asp87 and its neighbors.
Configurational entropies of the closed simulations were estimated to investigate the entropic penalty upon ATP binding. Entropy values calculated from quasi-harmonic analysis of MD simulations are sensitive to simulation length and the number of frames in which the calculation is based on [50]. Moreover, significant variation between individual MD has been reported [41], [51]. We thus performed entropy calculations on the multiple closed MD simulations for the equatorial domain and for the whole subunit (Figure 6). We observe only small changes in entropy within the equatorial domain upon nucleotide binding. Moreover, during the first 60 ns of the simulations, no entropy difference is found for the whole subunit, while after 80–100 ns the entropy of the ATP-bound simulation is slightly higher than that of the apo simulations.
Comparing distance maps between different conformations has the potential of revealing unique atomic interactions potentially important for the conformational transitions within the GroEL subunit. We used difference contact maps (DCMs) to identify atomic interactions (all heavy atoms) unique to the t, r, and r″ X-ray conformers (Figure S12A-E) complemented with DCMs obtained from the MD simulations of the closed form (Figure 7). Of particular interest are those residues which change interaction partner during the transition from t to r″.
A large number of contacts are found to be unique for either the t or the r″ forms (Figure S12A). High density of differential atomic interactions are observed within the equatorial domain, and in particular the interaction pattern between helices B, C, and D is altered upon the t-r″ transition. Contacts between helices B-C and C-D are found to be unique for the r″ conformers. Conversely, the t conformers show tighter interactions between helix C and the loop connecting helix C and D. Here, Asn82 contacts with Asp87 in the t conformers but with Arg58 in the r″. Moreover, Asp87 changes its contact partners from Asn82 in the t state to Ser151 and Asp398 in the r″ state, which is the main interaction between helix M of the intermediate domain and the equatorial domain.
Unique for the r″ forms is also a set of hydrophobic interactions between helices C and D (Val74-Ile100, Val77-Ala96, Ala81-Ala92, Val77-Ala92). Conversely, the t forms have unique contacts between helix C and R (Val74-Val510, Val77-Tyr506, Ala85-Val499).
The intermediate domain also undergoes conformational rearrangements during the transition. Residues downstream of helix G initially interact with residues of helix M in the t form, but change interaction with helix L and K during the transition, e.g. the interaction Glu172-Arg404 in the t forms is changed to Glu172-Arg350 in the r″ forms. Moreover, Asp179 contacts Lys390 at the intermediate domain in the t forms, and the Thr48 at the equatorial domain in the r″ forms.
While the ensemble of static X-ray structures provides the differential atomic interactions between the two end points of the reaction cycle, the MD simulations can reinforce the findings based on X-ray and further complement this by providing the differential atomic interactions at an earlier phase in the transition. Consistent with the findings from the X-ray-based DCM, the areas of high density differential contacts are situated in the equatorial and intermediate domains (Figure 7). Twenty differential contacts are consistently identified (Table 2). Perhaps the most interesting of these are the interactions Val412-Asn475 and Leu134-Asn475 located near the lower hinge (Figure 8A). These interactions show similar average distance difference in the X-ray and MD distance matrices. Moreover, they are located in the equatorial domain, close to the lower hinge, making them potentially important for the interaction between the equatorial and intermediate domain.
In the intermediate domain Asp179 and Lys390 connect helix M to three sheets (, , ) close to the apical domain in the t forms. These residues are 0.9 Å closer in the apo simulations than in the holo ones. This is consistent with the X-ray DCMs, which shows that Lys390 interacts with Thr48 in the equatorial domain.
The distance matrices can also be helpful to probe the underlying mechanisms for the changes in fluctuation pattern of the equatorial domain (Figure 4A–B). Table 3 summarizes the average difference in atomic distances for residues identified to have altered fluctuations. Of particular interest is Arg421 and Asn475 which show higher RMSF values in the holo simulations. Arg421 shows a tighter binding to Gly471 in the apo simulations than in the holo ones (). This is consistent with X-ray data, although the difference is about 1 Å smaller ( and ). Perhaps the most evident difference affects Asn475 which in the MD simulations appears 2.2 Å closer to Val412, 2.0 Å closer to Leu134, and 1 Å closer to Ala413. These findings are again consistent with the differences found in the X-ray structures; i.e. 1.8, 1.8, and 0.8 Å for the t-r″ difference, and 1.4, 2.5, and 0.7 Å for the t-r difference.
Interactions between 2-loop-3 and 16-loop-17, close to the ATP-binding site, are strengthened upon nucleotide binding which might explain the smaller RMSF values in this area (Figure 8B). The interaction Lys34-Glu483 is particularly affected; the average distance difference is 2.71 Å for the MD structures, while 0.8 and 1 Å for the X-ray structures. A significant difference is also found for Val54-A78 which are 1.4 Å closer in the holo simulation than in the apo simulations (3.5 and 1.7 Å for the X-ray differences).
Determining the residue cross correlation to investigate whether the motions of one residue are related to the motions of another can aid in deciphering the underlying mechanisms for the observed conformational transitions. We have previously reported the cross-correlation map for the entire GroEL subunit revealing global correlations [23]. In the present work we focus on the equatorial domain which has the advantage of revealing more detailed correlations around the binding site. The corresponding residue-residue correlation map for the equatorial domain is shown in Figure 9A, and highlights 7 off-diagonal correlated areas. These areas are consistently described as correlated throughout all subsets of the simulation (10, 20, and 50 ns intervals), and are the following: (1) residues Lys16-Gly20 (helix A) and Phe66-Asn68 (helix C), including two positively charged residues in helix A, and one negatively charged residue in helix C; (2) residues Gln454+Ile455+Asn458 (helix P) and Thr31-Gly33 (loop region between helix A and 2); (3) The latter region is also correlated to the phosphates of ATP; (4) the phosphates of ATP and residues Gly89-Thr92 (helix D); (5) residues Asp115-Ile119+Ala123 and Asp435-Asn437+Ile440+Ala443; (6) Val411-Val412 (15) and Leu494-Thr497 (18) close to the lower hinge; and (7) the adenine ring of ATP and residues Asn479-Thr482 (16-loop-17 region).
Several computational structural biology attempts have focused on the investigation of the ligand-induced conformational changes in GroEL [28], [30], [31], [33], [34]. Comparison of crystal structures at different stages of the functional cycle, NMA analysis and targeted MD simulations provide important information on the conformational space and motions associated to ligand-binding. Nevertheless, all-atom unbiased MD simulations have a superior potential to reveal the mechanisms involved in conformational transitions in proteins since using targeted MD simulations applying fictitious driving forces bias the motions toward the target and may drive the transitions along unrealistic deformations. However, GroEL has been considered to be beyond the analysis by unbiased MD simulations which have been held back due to the large size of the protein. Nevertheless, Sliozberg and Abrams have previously performed a 20 ns long unbiased MD simulation of one subunit, investigating the ATP-driven conformational changes [33]. This simulation manages to sample the transition from the t unbound, tense conformation to r, defined as a semi-relaxed conformation intermediate towards the open r″ conformation. This transition includes a counterclockwise rotation of the apical domain along with closing of the nucleotide binding pocket, and was associated to the initial transition in a 2-stage activation by ATP [13], [30].
The T to R transition has been associated to the intra-ring cooperative allosteric response to ATP binding, since functionally the R conformer shows high affinity for ATP [13]. Following this reasoning Sliozberg and Abrams interpreted their results on the way the r structure would be concertedly transmitted to all seven subunits in the cis ring. In our present simulations of the isolated subunit, we sample close to the r structure (minimum deviation of 2.6 Å), and in addition, these long MD simulations (300 ns) provide so far unexplored insight into the ATP-induced transition from r all the way to a structure that is, according to the RMSd values, close to r″ (minimum deviation of 5.8 Å). An almost full transition in the opposite direction, from r″ to t, is revealed from a similar simulation of the ligand free apo state starting from the r″ structure (minimum deviation of 5.1 Å). This transition would correspond to the conversion occurring in the trans ring, which functionally is accompanied by ejection of the hydrolysis product ADP and the co-chaperonin GroES. We thus show for the first time with unbiased MD simulations that the conformational transitions in the GroEL oligomer are favored by the intrinsic behavior of the isolated GroEL subunit. The predictive power of the present study is reinforced by the multiple 100 ns long MD simulations.
The conformer plots and the fluctations along the simulations also reveal that the r conformation is sampled both in the presence and absence of ATP. Thus, our results further add to the accumulating experimental and theoretical proofs of the pre-existing equilibrium between inactive and active states (see e.g. [52]–[56]). A similar trend was also detected in a recent fluorescence study where the authors were able to measure the fraction of molecules in the T and R state of GroEL with increasing ATP concentrations [57]. They determined that about 50% of the molecules were in the T state even with high ATP concentrations, indicating a constant cycling between the two conformations. Moreover, Chaudhry et al. showed that the inherent motions of the unliganded GroEL were “biased along the transition pathway that leads to the folding-active state” [58]. X-ray structures of the nonhydrolyzable ATP analog , in which the domain rotations are not observed [9], give means to further support a t to r cycling.
These unbiased simulations also show that the open structure, similar to the r″ structure, is attained for the isolated subunit, just in the presence of ADP, and without the need of the co-chaperonin GroES. It is accepted that binding of GroES to the heptameric cis ring with the subunits in the r conformation is the main functional determinant for the induction of the r″ conformer, triggering the encapsulation of bound protein substrate into the cavity. Our results thus support that the subunit conformational changes along the functional cycle of GroEL, as revealed by X-ray crystallography and cryoelectron microscopy, are largely intrinsic to the 3D structure (sequence-structure-dynamics) of the subunit (intramolecular), though stabilization of intermediate conformations should be associated to intersubunit or interprotein (intermolecular) interactions in the GroEL-GroES oligomers.
Isolating a monomer from its natural biological assemblage may result in an altered dynamical behavior. However our results indicate that the dominant motion of a single GroEL subunit resembles that occurring in the ensemble of GroEL complex crystal structures. Thus, the dynamics of the subunit alone seems to be reflected in the larger biological assembly. This is seen by the RMSd analysis of the trajectories (Figure 3), as well as in the comparison between the X-ray and MD derived PCs (root mean squared inner product = 0.73) (Figure S13). The dominance of this intrinsic motion might be related to the relatively small inter-subunit contact area (6.6%, or ) [8], which is comparatively low for an oligomeric protein [59], [60]. Nevertheless, each of the 7 apical domains interacts with two neighboring apical domains, as well as with one intermediate domain in the initial closed T form [8]. This arrangement imposes constraints on the motion of the subunit, though at present to an unknown degree. It is therefore expected that, for similar timescales, simulating the entire GroEL assembly would lead to a more restricted conformational sampling of the subunit which would most likely be more consistent with the timescale upon which these conformational transitions are thought to occur [15].
Sliozberg and Abrams described that MgATP binding to the subunit in the t conformation initiates a conformational change mostly due to the strong interaction of and Asp398 [33]. This interaction induces a series of H-bond ruptures (such as the Asp155-Arg395 salt bridge) and H-bond formations that were described in an induced-fit, cascade-like way [61]. We were however unable to reproduce this pattern of interactions. As seen in the conformer plots (Figure 2) the effect of ATP binding is rather described by causing a shift of conformational equilibria towards the r″ conformation, with the peculiarity that in this case the ATP-bound conformation corresponds to the open conformation of the subunit, while usually ligand binding leads to a closed structure for the majority of proteins. In fact, in GroEL, ATP binding leads to a closed and less dynamic equatorial-domain structure but open and more dynamic (r and r″) subunit structure.
In the case of preexisting equilibria where the ligand binds selectively to an “active” conformation the energy barrier between the conformations in equilibrium should be low [52], and binding should not bring a high entropic penalty, as certainly seems to be the case for the GroEL subunit (Figure 6). Further experimental support for low energy barriers is also found for Hsp70 proteins [62]. Moreover, a thermodynamic study on nucleotide binding to GroEL provided a positive entropy change for the binding of ATP at temperatures higher than [63]. Notwithstanding the fact that these measurements were performed on the complete oligomer, it is noteworthy to relate the trend in entropy change to that found for the isolated monomer.
Computational studies employing advanced procedures in conjunction with NMA and brownian dynamics have probed allosteric networks in the GroEL system [31], [37], [39], [40]. They have been able to determine and highlight a large set of residues responsible for the transitions in the GroEL complex. Lys80-Asp359, Asp83-Lys327, Arg58-Glu209, Pro33-Asn153, and Gly257-Arg268 are among several intra-subunit interactions that have been identified to be important during the conformational transitions in GroEL [31], [39]. Of these, Arg58, Asp83, Gly209, and Lys327 were also highlighted by Tehver et al. [40], and shown to be highly conserved [64].
Our study of the MD trajectories and the large amount of X-ray data reinforces these studies as we identify many of the same interactions and residues. Moreover, we identify several other atomic differential interactions brought about by ATP-binding which have not previously been detected. Of particular interest is the tight binding between Lys34 and Glu483 in the presence of ligand (Figure 8A), which is captured in both the MD simulations and the X-ray crystallographic data. This effect appears as an important rearrangement induced by ATP-binding. These charged residues are situated close to the nucleotide binding site in two separate loops (16-loop+17 and A-loop-2), which movements interestingly are highly correlated to ATP; residues Asn479-Thr482 to the adenine ring, and Thr31-Gly33 to the phosphates of ATP. The correlation between these areas and ATP is likely to contribute to the stabilization of these loops as observed by fluctuation calculations (Figure 4), which in turn might aid in the salt-bridge formation between Lys34 and Glu483 (Figure 8B).
More peripheral to the ATP-binding site in the equatorial domain, close to the lower hinge, we observe weaker contacts between Leu134-Asn475, Arg421-Gly471, and Val412-Asn475 (Figure 8A). This is possibly due to rearrangements of helix N upon nucleotide binding which disrupts the initial interactions found in the t conformers and along our MD simulations of the unbound t forms. Moreover, the DCM analysis pinpoint an altered interaction pattern for helix C in the equatorial domain, possibly explaining the movements of helices A+C in the PCA analysis of X-ray structures. Among these interactions, hydrophobic amino acids are a common denominator; Ala84-Tyr506, Val77-Ala92, Val54-Ala78, Val77-Ala507 all show a tighter binding in the nucleotide bound conformers. Experimentally, this core of hydrophobic amino acids has been shown to be important, and e.g. the Ala92Thr protein vaiant shows a low ATPase activity [65]. Interestingly, the MD simulations also capture the rupture of the Asp179-Lys390 contact within the intermediate domain, and the formation of Arg231-Glu310 contact in the apical domain upon ATP-binding, both consistent with X-ray data.
Realizing that a deep understanding of the function of GroEL as a molecular machine certainly requires analysis of inter-subunit and inter-ring interactions in the oligomer, MD simulations of the complete GroEL complex would be required to probe the full effects of ATP-binding. While these are expected to reveal essential details on the allosteric regulation of GroEL function, multisubunit simulations of a system of 600.000 atoms are a very computationally-demanding task. Regardless of the capacity of GroEL to assemble into oligomeric structures, various studies have also shown the surprising facility to obtain stable folded monomers of GroEL by a variety of means [43]–[47], thay might display a weak chaperone activity of the GroEL monomer [46].
In the present work we have attained detailed and statistically proven results from MD simulations of the isolated GroEL subunit, which have highlighted the importance of considering the proper dynamics and response of the subunit in the context of the large scale transitions in the GroEL complex. Since the equatorial domain holds the ATP binding site, it constitutes the epicenter of the conformational changes in the GroEL complex. By paying extra attention to the detailed mechanisms in this region of the protein during the conformational transitions observed along the simulations we have been able to map important intramolecular rearrangements which potentially are a prerequisit for the intermolecular transitions to occur.
All available GroEL structures were collected from the RCSB protein databank [66]. Six structures were omitted due to large amount of missing coordinates (PDB-codes: 2CGT, 1GR5, 1IOK, 2C7C, 1GRL, 3CAU). A total of 21 crystal structures (287 GroEL subunits) were kept for further analysis (PDB-codes: 1PCQ, 1PF9, 1SVT, 3C9V, 1AON, 1GRU, 1MNF, 1XCK, 2C7D, 2NWC, 3E76, 2EU1, 1SS8, 1SX3, 1J4Z, 1KPO, 2C7E, 1KP8, 1OEL, 1WE3, 1WF4). Structural superposition was performed on the invariant “core” as defined by Grant et al. [49]. These structures were collected and analyzed using the Bio3D package [49].
The calculation of the PCA modes involves two main steps; (1) the calculation of the covariance matrix, , of the positional deviations, and (2) the diagonalization of this matrix [67], [68]. The dimensional covariance matrix is calculated based on an ensemble of protein structures, and the elements of are defined as(1)where and are atomic coordinates and the brackets denote the ensemble average. The diagonalization of the symmetric matrix involves the eigenvalue problem(2)where is the eigenvectors and the associated eigenvalues.
Our PCA calculations were based on the C coordinates of the ensemble of the 287 GroEL crystal structures and was performed with the Bio3D package [49]. The X-ray and MD conformers were projected into the sub-space defined by PC1 and PC2, where the maximum variation of the conformational distribution was observed. Plotting these projections results in ‘conformer plots’ which displays a low dimensional representation of the conformational change in terms of the two principal components.
All-atom MD simulations of the closed and open forms of the isolated GroEL subunit were performed both with and without bound nucleotide (designated holo and apo, respectively). The atomic models were prepared from the high-resolution crystal structures with PDB codes 1XCK chain A [69] and 1SVT chain A [58], for the closed and open forms, respectively. MgATP coordinates for the closed holo (with nucleotide) simulation was collected from the crystal structure with PDB id 1KP8 chain A [70]. All atomic models were prepared with Amber10 [71] and the corresponding Amber03 forcefield [72], [73]. ATP and ADP parameters were obtained from Meagher et al. [74]. For each of the simulations, the protein was solvated in a periodic truncated octhahedron box with TIP3 water molecules [75], providing 16 Å of water between the protein surface and the periodic box edge. The solute was minimized for 10,000 steps, followed by 10,000 steps of minimization of the entire system. The protein was then heated to 100 K with weak restraints for 100 ps, and to 300 K in 200 ps. 2 ns of equilibration with constant pressure and temperature (NPT) of the system was performed prior to the production run in order to ensure correct density. The production runs were performed with constant volume and energy (NVE) with a 1 fs time step, using SHAKE constraints on hydrogen-heavy atom bonds.
A total of 14 MD simulations of the GroEL subunit were carried out. Of these, we performed four 300 ns long simulations: (A) closed t unbound, (B) open r″ unbound, (C) closed t ATP-bound, and (D) open r″ ADP-bound. Additionally ten 100 ns long simulations of the closed t form were performed: 5 (E-J) unliganded, and 5 (K-P) ATP-bound. Simulations (D) and (K) consist of the 0–100 ns interval of simulations (A) and (C), respectively.
Distance matrices were calculated between all 3856 pairs of atoms in order to monitor the atomic interactions (between residues at least four residues apart in sequence) [49]. The distance matrix was issued to residue grouping by only considering the minimal atomic distance between the residue pairs. Residue pairs closer than 4 Å are assumed to be in contact, and constitute the contact matrix for one particular conformation. Contact matrices were calculated for 28 closed (T) apo subunits (PDB code: 1XCK, 1SS8, 1OEL), 28 closed (T) ATP bound subunits (PDB code: 1KP8, 1SX3) and 28 open (R″) ADP bound subunits (PDB codes: 1AON, 1SVT, 1SX4, and 1PF9), and 6000 snapshots obtained from the last 50 ns of the 12 independent MD simulations on the closed GroEL subunit (6 apo and 6 holo). Only contacts with at least 50% occupancy and an average distance difference of 0.5 Å were considered. The difference of two contact maps (DCM), i.e. difference between apo and holo, then defines side-chain contacts which exist in one form, but not in the other.
Clustering analysis, correlation maps, entropy calculations, and hydrogen bond analysis were performed with the the ptraj module of AmberTools [71]. Clustering was performed on the MD conformers using the average-linkage clustering algorithm [76]. Cross-correlation calculations were performed on several subsets of the closed ATP bound simulations (10, 20, and 50 ns intervals) in order to obtain areas of consistent correlations. All figures of GroEL are made in Pymol [77].
|
10.1371/journal.pntd.0002222 | Fine Analysis of Genetic Diversity of the tpr Gene Family among Treponemal Species, Subspecies and Strains | The pathogenic non-cultivable treponemes include three subspecies of Treponema pallidum (pallidum, pertenue, endemicum), T. carateum, T. paraluiscuniculi, and the unclassified Fribourg-Blanc treponeme (Simian isolate). These treponemes are morphologically indistinguishable and antigenically and genetically highly similar, yet cross-immunity is variable or non-existent. Although all of these organisms cause chronic, multistage skin and systemic disease, they have historically been classified by mode of transmission, clinical presentations and host ranges. Whole genome studies underscore the high degree of sequence identity among species, subspecies and strains, pinpointing a limited number of genomic regions for variation. Many of these “hot spots” include members of the tpr gene family, composed of 12 paralogs encoding candidate virulence factors. We hypothesize that the distinct clinical presentations, host specificity, and variable cross-immunity might reside on virulence factors such as the tpr genes.
Sequence analysis of 11 tpr loci (excluding tprK) from 12 strains demonstrated an impressive heterogeneity, including SNPs, indels, chimeric genes, truncated gene products and large deletions. Comparative analyses of sequences and 3D models of predicted proteins in Subfamily I highlight the striking co-localization of discrete variable regions with predicted surface-exposed loops. A hallmark of Subfamily II is the presence of chimeric genes in the tprG and J loci. Diversity in Subfamily III is limited to tprA and tprL.
An impressive sequence variability was found in tpr sequences among the Treponema isolates examined in this study, with most of the variation being consistent within subspecies or species, or between syphilis vs. non-syphilis strains. Variability was seen in the pallidum subspecies, which can be divided into 5 genogroups. These findings support a genetic basis for the classification of these organisms into their respective subspecies and species. Future functional studies will determine whether the identified genetic differences relate to cross-immunity, clinical differences, or host ranges.
| Pathogenic treponemes include three subspecies of Treponema pallidum (pallidum, pertenue, endemicum), T. carateum, T. paraluiscuniculi, and the unclassified Fribourg-Blanc treponeme. Although they share morphology and have very similar antigenic profiles, they have traditionally been distinguished by mode of transmission, host specificity and the clinical manifestations that they cause. The molecular basis for these disease characteristics is not known. Comparative genomics has revealed that sequences differences among the species and subspecies are found in very localized regions of the chromosome. Many of these regions of sequence variation are found in the tpr genes, which encode a family of twelve candidate virulence factors, many of which are predicted to be outer membrane proteins. Most of the tpr-specific sequence changes are consistent within subspecies or species, supporting the historical classification of these organisms into separate subspecies and species. Functional studies are needed to determine whether any of the tpr gene differences are related to differences in host range, immunity, or clinical manifestations.
| Non-cultivable pathogenic treponemes include three subspecies of Treponema pallidum: T. pallidum subsp. pallidum (T. p. pallidum), T. pallidum subsp. pertenue (T. p. pertenue) and T. pallidum subsp. endemicum (T. p. endemicum). These subspecies are human pathogens and cause venereal syphilis, yaws and bejel, respectively. Other very closely related species or isolates are Treponema paraluiscuniculi and the Fribourg-Blanc or Simian treponeme. T. paraluiscuniculi causes venereal syphilis in rabbits and is reportedly not infectious for humans [1], [2]. The unclassified Simian treponeme was isolated from a baboon, causes a yaws-like disease in non-human primates, and is able to cause active infections in humans [3]–[5]. All of these organisms can be propagated in rabbits and cause disease following experimental inoculation of rabbits. Treponema carateum causes the human disease, pinta, but no strains of this organism are available.
The infections caused by T. pallidum organisms are characterized by chronic infection with distinct early and late clinical manifestations. Syphilis, usually a sexually transmitted infection, is a highly invasive process and can involve virtually any organ or system including the central nervous system. In pregnant women, early syphilis infection often results in transmission to the fetus. Each year, approximately twelve million new cases of syphilis are estimated to occur globally [6], [7]. Yaws and bejel affect approximately 3 million people worldwide and are transmitted by non-sexual direct contact, usually during childhood and largely affecting people living in remote villages in developing countries. Yaws and bejel have predominantly skin or mucous membrane and osseous manifestations [8]–[10], with tissue destruction late in infection. Pinta causes significant skin discoloration in the late stages, but rarely causes tissue destruction. Unlike syphilis, these infections are said not to affect the central nervous or the fetus [9], although some scientists question this statement [11]. T. paraluiscuniculi infection in rabbits appears to be a chronic, but clinically mild, process characterized by long-lasting crusty lesions of the genitalia, nose, and mouth [12]. Treponemal infections in non-human primates have not been traditionally associated with genital disease; however, a recent study by Knauf et al. [13] reports asymptomatic, moderate or severely destructive genital lesions (and perhaps sexual transmission) resembling human syphilis, caused by organisms classified phylogenetically as more closely related to the Fribourg-Blanc and T. pallidum subsp. pertenue isolates.
The molecular basis for host specificity and the different clinical manifestations caused by these treponemes is not known. These organisms are morphologically identical [1], [3], [14]–[17] with very similar antigenic composition [18]–[23], stressed by the fact that, to date, infection-induced antibody or cellular immune responses cannot distinguish species, subspecies or strains. Protective immunity is induced only by long-term infection and is subspecies-specific [24]. In cross-immunity experiments [1] in which initial infections in the rabbit model lasted at least 3 months, three scenarios are observed: 1) inoculation with a particular strain results in complete protection against re-infection with the homologous strain, 2) protection against re-infection with another strain of the same subspecies is variable or non-existent, and 3) protection against challenge with other species or subspecies is absent. These cross-immunity observations are in concordance with inoculation studies in humans conducted by Magnuson et al. [25]. Subjects with treated late latent syphilis challenged with the Nichols strain had either of two outcomes: 1) those that did not develop either clinical signs or serological evidence of re-infection, indicating immunity; and 2) those that had increases in serological titers and/or development of darkfield positive lesions after inoculation, interpreted as active reinfection with the challenge strain. Although there was no evidence for waning immunity in the subjects who were susceptible to reinfection, this is a possible explanation. However, the lack of cross-immunity among highly similar species/subspecies may also reflect differences in a set of immunologically “inconspicuous” epitopes, underlying immunodominant, but not protective, antigens such as Tp47 (TP0574).These immunodominant antigens may act as decoy systems as described for other bacterial pathogens [26].
Recent comparative analyses of whole genome sequences [27]–[30] (Giacani et al., unpublished) reported <0.1% sequence differences among T. p. pallidum strains [29]; <0.2% between T. p. pertenue and T. p. pallidum subspecies [29]; and <1.2% between T. paraluiscuniculi and the human treponemes [31], [32]. Sequence diversity is primarily localized to six hot spots [29], which include regions encoding several members of the tpr gene family. The Tpr proteins represent candidate virulence factors, and have been the focus of intense research for the last decade. As a consequence, the distinct clinical presentations, host specificities, and variable cross-immunity studies suggest that foci of sequence diversity, including the tpr genes, may be the basis for explaining the differences described above for the treponemal infections.
Sequence homology divides the tpr family, a group of twelve paralogs, into three subfamilies: Subfamily I (tpr C, D, F and I), Subfamily II (tprE, G and J) and Subfamily III (tprA, B, H, K and L). As we progressively gain a better understanding of this gene family, an essential role for many of these genes is more apparent. Several studies show that the Tpr antigens are expressed during infection and are able to elicit marked antibody and cellular immune responses in the infected host [33]–[40]. Of the encoded Tpr antigens, TprA, B, C, D, E, F, I, J and K have been predicted to be outer membrane proteins (OMP) [33], [40], [41]. Opsonization and/or vaccine studies with these proteins support surface exposure [33], [37], [42]–[44] and both antigenic variation (TprK) [45] and phase variation (TprE, G, J) [46] mechanisms have been identified in tpr members. Yet, the high invasiveness and ability of T. pallidum to persist for decades in the host suggest that this spirochete may rely not only on antigenic and phase variation for survival. To influence infection outcomes, T. pallidum may also employ other strategies including genetic drift, genetic shift and or pathoadaptive point mutations, which can arise either during long term evolution or rapidly during a single infection. An important body of evidence has accumulated showing genetic variation in specific regions of the T. pallidum genome among subspecies and among strains [47]–[56]. The present study demonstrates significant sequence diversity in the tpr gene family, which can have important implications in understanding evolution of these organisms, as well as cross-immunity, strain typing and vaccine design.
No investigations were undertaken using humans or human samples in this study. New Zealand white rabbits were used for strain propagation. Animal care was provided in accordance with the procedures outlined in the Guide for the Care and Use of Laboratory Animals, and all work was conducted under protocols approved by the University of Washington Institutional Animal Care and Use Committee.
T. pallidum subspecies, T. paraluiscuniculi, and the Fribourg-Blanc treponeme were propagated in New Zealand white rabbits by intratesticular inoculation as previously described [57]. DNA was extracted for PCR amplification from the following isolates: T. pallidum subsp. pallidum (Sea 81-4, Mexico A, Bal 3), T. pallidum subsp. pertenue (Gauthier, CDC2, Samoa D), T. pallidum subsp. endemicum (Iraq B, Bosnia A), the Fribourg-Blanc treponeme (Simian isolate) and T. paraluiscuniculi (Cuniculi A). These strains were selected to represent different species/subspecies, geographical regions of origin, years of isolation, and anatomical sources (Table 1). The sequences of the tpr genes for the T. p. pallidum Nichols and Street 14 strains were downloaded from their corresponding genome sequences, GenBank accession numbers NC_000919.1 and NC_010741.1, respectively [27], [28]. Although we determined tpr sequences for a number of other strains of T. pallidum subsp. pallidum, only strains defining the 5 identified genogroups of T. p. pallidum are included in this manuscript. To ensure that the correct strain was propagated and extracted, only one strain of treponeme was handled at any time during the propagation and freezing process, and rabbit ear tags as well as labels on tubes were double-checked. Bacteria were extracted from infected rabbit testes in sterile saline, collected in sterile 1.7-ml microcentrifuge tubes, taking precautions to prevent cross-contamination between samples, and spun immediately in a microcentrifuge at 1,000g for 10 minutes to remove rabbit debris, followed by centrifugation of the supernatant at 12,000g for 30 min at 4°C [57]. Pellets were resuspended in 200 µl of 1X lysis buffer (10 mM Tris [pH 8.0], 0.1 M EDTA, 0.5% sodium dodecyl sulfate), and DNA was extracted with the Qiagen (Chatsworth, Calif.) kit for genomic DNA extraction as described in the manufacturer's instructions, but adding 50 ul of proteinase K (100 mg/ml stock solution) and incubating the sample for 2 h at 65°C. After the final elution step in 200 µl of H2O, DNA was used for analysis by PCR and sequencing.
The Nichols T. pallidum genome sequence [27] was used to design primers in the 5′ and 3′ flanking regions of the tpr genes to amplify the corresponding DNA regions from genomic DNA of the 10 treponemal strains. Table S1 lists the primers used for amplification and sequencing. Using genomic DNA as a template, whole ORF amplifications were performed in a 50-µl final volume containing 200 µM deoxynucleoside triphosphates, 1.5 mM MgCl2, and 2.5 U of GoTaq DNA polymerase (Promega, USA). For larger amplicons such as the tprG-F or tprJ-I operons [38], the LongAmp Taq PCR Kit was used as instructed by the manufacturer (New England Biolabs, USA). The products were cloned into the pCRII-TOPO or TOPO-XL (long amplicons) cloning vectors (Invitrogen, USA) according to the manufacturer's instructions. Plasmid DNA was extracted by using the Qiagen Plasmid Minikit (Qiagen, USA), and two to ten clones for each strain were sequenced with the Applied Biosystems dye terminator sequencing kit (Perkin-Elmer, USA). Consensus sequences were obtained with the CAP sequence assembly program [58] and ORFs from each strain at each locus were aligned using the MAFFT alignment program [59]. GenBank accession numbers are listed in Table S2. Structural homologs were identified using the 3D jury approach [60]. Structural (3D) models for TprC, TprD and TprI were generated using the TMBpro algorithm [61]. The orientation of the predicted loops in the TMBpro models, surface exposed vs. periplasmic, was determined as previously described by Randall et al. [61]. Signal peptide predictions were performed using the Predisi algorithm [62].
The tpr loci defined in the original Nichols genome sequence [27] were used to determine corresponding genes in 11 additional treponemal strains, including four T. p. pallidum, three T. p. pertenue, two T. p. endemicum, one T. paraluiscuniculi, and the Fribourg-Blanc strains (Table 1). TprK is excluded from this analysis because of the already extensive work that has been done on this gene [33], [34], [42], [43], [45], [53], [63]–[67]. Sequence analyses of the tpr loci from these strains identified significant heterogeneity within and among pallidum subspecies, the Fribourg-Blanc isolate and T. paraluiscuniculi. Figure 1 summarizes our findings. While Subfamilies I and II display a wide range of changes, diversity in Subfamily III is limited largely to the tprA and tprL loci. The observed changes are quite diverse, including SNPs of synonymous and non-synonymous character, indels, deletions of entire ORFs, chimeras, and alleles with large unique regions. Readers interested in the entire spectrum of sequence modifications identified in this study are referred to DNA and amino acid sequence alignments for each locus appended as Figure S1. For practical and comparison purposes, the tpr loci annotated in the Nichols genome sequence [27] will be considered as the reference ORFs.
It is noteworthy that many of the sequence changes divide the T. pallidum strains cleanly by subspecies or species: examples include no tprI ORF in endemicum strains; and tprG/J chimeras in both tprG and tprJ loci in pertenue strains while these two loci contain GI and GJ chimeras, respectively, in endemicum strains. In some cases, the sequences clearly divide syphilis vs. non-syphilis T pallidum subspecies: as an example, the absolute conservation of tprF (with frameshift) and intact tprI sequences in the syphilis strains, while tprI-like sequences are found in these loci in non-syphilis strains. Within subspecies, the syphilis strains demonstrated the most heterogeneity, being divided into five genotypes.
Subfamily I tprs include the tprC, D, F, and I loci. Initial examination of deduced protein alignments from the Nichols strain showed a significantly high degree of sequence conservation within Subfamily I at the amino and carboxyl termini, with central unique regions [33]; however, discrete heterogeneity was later evident in the amino and carboxyl regions when additional strains were analyzed. The Nichols TprC/D proteins reportedly have porin activity and an OM localization [68]. Although not yet experimentally demonstrated, TprF and I are also predicted to have a cleavable signal peptide and to be surface exposed [33], [40], [41], [44].
The tprC and tprD loci in the reference Nichols genome contain two identical coding sequences [27]. Earlier studies [37], [56] identified tprC and tprD variants among strains and among the three pallidum subspecies. The present study significantly expands our knowledge of the sequences in the tprC and tprD loci, and a schematic representation of all variants at the C and D loci identified to date is presented in Figure 2. Among the treponemal strains tested in this study, four alleles are found at the tprD locus: the reference tprD (Nichols), the tprD2 allele (Bal 3, Mexico A, Sea 81-4, Street 14, Samoa D, Iraq B, Bosnia A, Fribourg-Blanc), a predicted truncated tprD2 (Cuniculi A), and the tprD-like variants (Gauthier, CDC2). We previously referred to the sequence in the tprD locus of Gauthier as tprD3 [56]. However, we have now found a very similar (but not identical) sequence in the CDC2 strain, and we have chosen to call these “tprD-like” sequences, which are further described below. As in Nichols, those T. p. pallidum strains that have the Nichols tprD allele in the D locus also contain an identical copy of tprD in the C locus [37], and none of the non-syphilis treponemes carries tprC/D ORFs identical to the Nichols strain. As previously reported by our group, tprD2 has four unique regions that differentiate it from Nichols tprD and the tprD-like sequences: a 330-bp central region and three smaller regions toward the end of the open reading frame (Figure 2) [56]. The tprC locus of the tprD2-containing Bal 3, Sea 81-4, Street 14 and Mexico A T. p. pallidum strains contains tprC-like ORFs, with small sequence changes compared to the Nichols tprC. [37]. Overall, the sequence homology among tprC alleles is >95%. All pertenue, endemicum and the Fribourg-Blanc strains also have tprC-like sequences. As previously reported for T. paraluiscuniculi, Cuniculi A strain [31], [32], [36], the tprC and D loci are occupied by two truncated tprD2 variants.
In both tprC and tprD, sequence variation does not occur randomly, but rather is found in discrete variable regions (DVRs; Supplemental Figures 2.1.and 2.2 in Figure S2). In the majority of cases, these base pair changes result in amino acid changes. Pore-forming activities for TprC/D have been recently reported by Anand et al. [68]. 3D predictions of peptides without signal peptides suggest typical β-barrel structures of 22 antiparallel transmembrane regions resulting in 11 loops at each end of the structure (Figure 3, top panel). Our analysis of 22 TprC/D sequences demonstrated seven DVRs, all of which co-localize with surface-exposed external loops predicted by the 3D models (Figure 3, and Supplemental Figure 2.1 and 2.2 in Figure S2). In addition, these 3D predictions suggest four external loops with conserved sequences, located primarily in the amino-half of the proteins. This sequence variation in predicted surface-exposed peptide loops could have significant implications for cross-immunity.
In the Nichols genome, tprF and tprI loci are 1107 and 1827 nucleotides long, respectively. Their sequences are identical except that tprF is a truncated version of tprI due to a 720 nucleotide deletion (spanning the central and most of the 3′ region) in tprF, resulting in a shorter ORF, frameshifting and a premature termination [27]. In T. p. pallidum strains, tprF genes are identical in all isolates sequenced to date (Figure 1 and Supplemental Figures 1 and 2.3 in Figure S1 and S2). In contrast to the syphilis strains, the pertenue and Fribourg-Blanc isolates have a full length (not frameshifted) duplicated tprI-like gene at the tprF locus. Interestingly, however, the tprF locus is deleted in the endemicum strains Iraq B and Bosnia A, and in T. paraluiscuniculi. tprI loci are virtually identical to each other in T. p. pallidum strains except for the presence of a few synonymous SNPs in the 5′ and central regions reported in Street 14 [28]. In contrast, however, tprF or tprI ORFs are absent in the rabbit pathogen T. paraluiscuniculi. [36].
For a more detailed analysis of the polymorphism observed in the tprF and tprI loci, a sequence alignment was generated including all genuine (not truncated or replaced) tprF and tprI loci from 11 strains (Supplemental Figure 2.3 in Figure S2). TprF and TprI in syphilis and non-syphilis organisms display DVR patterns resembling the heterogeneity observed in TprC and TprD above, though to a lesser extent (Supplemental Figure 2.3 in Figure S2). Changes are clustered in 9 DVRs spread throughout the protein sequences. Deduced TprF and TprI proteins are also predicted to be outer membrane proteins [33], [40], [41], [68]. Structural predictions also suggest that TprF and TprI are homologs of transport porins with OM localization, and 3D predictions of TprF/TprI peptides without signal peptides (Figure 3, bottom panel) yield typical β-barrel structures. Similar to TprC/D, all TprF/I DVRs show co-localization with predicted surface exposed loops (Figure 3 bottom panel, and Supplemental Figure 2.3 in Figure S2), again suggesting an important role for these variable regions during infection.
The Subfamily II genes include tpr E, G, and J, which code for proteins nearly 800 amino acids in length with highly conserved amino termini, unique central regions and carboxyl ends with small unique gene-specific signatures [33]. tprE shows very limited sequence variation among strains and subspecies, however, the observed changes clearly segregate syphilis from non-syphilis treponemes and the Fribourg-Blanc strain (Supplemental Figure 1.4 in Figure S1). T. paraluiscuniculi has a tprGJ chimera (predicted truncation) in the tprE locus [36].
In contrast, the tprG locus is more diverse in its gene sequence, in that five different groups of ORFs can be found (Figure 1, Figure 4 and Supplemental Figure 1.6.1 and 1.6.2 in Figure S1): 1) tprG sequences as described in the Nichols genome (Bal 3, Street 14); 2) a truncated tprG due to two single and one 3-nucleotide insertions (position range 1885–1956), frameshifting, and a premature stop at its 3′ end (Sea81-4); 3) a tprGJ chimera, in which the 3′ end of tprG has been replaced by the corresponding region of tprJ as evidenced by the presence of a tprJ-specific signature (TAACGGGAACCCTCTCCCTTCCGGCGGTTCCTCAGGGCACATTGGCCT) near the 3′ end of tprJ (Mexico A and all T. p. pertenue strains); 4) a tprGI chimera in which the 5′ end of the ORF is homologous to the corresponding region of tprG, and its central and 3′ regions of the gene are homologous to the corresponding regions of tprI (all T. p. endemicum strains); and 5) a truncated tprGI chimera due to a single nucleotide insertion (T. paraluiscuniculi and the Fribourg-Blanc strain).
While four T. p. pallidum strains have the reference Nichols tprJ sequence, the T. p. pallidum Sea 81-4 strain and all non-pallidum treponemes studied to date contain a tprGJ chimera in the tprJ locus (Figure 1 and Figure 4). The rabbit pathogen, however, contains a tprGJ chimera that codes for a truncated protein due to an insertion in its 5′ end [36].
Subfamily III tprs show a reduced degree of homology among family members, compared to Subfamilies I and II, with only small regions of sequence identity scattered throughout the coding sequences [27], [33]. This is contrasted by a lower level of sequence heterogeneity at each locus among strains, subspecies, and species. tprB shows no variation among all strains (Supplemental Figure 1.2 in Figure S1). Among strains and subspecies, the tprH locus also contains highly homologous sequences, with only a few point mutations, of which 3 SNPs consistently distinguish syphilis vs. non-syphilis organisms (Supplemental Figure 1.7 in Figure S1).
In the tprA locus, at positions 706 to 711, there is a short region containing either three or four CT dinucleotide repeats. Strains containing only three CT repeats carry a gene that codes for a truncated protein due to a frameshift leading to a premature stop (Nichols, Mexico A, Street 14 and Bal 3). In contrast, strains carrying tprA genes with four CT repeats (the syphilis Sea 81-4 and all non-syphilis isolates) have no predicted frameshift and generate a sequence encoding a full length TprA product (Figure 5).
tprL (tp1031) shows major changes among strains and subspecies. Re-analysis of this region in the all of the endemicum and pallidum strains and 8 additional syphilis strains (Brinck Reid et al., unpublished) revealed a larger putative tprL ORF coding for a protein sequence of 602 amino acids, compared to 514 amino acids as previously reported for the Nichols and Street 14 strains [27], [28]. In this extended ORF (Figure 6), an alternative start codon (CTG) was identified with a typical ribosomal binding site (RBS, GGAGG). Furthermore, beginning at position −31, a 15 to 17 nucleotide poly-G tract flanked by −10 and −35 σ70 signatures (TAGACA and TGTTGT) is evident (Figure 6). Unlike the TprL product annotated in the Nichols genome sequence, the extended TprL is predicted to have a putative OM localization, with a predicted cleavable signal peptide (cleavage between positions 25 and 26, VFS-EQ). Compared to T. p. pallidum and T. p. endemicum sequences, our analysis revealed a gene fusion in the T. p. pertenue and Fribourg-Blanc strains caused by a deletion of 278 nucleotides (Figure 6), encompassing the 5′ end and central regions of the tp1030 ORF and a small fragment of the 5′ end of tprL including its start codon. This deletion creates a hybrid sequence (tp1030 and tprL, here called tprL1) of 1668 bp with the start codon (ATG) in the plus strand of tp1030 (the tp1030 coding sequence is located on the minus strand of the chromosome) in frame with the rest of tprL (tp1031). As a consequence, the first 130 nucleotides of this new pertenue tprL1 (Figure 1, Figure 6 and Supplemental Figure 1.9 in Figure S1) are unique, not found in T. p. pallidum or endemicum tprL. The new extended TprL (in T. p. pallidum and T. p. endemicum and T. paraluiscuniculi) and the newly predicted TprL1 proteins (T. p. pertenue and the Fribourg-Blanc treponeme) are 602 and 556 amino acids long, respectively. Because the first 44 amino acids of TprL1 are encoded by the plus strand, this region is unique to the yaws and simian strains, with no homologous peptide in the pallidum and endemicum proteins (Supplemental Fig. 1.9 Figure S1). This unique peptide sequence is also not found elsewhere in the chromosome. Unlike the newly predicted extended TprL, TprL1 does not have a predicted signal peptide (Figure 6). This raises the possibility that the pallidum and endemicum subspecies may have an OM-localized TprL, while this would be predicted to be absent in the pertenue subspecies.
The 12 treponemal isolates from the three T. pallidum subspecies (pallidum, pertenue and endemicum), the Fribourg-Blanc treponeme, and T. paraluiscuniculi show pleomorphic genetic changes in the tpr family characterized by SNPs, indels, chimeric sequences, and even absence of entire ORFs. Initial comparisons of the currently available full genome sequences of the Nichols, Chicago C, Sea81-4 and Street 14 syphilis strains revealed a high degree of sequence identity and a remarkable conservation of their genome organization [30] (and Giacani et al., unpublished). The study by Mikalova et al. [29] confirmed these observations, reporting clustering of sequence divergence in only a handful of distinct genomic regions among syphilis and non- syphilis strains, similar to those identified previously by Weinstock and colleagues [69]. Many of the hot spots of diversity are located in genes encoding members of the Tpr antigen family. The present study, however, provides a detailed description of sequence diversity within this paralog family and uncovers a rich number of sequence modifications among species, subspecies and strains. Importantly, our analyses also indicate some alternative genes or modified loci.
It is striking that much of the sequence diversity identified in the tpr genes segregates the strains into the same subspecies and species groups that were originally defined according to their modes of transmission, their natural hosts, and the diseases they cause. This is most effectively seen in the colored blocks in Figure 1. Given that the tpr loci represent the primary regions comprising the extremely low genomic diversity among the T. pallidum subspecies, it is likely that the proteins encoded by these variant genes play a major role in the differing pathogenesis of syphilis vs. yaws vs. endemic syphilis. Assigning a definitive role for individual proteins or combinations of proteins in determining clinical outcomes, however, awaits the determination of the functions of the Tpr proteins and the ability to genetically manipulate these genes within the organism. To inform studies of possible location and function, computational and immunological studies can provide clues for individual gene products.
Several arguments emphasize a key role for TprC and TprD during syphilis infection: 1) they are the targets of strong antibody and cellular immune responses [35], [37], [40], [56]; 2) immunization with recombinant TprC/D induces partial protection against infectious challenge [37]; 3) their surface exposure is supported by opsonophagocytosis assays [68] (Lukehart et al., unpublished); 4) TprC and D show sequence diversity among strains [37] (and this study); and 5) 3D models predict a typical β-barrel structure with surface-exposed loops that contain each of the regions where sequence diversity is localized (this study). It is highly unlikely that the co-location of sequence diversity and predicted surface-exposed loops is coincidental. A recent study by Anand et al. [68] proposes an alternative model for TprC and TprF, suggesting that the amino terminus of these two proteins is localized in the periplasmic space. However, experimental evidence argues against this model. Recombinant amino terminal TprF/I peptide induces partial protection against homologous challenge in immunization experiments in the rabbit model [37] and elicits opsonizing antibodies upon immunization (Lukehart et al., unpublished), observations supportive of surface exposure. However, the TprC and D sequence diversity (localized in the exposed DVR) identified among subspecies in the present study may contribute to the variable degree of cross-protection observed among T. pallidum strains and subspecies in infection-induced immunity. In this context, it is possible that sequence differences in the DVRs of TprC and D could lead to subspecies- or strain-specific surface-exposed epitopes that are critical to opsonic function or other mechanisms of protection. Studies are ongoing to test this hypothesis. A recognized example of functionally important strain-specific epitopes is loop 5 of the OMP P2 protein of non-typeable Haemophilus influenzae, which is associated with elicitation of bactericidal antibodies and protective immunity [70]. An alternative, or complementary, function of variable surface-exposed loops (e.g. DVR) could be that of providing steric hindrance to prevent the immune system from recognizing conserved external loops on the antigen, which are perhaps essential for correct protein structure or function. It is noteworthy that TprC and TprD are each predicted by 3D analysis to contain 4 conserved external loops.
During natural human infection and experimental infection of rabbits [37], [56], antibodies are made against TprC/D and TprD2. In addition to TprC and D, the TprD2 variant is also predicted to have surface exposure [37], [40], [56], and is found in both syphilis and non-syphilis treponemes (Figure 1). The regions unique to TprD2 also contain predicted external loops, thus adding another layer of complexity to the already existing set of predicted loops for TprC and D (not shown). Our structural predictions of TprC/D showing co-localization of external loops with DVRs is strong support for our hypothesis that antigenic differences in surface exposed loops of TprC and D have functional significance in immunity to the T. pallidum subspecies, and may be determinants of cross-immunity among subspecies and strains.
Of interest is the observation that the CDC2 strain maintained in Seattle (originally obtained in 2005 from Rob George and Victoria Pope from the Centers for Disease Control in Atlanta, GA) contains a tprD-like allele while the corresponding sequence reported by Mikalova et al. [29] contains a tprD2 sequence. Re-sequencing of the tprD locus of this strain using our original frozen stocks confirmed that the CDC2 strain indeed contains a tprD-like allele. Also, we have sequenced the tprD locus of the pertenue CDC1 strain, isolated in a neighboring village in Africa from where the CDC2 strain was obtained, and found that the CDC1 strain also contains a tprD-like gene. It may demand a significant effort to identify the source of discrepancy between our data and that of Mikalova et al., perhaps requiring the analysis of the two CDC2 lineages over the last several years.
In contrast to syphilis treponemes, the tprF and I loci in T. p. pertenue and the Fribourg-Blanc treponemes each contain identical full-length ORFs. Although their coding sequences are identical within each location, tprF and tprI are located in separate tprG-F and tprJ-I operons, respectively, and their expression may be differentially modulated. The number of G residues in a polyG string in their promoters controls phase variation of these operons [46], and the binding of TpCRP (Tp0262) to the promoters was shown to either increase (tprJ) or decrease (tprG) transcription of the operon [71]. The implications of a “double dose” of tprI in the non-pallidum strains might be reflected in the total amount of message made in tissue specific locations or in differential expression over time during infection. Preliminary studies of antibody reactivity in rabbits infected with T. p. pertenue Gauthier strain demonstrate high levels of antibody to TprI, consistent with high (or double) expression of the protein (Lukehart et al., unpublished). The strong resemblance of the TprI/F 3D predictions to the TprC/D structural models, and the co-localization of DVRs and external loops suggest analogous roles at the microbe-host interface.
T. pallidum tprGI chimeras were identified by Giacani et al. [36] in T. paraluiscuniculi and also present in the whole genome sequences later reported by Strouhal et al. [31] and Smajs et al. [32], whose unique sequence composition was also recognized by these authors. Our analysis shows that, in all strains of T. p. pertenue, T. p. endemicum and the Fribourg-Blanc treponeme, the G and J loci are occupied by either tprGJ or tprGI chimeric genes. In contrast, the Nichols reference tprG and tprJ genes are frequently found in syphilis isolates, but not in any pertenue, endemicum or the Fribourg-Blanc strains tested to date. Only the T. p. pallidum Mexico A and Seattle 81-4 strains carry the GJ chimeric gene in the tprG and tprJ loci, respectively. Of interest is the presence of three truncated chimeras encoded by the tpr E, G, and J loci in T. paraluiscuniculi. This, in addition to predicted truncations or absences of Subfamily I Tprs (Figure 1), is perhaps related to the inability of T. paraluiscuniculi to infect humans, although further study is needed to explore this issue more thoroughly.
One might wonder whether the tpr chimeras identified in this study are artifactual, due to “jumping” between highly similar sequences during PCR amplification [72]–[74]. In our study, tpr chimeras are unlikely to be artifacts for two reasons: 1) independent PCR amplifications of treponemal DNA obtained from different strain harvests rendered identical sequences, and 2) published sequences obtained by multiple sequencing approaches also show the same chimeras [31], [32], [36], [75].
With the exception of TprK, little is known about the other members of Subfamily III Tprs (tprA, tprB, tprH, and tprL). TprA, B and L are predicted to be OMPs [40], [41], and Tpr B induces antibodies that promote opsonophagocytosis (Lukehart et al., unpublished). Sequence conservation of tprB and tprH across species, subspecies, and strains suggests a required function for these proteins in the biology of T. pallidum. Nucleotide repeats, whether in regulatory or coding regions, are frequently associated with modulation of gene expression in an ON-OFF manner. The structure of the promoter region of the newly proposed extended tprL ORF is highly reminiscent of modulation of gene expression by single nucleotide repeats in the promoters of porA and opc loci of Neisseria meningitidis [76]–[78]. One could argue that predictions of an extended tprL ORF may lack accuracy because of the assumption of CTG as start codon, an underrepresented start codon in the annotated Nichols T. pallidum genome. However, our predictions are supported by the identification of a typical RBS, as well as −10 and −35 σ70 signatures with intervening homopolymeric G repeats of variable lengths resembling classic bacterial phase variation systems. In tprA, the variable number of CT dinucleotide repeats creates frameshifting and premature termination, dividing strains carrying tprA genes coding for full length product from those encoding predicted truncated products (Figure 5 and Supplemental Figure 1.1 in Figure S1). This is another mechanism for possible phase variation.
Our analysis of the tpr gene sequences is based on an approach of targeted PCR amplification, cloning, and sequencing a number of clones to obtain consensus sequences. The tpr ORF sequences appear to be unchanging within a given strain during infection. However, limited information at the population level invites speculation about the possible presence of genetically distinct subpopulations within isolates. Smajs et al. [28], [79] reported that at least two subpopulations are present within the Nichols strain as defined by a ∼1 Kb deletion in the flanking region of tp0131. Our approach could have overlooked underrepresented variant organisms within isolates and, if intrastrain variation indeed exists, our findings might then reflect amplification of the most predominant subpopulation. Small mutational changes, even SNPs, in coding or non-coding regions can affect transcription, translation, or folding of the protein themselves, of neighboring genes, or those at more distant sites [80]–[83]. This could explain, for example, some of the differences in transcription observed among treponemal strains [40]. On the other hand, the now standard use of template-based assembly of short stretches of sequence generated by newer sequencing technologies can overlook the existence of hybrid genes or missing ORFs, whereas our individual-ORF sequencing approach can clearly identify these variations. Renewed efforts to address all of the above questions may be effectively resolved using next generation approaches such as deep sequencing of targeted regions, single cell isolation, or whole transcriptome sequencing.
How might knowledge of tpr sequence diversity be translated into tools that are relevant to persons who are infected with one of the pathogenic treponemes? The geographical distribution of yaws and syphilis is not as distinct as decades ago, and travel or migration can serve to transport an infection between urban and rural settings, complicating diagnosis. Because of the re-emergence of yaws over the past 20 years [84], etiological differentiation of yaws vs. syphilis infections is desirable, and a practical approach for diagnosis is needed. The overall reported genetic variability between syphilis and yaws treponemes (0.2%) makes these organisms almost genetically indistinguishable, and existing serological tests fail to differentiate the infections. Several small signatures that differentiate the distinct species/subspecies have already been identified in several genes [47]–[49], [51], [52], [85]. The unique sequence composition of TprL described here in pertenue vs. pallidum strains reveals a possible 90 amino acid sequence unique to non-yaws treponemes, which includes a 25 amino acid predicted signal peptide, as well as a 44 amino acid peptide specific to T. p. pertenue. Given that Giacani et al. [40] showed that the tprL ORF is actively transcribed in both syphilis and yaws treponemes during experimental infection, our findings could facilitate the development of targeted serological screening for differentiating these two infections.
Treponemal infections are chronic, yet only a minority of infected persons develops the severe late manifestations of disease. Is it possible that small genetic markers in the infecting could predict clinical outcome? We previously showed that rabbits infected intravenously with the Sea 81-4 strain had higher levels of cerebrospinal fluid (CSF) inflammation, compared to other infecting strains, while animals infected with Bal 7 had more severe skin disease [86]. Our more recent work in humans supports the hypothesis that disease outcome may be related to genetically defined strain types [55]. Subfamily II tprs and the arp genes were first utilized for strain typing purposes by Pillay et al. [50], although they were not able to correlate strain type with clinical outcome. Using an enhanced strain typing system developed by Marra et. al. [55], which includes the targets initially described by Pillay et. al. [50] and the tp0548 gene, we demonstrated that patients infected by 14d/f type strains were significantly more likely to have neurosyphilis [55]. Four of the pallidum strains shown to represent different genotypes in this report (Nichols, Street 14, Mexico A and Sea 81-4) fall into four different molecular types using the enhanced typing system. The correlation supports the possibility that sequence changes in the tpr genes may be related to specific disease manifestations.
It is noteworthy that T. paraluiscuniculi causes a very mild infection in its natural host, compared to syphilis, and is unable to infect humans [2], [12], [87]. One possible explanation for mild natural infection and the failure to infect other hosts is the dearth of functional Tpr proteins in this organism: there are seven truncated Tpr proteins (TprC, D, F, I, E, G and J) in T. paraluiscuniculi. In contrast, all T. pallidum subspecies and the Fribourg-Blanc treponemes, which have fuller Tpr repertoires, can multiply in more than one vertebrate host and can cause infection in humans. The Fribourg-Blanc treponeme, isolated from non-human primates from a yaws-endemic region in Africa [3], [4], resembles very closely the tpr repertoire of yaws strains (10 out of 12 ORFs are of the same type), although it resembles T. p. endemicum at the G locus, implying shared evolutionary pathways, as previously proposed [29], [65], [88], as well as common strategies of interaction between microbes and their host.
Although the clinical outcome of infection is likely dependent upon several factors, including individual host immunity, inoculum size, and route of infection, sequence changes in the tpr genes could determine differences in antigenicity or function, resulting in different adaptive strategies and differences in pathogenicity. While the distribution of tpr gene variants among the 12 isolates studied here appears, in most cases, to be clustered by subspecies, some isolates in the T. p. pallidum group share tpr variants that are otherwise restricted to non-syphilis organisms. For example, Sea 81-4 contains four tpr ORFs present in the endemicum subgroups (Figure 1), and Mexico A contains the tprGJ chimera in the tprG locus. The recent demonstration of syphilis-like genital lesions and purported sexual transmission of a yaws-like treponeme in wild baboons [13] suggests that pathogenicity and mode of transmission may not, however, be completely hard-wired in the genome. The sharing of some tpr variants among individual pallidum strains and the non-pallidum strains confounds the concept of a purely genetic basis for the nature of the disease. These findings again raise the 1960's nature vs. nurture controversy between Hudson and Hackett with regard to the biological or environmental/epidemiological basis for the differing clinical manifestations seen among the treponematoses [89], [90]. Based upon tpr sequencing, there is genetic heterogeneity (five genogroups) within the pallidum subspecies, as well as some overlap among subspecies and species. Rather than having discrete organisms for each treponemal disease, there may in fact be a genetic continuum of the pathogenic Treponema, individual components of which affect pathogenesis in an individual host in concert with social or environmental factors that influence routes of transmission and disease manifestations. Finding the answer to this question will depend upon the ability to genetically manipulate T. pallidum so that the effects of individual genes can be definitively assessed.
|
10.1371/journal.pntd.0004310 | Trichinella spiralis Paramyosin Binds Human Complement C1q and Inhibits Classical Complement Activation | Trichinella spiralis expresses paramyosin (Ts-Pmy) as a defense mechanism. Ts-Pmy is a functional protein with binding activity to human complement C8 and C9 and thus plays a role in evading the attack of the host’s immune system. In the present study, the binding activity of Ts-Pmy to human complement C1q and its ability to inhibit classical complement activation were investigated.
The binding of recombinant and natural Ts-Pmy to human C1q were determined by ELISA, Far Western blotting and immunoprecipitation, respectively. Binding of recombinant Ts-Pmy (rTs-Pmy) to C1q inhibited C1q binding to IgM and consequently inhibited C3 deposition. The lysis of antibody-sensitized erythrocytes (EAs) elicited by the classical complement pathway was also inhibited in the presence of rTs-Pmy. In addition to inhibiting classical complement activation, rTs-Pmy also suppressed C1q binding to THP-1-derived macrophages, thereby reducing C1q-induced macrophages migration.
Our results suggest that T. spiralis paramyosin plays an important role in immune evasion by interfering with complement activation through binding to C1q in addition to C8 and C9.
| Trichinellosis is one of the most important food-borne parasitic zoonoses worldwide. The key factor for Trichinella spiralis to survive in its host is evading from the attacks by the immune defense system. Our previous study revealed that paramyosin from Trichinella spiralis (Ts-Pmy) played a role in evading host immune attacks by binding to human complement C8 and C9. Here, we demonstrated that Ts-Pmy inhibited classical complement activation by binding to human complement C1q. As a result, classical complement pathway-mediated hemolysis was inhibited in the presence of Ts-Pmy. Additionally, Ts-Pmy inhibited C1q binding to THP-1-derived macrophages and C1q-induced macrophages migration. These results suggest that Trichinella spiralis paramyosin is a potential immunomodulator involved in the evasion of the host complement attack by binding to C1q in addition to C8/C9, and therefore is a potent vaccine target against trichinellosis.
| Trichinellosis is a serious zoonotic disease caused by the ingestion of undercooked meat contaminated with the larvae of Trichinella spiralis. Heavy infection can result in death [1]. Recently, trichinellosis has been regarded as an emerging or re-emerging disease in some countries due to improvements in people’s living standards and changes in eating habits [2,3]. To establish parasitism in the host, T. spiralis has evolved sophisticated mechanisms to avoid immune attack from the host. Elucidating the mechanisms developed by the parasite to survive in the host would facilitate the development of strategies to interrupt parasitism and prevent infection.
The complement system is considered to be the first line of defense against invaded pathogens and plays a crucial role in human innate immunity [4]. Many pathogens have evolved diverse strategies to evade host immune attacks and that commonly encounter the complement system first. The human astrovirus coat protein inhibited classical and lectin pathway activation by binding to C1q and mannan binding lectin (MBL) [5,6]. Other pathogenic proteins,such as Pseudomonas aeruginosa alkaline protease and Trypanosoma carassii calreticulin, also interfere with complement activation by binding to complement components [7,8]. Many parasitic helminths release molecules that interfere with the functions of complement and assist in the parasite’s survival in the host [9,10]. One protein that has been well studied for its immunomodulatory effect on the host complement system is paramyosin [11–13]. Paramyosin is a protein dimer that forms thick myofilaments and found exclusively in invertebrates [14]. Recent studies on paramyosin suggested that it was a functional protein involved in helminth infection as well as a structural protein [12–16]. Many helminth parasmyosins have been reported to be capable of directly reacting with human complement C8 or/and C9. Schistosoma mansoni paramyosin protected the parasites against host attack by binding to complement C8 and C9 [12,15]. Clonorchis sinensis paramyosin bound both human collagen and C9 [16]. In our previous study, we have identified that T. spiralis paramyosin (Ts-Pmy) was expressed on the surface of T. spiralis adult and larval worms [13]. Mice immunized with recombinant Ts-Pmy (rTs-Pmy) achieved protective immunity against T. spiralis infection [17], suggesting that it was a good vaccine candidate. Further investigations into its role in the survival of parasites in the host demonstrated its inhibitory effect on the formation of the complement membrane attack complex (MAC) by interacting with complement C8 and C9. As a consequence, the invaded T. spiralis could evade the host complement attack by inhibiting MAC formation [13,18]. The C9 binding site on Ts-Pmy was determined to be located within the C-terminus of the protein (between 866Val and 879Met) [18]. A monoclonal antibody (mAb 9G3) targeting the binding site could block the binding of Ts-Pmy to human C9, resulting in a significant increase in the complement-mediated killing of newborn larvae of the parasite in vitro [18,19].
In addition to targeting C8 and C9 by helminth-expressed paramyosin, it was reported that S. mansoni and Taenia solium produced paramyosin proteins could bind to complement C1q [11]. C1q is the first complement component and initiates the classical activation pathway. To determine whether Ts-Pmy also inhibited classical complement pathway through binding to C1q, except for C8/C9, as a sophisticated strategy to evade host complement attack, the interaction between Ts-Pmy and complement C1q was investigated. In this study, we demonstrated that rTs-Pmy was able to bind to C1q indeed, resulting in the inhibition of classical complement activation. Thus, C1q represented a complement component and pathway targeted by Ts-Pmy in addition to C8 and C9 as a strategy to escape host immune response.
Female BALB/c mice aged 6–8 weeks and free of specific pathogens were obtained from the Laboratory Animal Services Center of Capital Medical University (Beijing, China). The mice were maintained under specific pathogen-free condition with suitable humidity and temperature. All experimental procedures were approved by the Capital Medical University Animal Care and Use Committee (approval number: 2012-X-108) and comply with the NIH Guidelines for the Care and Use of Laboratory Animals.
Normal human serum (NHS) was derived from the blood of 20 healthy human volunteers, aliquoted and frozen at −80°C. All human blood samples were collected according to the protocol approved by the Institutional Review Board (IRB) of Capital Medical University. Human C1q-deficient serum (C1q D) and C3-deficient serum (C3 D) were purchased from Merck (Germany).
T. spiralis (ISS533) was maintained in female ICR mice. Muscle larvae (ML) were recovered from infected mice using a modified pepsin-hydrochloric acid digestion method as previously described [20]. Adult worms were collected from the intestines of infected mice four days following oral larval challenge. Crude adult worm antigens were prepared from homogenized worm extracts based on a previously described protocol [18]. The anti-Ts-Pmy monoclonal antibody (mAb) 9G3 that specifically recognized Ts-Pmy was previously produced [19]. Recombinant Ts-Pmy (rTs-Pmy) with a His-tag at the C-terminus was expressed in baculovirus/insect cells (Invitrogen, USA) and purified by Ni-affinity chromatography (Qiagen, USA). Ts87 (38 kDa) was an excretory-secretory antigen of T. spiralis identified previously [21]. In this study, recombinant Ts87 (rTs87) was used as a non-relevant protein control.
The human leukemia monocytic cell line THP-1 was purchased from China Infrastructure of Cell Line Resource. THP-1 cells were induced into M0 phenotype macrophages by incubating with phorbol-12-myristate-13-acetate (PMA, Sigma, USA) for 48 h and M2 by stimulating with human IL-4 (PeproTech, USA) in RPMI 1640 medium containing 10% FBS at 37°C in 5% CO2 for another 24 h [22].
To evaluate whether the binding of rTs-Pmy to C1q inhibited complement activation, C3 deposition following complement activation were analyzed [23]. Plates were coated with 2 μg/ml of human IgM in 100 μl of coating buffer (100 mM Na2CO3/NaHCO3, pH 9.6) at 4°C overnight. After washing three times with PBST, the plates were blocked with 1 × PBS containing 5% BSA for 2 h at 37°C. Two μg of C1q was pre-incubated with different amounts of rTs-Pmy (0, 2, 4 μg) and BSA (4 μg, as a control) for 2 h at 37°C before adding to the plates coated with IgM (the activator) for 1 h at 37°C. After washing three times with PBST, C1q-deficient serum (C1q D) diluted to 2% in GVBS++ (Veronal-buffered saline containing 1 mM MgCl2, 0.15 mM CaCl2, 0.05% Tween-20, and 0.1% gelatin, pH 7.4) was added as a source of rest complement components to the plates for 1 h at 37°C and then washed with PBST three times. C3 deposition was determined with anti-C3 polyclonal antibody (Abcam, USA; 1:5,000 in 1% BSA/PBS). HRP-conjugated goat anti-rabbit IgG (BD Biosciences, USA) was used as the secondary antibody and OPD (Sigma, USA) was used as the substrate. The absorbance of the supernatants was measured at 450 nm with a MultiskanGO plate reader (Thermo, USA).
To determine the inhibition of classical complement activation-mediated hemolysis by rTs-Pmy, freshly prepared sheep red blood cells (RBC) (109 cells/ml) were sensitized with an anti-sheep RBC antibody (Sigma, USA) at a 1:200 dilution in 1× HBSS++ (Hank’s balanced salt solution containing 1 mM MgCl2, 0.15 mM CaCl2. Thermo, USA) at 37°C for 30 min, then washed with 1× HBSS++. Different amounts of rTs-Pmy (0, 1, 2, 4 μg) were pre-incubated with NHS (5% in 1× HBSS++) for 1 h at 37°C and then added to the antibody-sensitized erythrocytes (EAs) (5×107 cells/well) for 30 min at 37°C. Cold HBSS++ containing 10 mM EDTA was added to stop the reaction. The cells were centrifuged at 3,000 rpm for 10 min. The absorbance of the supernatants was measured at 412 nm with a MultiskanGO plate reader (Thermo, USA). The percent lysis was calculated relative to cells completely lysed in water.
To determine the inhibition of rTs-Pmy on the binding of C1q to IgM, the ELISA assay was performed. Plates were coated with 2 μg/ml of human IgM in 100 μl of coating buffer (100 mM Na2CO3/NaHCO3, pH 9.6) at 4°C overnight. After washing three times with PBST, the plates were blocked with 1 × PBS containing 2% BSA for 2 h at 37°C. One μg of C1q was pre-incubated with different amounts of rTs-Pmy or BSA (0, 2, 3, 4 μg) for 2 h at 37°C, then added to the plates coated with IgM for 1 h at 37°C. After washing three times with PBST, the binding of C1q to IgM was determined with anti-C1q polyclonal antibody (Abcam, USA; 1:3,000 in 1% BSA/PBS). HRP-conjugated goat anti-rabbit IgG (BD Biosciences, USA) was used as the secondary antibody and OPD (Sigma, USA) was used as the substrate. The absorbance of the supernatants was measured at 450 nm with a MultiskanGO plate reader (Thermo, USA).
To evaluate whether rTs-Pmy could inhibit C1q binding to macrophages, THP-1 cells (containing C1q receptors, 2×105 cells/ml) [24,25] were induced into M2 type macrophages with PMA and human IL-4 as previously described [22] because M2 phenotype macrophages play a role in the immune response to helminth infections [26]. The cells were fixed with 4% paraformaldehyde (PFA) for 20 min at room temperature and then washed with PBS. The cells were blocked with goat serum (ZSGB-BIO, China) for 30 min at room temperature before adding C1q (80 μg/ml) that was pre-incubated with rTs-Pmy (80 μg/ml). The incubation was continued at 37°C for 1 h. After washing with PBS, rat anti-C1q mAb (Abcam, USA; 1:100 in PBS) was added; Dylight 488-labeled goat anti-rat IgG (KPL, USA; 1:100 in PBS) was used as the secondary antibody. The control group incubated with rTs-Pmy was detected by anti-Ts-Pmy antibody 9G3. The cell nuclei were stained with DAPI (ZSGB-BIO, China). Images were acquired with an inverted fluorescence microscope (Leica, DM4000B), and the fluorescence intensity of C1q binding to macrophages was measured with high content analysis (Thermo, USA).
The effect of rTs-Pmy on the C1q-induced migration of THP-1-derived macrophages was determined using a Transwell insert with an 8 μm membrane (Corning, USA) [27]. A total of 1×107 THP-1 cells were added into the upper chamber and stimulated with 100 nM PMA for 48 h; then, human IL-4 (20 nM) was added for another 24 h to induce into M2 type macrophages. Human C1q (10 nM) with different amounts of rTs-Pmy (0, 3, 6, or 12 μg) was added into the lower chamber. The incubation was continued at 37°C in 5% CO2 for 24 h to allow the cells to migrate through the membrane. After washing with PBS, the cells on the membrane were fixed with methanol and stained with Giemsa. The cells that remained in the upper surface of the membranes were removed, and the cells that migrated to the bottom surface of the membranes were counted using a phase contrast microscope (Leica, 1X71). A total of 8 randomly selected fields were counted, and the average of each field was calculated using a previously described method [28]. Non-relevant BSA (12 μg) was added as a negative control, and LPS (100 ng/ml) was used as a positive control.
The data were expressed as the means ± standard deviations (S.D). Differences between groups were evaluated with the GraphPad Prism 5 software (San Diego, CA, USA) using one-way ANOVA. p < 0.05 was considered statistically significant.
The binding of rTs-Pmy to human complement C1q was determined by ELISA and Far Western blotting. ELISA results clearly showed that rTs-Pmy bound to human C1q coated plates in a dose dependent manner while BSA coated plates (2 μg/well) showed no any binding to rTs-Pmy (Fig 1A). Wells coated with 0.5 μg of C1q had showed saturate binding with rTs-Pmy. SDS-PAGE results showed that C1q was separated into 3 chains (A, B and C chain) under reducing condition (Fig 1Ba). Interestingly, Far Western blotting demonstrated that only A chain of C1q was bound to rTs-Pmy, as detected by the anti-His antibody (rTs-Pmy contain a 6His-tag at the C-terminus) (Fig 1Bb), no binding was observed to the non-relative control BSA. Vice versa, rTs-Pmy under reducing condition was bound to C1q as detected by the anti-C1q antibody (Fig 1Bc). C1q did not bind to the same amount of rTs87 or BSA. The results confirmed that rTs-Pmy specifically bound to the A chain of human C1q.
The binding of native Ts-Pmy from T. spiralis adult worms to human C1q was investigated by immunoprecipitation and Western blotting (Fig 2). The results clearly demonstrated that C1q bound to native Ts-Pmy from worm extracts and that the binding complex was pulled down by the anti-Ts-Pmy mAb 9G3. No C1q was pulled down by the mAb 9G3 alone, indicating that C1q bound specifically to native Ts-Pmy.
To evaluate whether the binding of rTs-Pmy to C1q inhibits classical complement activation, we analyzed C3 deposition on plates coated with human IgM in the presence of different amounts of rTs-Pmy. The result showed that activation of C1q deficient serum (C1q D) was able to be reconstituted with the addition of C1q to the similar level of NHS by detecting C3 deposition (C1q D+C1q). However, the addition of increasing amounts of rTs-Pmy (0, 2, 4 μg) to C1q decreased C3 deposition in a dose dependent manner and the difference between the doses was significant (Fig 3). BSA (4 μg) had no any inhibitory effect with the C3 deposition similar to the group without any Ts-Pmy added (C1q D+C1q+Ts-Pmy 0 μg). C1q D itself didn’t cause significant C3 deposition without addition of C1q. The result demonstrated that activation of the classical complement pathway was inhibited by the binding of rTs-Pmy to C1q in this study.
To further determine whether rTs-Pmy inhibited classical complement activation, antibody-sensitized sheep erythrocytes (EAs) were incubated with fresh NHS pre-incubated with different amounts of rTs-Pmy. The classical complement-mediated hemolysis results showed that the lysis of EAs was significantly inhibited by the addition of rTs-Pmy in a dose-dependent manner (Fig 4). There was no significant hemolysis in the presence of C1q D or C3 D serum (C1q- or C3-deficient) because the classical pathway could not be activated without C1q or C3. BSA had no inhibitory effect on classical complement activation.
In order to understand how rTs-Pmy is involved in the inhibition of classical complement activation, different amounts of rTs-Pmy were incubated with C1q before adding into IgM coated plate. It was reported that IgM bound to the head region of C1q [29]. Interestingly, our result demonstrated that the binding of human C1q to IgM was inhibited in the presence of rTs-Pmy in a dose dependent manner. There was no inhibitory effect was observed when the same amount of BSA was added (Fig 5). This result implied that IgM and rTs-Pmy bound to the same region of human C1q and pre-incubation with rTs-Pmy blocked the region on C1q that binds to IgM, suggesting the binding site of rTs-Pmy was on the head region of C1q.
To assess whether rTs-Pmy affected C1q binding to THP-1-derived macrophages [25,30], C1q was pre-incubated with rTs-Pmy before adding to THP-1-derived macrophages. Immunofluorescence staining with anti-C1q mAb showed that the fluorescence intensity on macrophage cells was decreased after C1q was incubated with rTs-Pmy (C1q+rTs-Pmy) (Fig 6A). No fluorescence was detected in the PBS and rTs-Pmy alone control group. The quantitative measurement showed that the fluorescence intensity was significantly decreased in C1q with rTs-Pmy group compared with C1q only group (Fig 6B). The results indicated that rTs-Pmy interfered with the binding of C1q to macrophages.
To investigate the effect of rTs-Pmy on C1q-induced chemotaxis of THP-1-derived macrophages, a migration assay using a transwell chamber was performed. Both human C1q and LPS significantly induced the migration of THP-1-derived M2 macrophages through the membrane (Fig 7). After incubating C1q with increasing amounts of rTs-Pmy (0, 3, 6, or 12 μg), the cell migration through the membrane was significantly reduced in a dose-dependent manner (***p<0.001). No obvious inhibition was detected in the group incubated with BSA at high concentration of 12 μg. The result revealed that rTs-Pmy inhibited the chemotaxis of M2 phenotype macrophages towards C1q.
Complement activation is regarded as the initial guardian for pathogen elimination. Due to the fundamental role of the complement system in immune defense, evading complement system attack is a crucial step for the survival of pathogens. Many studies have reported immune evasion strategies developed by pathogens targeting complement. For example, the Staphylococcus complement inhibitor (SCIN) inhibited complement C3 convertases, and Pseudomonas elastase (PaE) inhibited C3 in a proteolytic degradation-dependent manner [31]. Pathogens including bacteria, viruses and parasites seem to share similar strategies to escape the immune attack by complement. However, the mechanisms underlying the evasion from complement attack developed by T. spiralis were not well investigated.
Paramyosin is a structural muscle protein that is expressed only in invertebrates. In addition to forming thick myofilaments, paramyosin is also expressed on the surface of S. mansoni [12], Echinococcus granulosus [32] and T. spiralis [13]. Recent studies revealed that paramyosin expressed on the helminth surface might act as a potential immunomodulatory effector by targeting complement. S. mansoni paramyosin could inhibit complement activation and the immune response by binding to complement C8, C9 [12], C1q and IgG antibody [33]. T. solium paramyosin blocked the activation of C1 by binding to complement C1q [11]. In our previous study, we demonstrated that T. spiralis paramyosin was able to inhibit the formation of MAC by binding to C8 and C9 and therefore protected the parasites from attack by activated complement [13]. In this study, we demonstrated that T. spiralis paramyosin also targeted C1q, which is the initiator of the classical complement activation pathway.
C1 is the first component of the classical pathway and comprised of three subcomponents: C1q, C1r and C1s. C1q is a versatile pattern recognition molecule that can interact with different types of ligands and perform various biological responses in addition to the initiation of the classical complement pathway [34]. Both the complement and non-complement functions of C1q play a crucial role in the host immune response. In the present study, we demonstrated that Ts-Pmy (both the natural protein from adult worms and the recombinant protein) could bind to C1q, more specifically to the A chain of C1q, indicating that Ts-Pmy might interfere with classical complement activation. Subsequent results showed that C3 deposition onto classical pathway activator human IgM were reduced in the presence of rTs-Pmy, confirming that the binding of rTs-Pmy to C1q could inhibit the activation of the classical complement pathway indeed. Further investigation demonstrated that rTs-Pmy inhibited the binding of C1q to IgM, suggesting that Ts-Pmy and IgM share the same binding site on the head region of C1q. Ts-Pmy’s binding on C1q blocks C1q’s binding to IgM or other immune complex, therefore inhibits the complement classical pathway activation. It may reflect one of the mechanisms that parasite-produced paramyosin inhibites the complement classical activation as a strategy to evade the complement-involved immune attack. The inhibition of C1q activation through binding to rTs-Pmy may affect the final formation of MAC, which was directly reflected by reduced hemolysis compared to C1q without rTs-Pmy. However, our previous study showed that Ts-Pmy also bound to C8/C9 except for C1q identified in this study. Therefore the Ts-Pmy induced inhibition of hemolysis may have resulted from the synergetic consequences of inhibiting both C1q and C8/C9 that reduce final MAC formation [13]. The results suggest that parasite-produced molecule(s) such as Ts-Pmy play roles in immunomodulating the host immune system by targeting a number of immune molecules and pathways.
In addition to complement which acts as the first line of innate immune defense, macrophages also play important roles in TH1- and TH2-mediated responses and eliminate pathogens directly or by associating with neutrophils and complement [35]. Macrophages or monocytes express complement receptor 1 (CR1) and other receptors on their surfaces [30,36,37] to interact with complement. It has been reported that C1q can not only directly bind to CR1 to activate macrophages, but also act as a chemokine to induce macrophage migration to inflammatory regions; therefore, C1q may play roles in the process of tissue damage and repair [27] and the elimination of the pathogen by phagocytosis [38]. In this study, we confirmed that C1q was able to bind to the surface of THP-1-derived M2-like macrophages, the addition of rTs-Pmy reduced the binding of C1q on macrophages possibly through blocking C1q binding to the CR1 or other receptors on macrophages. The addition of rTs-Pmy also reduced the C1q induced chemotaxis of THP-1-derived M2-like macrophages through a filter chamber, indicating that the binding of Ts-Pmy to C1q not only inhibited the C1q-initiated classical complement activation cascade but also impaired the C1q-induced migration of macrophages.
Together with our previous study, our results provide strong evidences that T. spiralis produces paramyosin as a potent immunomodulatory protein involved not only in the inhibition of complement activation through binding to C1q and C8/C9, but also in reducing the migration of macrophages to human C1q. Thus, paramyosin plays an important role in the defense against the host innate immune response and the survival of the parasite in the host, making it as a good preventive or therapeutic vaccine target against Trichinella infection. The C1q binding domain on rTs-Pmy and how rTs-Pmy inhibits the functions of C1q and other complement component(s) are under investigation.
|
10.1371/journal.pntd.0006440 | Implications of current therapeutic restrictions for primaquine and tafenoquine in the radical cure of vivax malaria | The 8-aminoquinoline antimalarials, the only drugs which prevent relapse of vivax and ovale malaria (radical cure), cause dose-dependent oxidant haemolysis in individuals with glucose-6-phosphate dehydrogenase (G6PD) deficiency. Patients with <30% and <70% of normal G6PD activity are not given standard regimens of primaquine and tafenoquine, respectively. Both drugs are currently considered contraindicated in pregnant and lactating women.
Quantitative G6PD enzyme activity data from 5198 individuals were used to estimate the proportions of heterozygous females who would be ineligible for treatment at the 30% and 70% activity thresholds, and the relationship with the severity of the deficiency. This was used to construct a simple model relating allele frequency in males to the potential population coverage of tafenoquine and primaquine under current prescribing restrictions.
Independent of G6PD deficiency, the current pregnancy and lactation restrictions will exclude ~13% of females from radical cure treatment. This could be reduced to ~4% if 8-aminoquinolines can be prescribed to women breast-feeding infants older than 1 month. At a 30% activity threshold, approximately 8–19% of G6PD heterozygous women are ineligible for primaquine treatment; at a 70% threshold, 50–70% of heterozygous women and approximately 5% of G6PD wild type individuals are ineligible for tafenoquine treatment. Thus, overall in areas where the G6PDd allele frequency is >10% more than 15% of men and more than 25% of women would be unable to receive tafenoquine. In vivax malaria infected patients these proportions will be lowered by any protective effect against P. vivax conferred by G6PD deficiency.
If tafenoquine is deployed for radical cure, primaquine will still be needed to obtain high population coverage. Better radical cure antimalarial regimens are needed.
| More than half of the malaria outside of Sub-Saharan Africa is caused by the parasite Plasmodium vivax which is characterised by multiple relapses of malaria from parasites which persist in the liver. The only drugs which prevent these relapses (radical cure) are the 8-aminoquinolines primaquine and tafenoquine, and they both cause haemolytic anaemia in G6PD deficiency, the most common enzymopathy of man. Neither can currently be prescribed in pregnancy or lactation. Tafenoquine is given as a single dose regimen and is a significant advance over primaquine (recommended as a 14 day regimen). However, a greater number of individuals, mostly females, will be ineligible for tafenoquine treatment due to a tighter restriction on the minimum G6PD enzyme activity considered safe for use of the drug. Using enzyme activity data from over 5000 individuals, we estimate the proportions ineligible due to G6PD deficiency as a function of the deficient allele prevalence. Adding this to simple estimates of pregnancy and lactation, we estimate the proportions of populations who cannot receive either tafenoquine or primaquine radical cure. For the elimination of vivax malaria in areas with a high prevalence of G6PD deficiency, then if tafenoquine is deployed primaquine will still be needed, so better regimens should be developed.
| Plasmodium vivax is an important cause of malaria outside Sub-Saharan Africa. The WHO estimates that P. vivax comprises 41% of the malaria burden outside of Africa. This translates into 6–11 million cases/year with an estimated 1800–4900 deaths. India, Indonesia and Pakistan account for just over 80% of the global vivax malaria case burden [1]. Relapse frequencies vary by geographical region. South East Asia and Oceania have the highest incidence with relapse rates exceeding 50% [2]. In this context relapse from liver hypnozoites is the main cause of P. vivax malaria illness and asymptomatic carriage [3].
The only currently available treatment to eliminate liver hypnozoites and thus prevent future relapses (‘radical cure’) of vivax or ovale malaria is primaquine, a rapidly eliminated 8-aminoquinoline. Radical curative efficacy depends on the total dose administered [4]. Treatment courses of 14 days are recommended by the WHO but the effectiveness of unsupervised primaquine is often poor [5,6]. Primaquine has one major adverse effect–it causes dose related acute haemolytic anaemia (AHA) in individuals with glucose-6-phosphate dehydrogenase deficiency (G6PDd) [7].
G6PDd is a common inherited X-linked red blood cell disorder prevalent in tropical and subtropical regions, where in some ethnic groups 35% of males are G6PDd [8]. Males are either deficient (hemizygotes) or normal, whereas females can be fully deficient (homozygotes), partially deficient (heterozygotes) or normal. Due to random X-inactivation (Lyonisation), heterozygous females have two red blood cell populations, one with normal G6PD enzyme activity and the other with reduced activity. On average heterozygote females have half the red cell enzyme activity of normal individuals [9]. However, because X-inactivation occurs early in embryogenesis, there is significant variation between individual heterozygous females in the ratio of deficient to normal red cells.
Standard radical cure regimens of daily primaquine (0.25 or 0.5 mg/kg/day x 14d) are not given to patients who test as G6PDd with currently available qualitative rapid diagnostic tests (RDTs). These tests identify subjects with < 30% of normal activity [10–12] and so detect all male hemizygotes and female homozygotes but only some heterozygous females [13]. The majority of heterozygous females have G6PD enzyme activities above 30% but they may still experience AHA when given daily primaquine or single dose tafenoquine [14–16]. Data in this vulnerable group are limited.
The WHO currently recommends eight weekly primaquine [0.75 mg/kg (45 mg in adults)] doses for those with mild G6PD deficiency variants but its safety is uncertain in the more severe G6PDd variants such as the Mediterranean variant in the Middle East and west Asia, and the variants (e.g. Mahidol, Viangchan, Vanua Lava, Canton) prevalent in South East Asia and Oceania [17]. Weekly primaquine in Cambodia, where G6PD Viangchan predominates [11], resulted in one of 18 vivax infected G6PDd patients requiring a blood transfusion [18]. Accordingly, the WHO recommends a careful risk benefit assessment and medical supervision if weekly primaquine is given.
Tafenoquine is a slowly eliminated primaquine analogue (half life ~14d) that will soon be introduced as a single dose regimen to provide radical cure. Tafenoquine also causes dose dependent AHA in G6PDd individuals [16], but because its slow elimination provides a protracted oxidant effect, its use will be restricted to individuals whose G6PD enzyme activity is > 70% of normal. For tafenoquine, this will require the use of a G6PD test that can quantify the G6PD activity, posing a significant challenge for malaria control programmes.
Both primaquine and tafenoquine are contraindicated in pregnancy and during lactation for fear of causing AHA in the foetus or in a G6PDd breast-fed infant. However, recent work has shown that concentrations of primaquine in breast milk are very low and likely to be safe for G6PDd infants outside the neonatal period [19]. Current dosing restrictions curtail the use of primaquine and tafenoquine [12]. We examined the effect of these restrictions on the potential coverage of radical cure with primaquine today and tafenoquine in the future.
All data used in this analysis were from trials with ethical approval where all subjects gave fully informed consent.
G6PDd is X-linked. The polymorphic variants prevalent in areas where malaria is or was prevalent confer varying levels of enzyme deficiency. We assume that the population distribution of the polymorphic G6PDd genotypes conforms to the Hardy-Weinberg equilibrium [20]. Hereafter, `allele frequency’ refers to the allele frequency in hemizygous males. In the studies (all studies except for [10,21,22]) where genotype data were not available, the expected number of homo- and heterozygous females was calculated from the Hardy-Weinberg proportions with deficient allele frequency estimated from the sampled male population.
There are few large data sets of G6PD quantitative activities. Raw and meta-data were collected from nine recent studies (2013–2017) that reported quantitative measurements of G6PD activity. Three of these datasets were from studies in malaria patients (P. vivax and P. falciparum); four were in healthy volunteers; one was in pregnant women; and one was from a mass primaquine treatment study [10,13,21–27]. These nine studies represent a convenience sample of G6PD activity data.
Raw data were available or made available for seven of these studies and for the other two studies, the corresponding authors kindly provided the meta-data. The full extracted meta-data are provided in the supplementary materials. The methods are calibrated to the adjusted population median activity [28].
Total G6PD enzyme activity in heterozygous females is assumed to be normally distributed (data based assumption from [22]). The mean and variance of this distribution is expected to depend on the severity of the deficiency conferred by the hemizygous/homozygous genotypes. As the majority of studies did not have genotype corrected data, a Bayesian hierarchical model was used to estimate the quantiles of the distributions of enzyme activity corresponding 30 and 70% of the population median as a function of two separate categories of severity using a Bayesian beta-binomial model. This categorisation allows for partial correction of the over-dispersion in the data. We defined two categories, which roughly separate out the severity of the deficiencies as follows: Category 1: A- and Mahidol; Category 2: Viangchan, Orissa, and Vanua Lava. These categories were defined by calculating the ratio of the median activity in hemizygous deficient males over the median adjusted normal activity. This visually clustered the genotypes into these two categories. Note that these categories do not correspond to the WHO categories [29].
The Bayesian beta-binomial model was fitted in R using stan [30]. See supplementary materials for full model specification and code.
To estimate the distributions of G6PD activities in hemizygous males and homozygous females (theoretically identical) we pooled all activity data from G6PD normal males (classified by phenotype) and G6PD normal females (classified by genotype to avoid bias from heterozygotes). To adjust for inter-study variability, each G6PD activity was then scaled by 10studymedianactivity so that the pooled data had a median of 10 (arbitrary value).
Currently, neither primaquine nor tafenoquine can be given during pregnancy or breast-feeding and are not recommended in children < 6 months. The incidence of vivax malaria is usually low in young infants so they were excluded from the calculations.
In our main scenario, we assume that the average woman of reproductive age (15–40 y) has three children who survive at least two years and that each child is breast-fed for two years (the minimum recommended by WHO [31]). Fertility rates in Indonesia, Pakistan and India (which comprise more than 80% of annual cases) range from 2.4–3.4 (data.worldbank.org/indicator/SP.DYN.TFRT.IN/). This totals a period of 8.25 years during which women on average cannot take radical curative regimens.
We assume that in countries where P. vivax is endemic, women of reproductive age comprise 40% of the total female population (this corresponds to current Asian populations, see www.populationpyramid.net/asia/2016/).
For both tested thresholds, we have assumed that in a given population the male to female ratio is 1:1 and that the G6PDd allele frequency in hemizygous males can vary between 0 and 25%. For primaquine dosing, qualitative G6PD tests with thresholds of 30% are in use currently. Quantitative point-of-care tests are still being developed. In these calculations, we have assumed a suitable quantitative test is available (i.e. detects accurately at least 70% of normal activity).
For males, it is assumed that the qualitative test has 100% specificity and sensitivity. For the quantitative test with a 70% of population median threshold all deficient males are correctly identified and 5% of normal males are misclassified as deficient (see Results). Thus, the proportion of males who cannot receive either radical cure is equal to the background allele frequency (q) for primaquine, and slightly higher for tafenoquine (0.05 + 0.95q).
The proportions of females who cannot receive radical cure (XPQ for primaquine & XTQ for tafenoquine) under current prescribing restrictions were calculated as follows:
XPQ=ΔdefPQ+0.4×PBF25−(0.4×PBF25)ΔdefPQ,
XTQ=ΔdefTQ+0.4×PBF25−(0.4×PBF25)ΔdefTQ.
Where ΔdefPQ and ΔdefTQ are the proportions of females who are classified as G6PD deficient at enzyme activity thresholds of 30 and 70% of the population median:
ΔdefPQ=q2+2pqQ30%,
ΔdefTQ=q2+2pqQ70%+0.05(1−q).
q: G6PDd allele frequency; p = 1-q; Q30% & Q70%: proportions of heterozygous females classified as G6PD deficient under a 30% & 70% cut-off, respectively; 0.4: the proportion of women between 15–40 years of age; PBF: mean number of years during which 8-aminoquinolines are restricted due to pregnancy or breast-feeding (this is a function of the fertility rate) and 25 is the number of years of reproductive age; 0.05: proportion of wild type (WT) homozygous females misclassified as deficient.
Extracted meta-data from studies included in analysis are available in the supplementary materials. Publicly available datasets used were:
https://doi.org/10.1371/journal.pone.0116143.s001
https://doi.org/10.1371/journal.pone.0169930.s002
https://doi.org/10.1371/journal.pone.0151898.s003.
Seven of the nine studies were in South East Asian populations (dominant G6PDd genotypes: Viangchan, Mahidol and Vanua Lava); one study was from Bangladesh (likely dominant genotypes Orissa, Kalyan-Kerala [32] and Mahidol); and one study was in African Americans (dominant genotype: A-) (Table 1).
These studies recorded quantitative enzyme data for a total of 2803 females and 2395 males. Estimated G6PDd allele frequencies varied from 7–15%. The estimated proportions of heterozygous females with enzyme activity < 30% of the population median varied considerably from 0% (A- variant: estimated 0 out of ~27 expected heterozygotes) to 29% (Vanua Lava, exact numbers not reported). The estimated proportions of heterozygous females with activity < 70% of the population median also varied considerably from 43% (A- variant: estimated ~12 out of ~27 expected heterozygotes) to 85% (Orissa/Mahidol are likely variants: estimated ~113 out of ~133 expected heterozygotes).
Pooled individual enzyme activity data on G6PD WT males (classified by phenotype) and G6PD WT females (only those classified by genotype) were used to estimate the proportion of all G6PD WT individuals who would have G6PD enzyme activity below a 70% threshold. There was substantial variation in this proportion across studies, with estimates varying from 1–20% in males and 2.6% in females, with a mean estimate of 5.6% (Fig 1).
We estimate that between 8% (African A- and Mahidol variants, 90% credible interval (C.I.): 2–19%), and 19% (Orissa, Viangchan and Vanua Lava variants, 90% C.I.: 9–36%) of heterozygous females test as G6PD deficient at a 30% threshold. The same model estimates that between 50% (African A- and Mahidol variants, 90% C.I.: 30–70) and 71% (Orissa, Viangchan and Vanua Lava variants, 90% C.I.: 51–86) of heterozygous females test as G6PD deficient at a 70% threshold. Both models conclude that an increasing level of G6PDd severity is associated with lower mean enzyme activities in heterozygotes [33,34].
To simplify results, all following calculations assume that 10 and 70% of heterozygous females classify as deficient at 30 and 70% thresholds, respectively.
Current prescribing restrictions imply that the proportion of males who cannot receive radical cure is the same as the background G6PDd allele frequency. For females, the relationship is more complex due to restrictions in pregnancy and lactation, and G6PD heterozygosity (Fig 2). Under our assumptions, independent of G6PD considerations, pregnancy and lactation result in ~13% of women ineligible for radical cure (6.5% of total population). If the background allele frequency is 10%, then 10 and 14.5% of males cannot receive radical cure (primaquine and tafenoquine, respectively), and 16 and 25% of females cannot receive radical cure (primaquine and tafenoquine, respectively). This results in 13 and 20% of the total population being excluded, respectively. The breakdown of excluded proportions for tafenoquine radical cure with a background allele frequency of 10% is shown in Fig 3.
Sensitivity to these estimates can be computed simply by adjusting the model parameters to fit a variety of epidemiological contexts via an interactive RShiny app found at:
https://moru.shinyapps.io/8_Aminoquinoline_Coverage/.
Lifting the breast-feeding restrictions for primaquine or tafenoquine would significantly increase potential radical cure coverage. Recent pharmacokinetic studies indicate that very little primaquine is excreted in breast milk, and therefore that primaquine is likely to be safe in breast-feeding mothers with infants older than 28 days [19]. Based on our assumptions [31], 9–10% of women will be breast-feeding at any given time point in the population.
If there were an alternative primaquine regimen which could be prescribed safely to G6PDd individuals, this would reduce substantially the proportion of excluded individuals [35] (Fig 2). All males could be treated and only 13% of females (those breast-feeding and pregnant) would be excluded from treatment. This would achieve 93.5% population coverage irrespective of background allele frequency. This compares with 80% coverage for tafenoquine when the background G6PDd allele frequency is 10% (Fig 4).
The majority of patients with vivax malaria who should receive radical curative regimens currently do not receive them. There are many reasons for this including the widespread unavailability of G6PD deficiency testing, concerns over haemolysis in both identified and unidentified G6PD deficiency, restrictions in pregnancy and lactation, the unavailability of primaquine, poor adherence to prescribed regimens and inertia. There is increasing recognition, however, that if vivax malaria is to be controlled and eliminated then the coverage of radical cure does need to increase.
Tafenoquine provides an excellent solution to the problem of poor adherence to standard 14-day primaquine regimens as it is given in a single dose treatment. However, this comes at the price of increased haemolytic risk. Tafenoquine is eliminated slowly and once taken persists at active concentrations in the blood for several weeks (mean half-life 2 weeks) whereas the rapidly eliminated primaquine can be stopped at the first signs of severe AHA. To mitigate the risk of severe AHA in G6PD heterozygous females, most of whom appear G6PD normal on current qualitative RDTs, biosensors have been developed that give a quantitative G6PD enzyme activity result. The percentage of G6PD activity in an individual with reference to the normal population median can be calculated easily. This will allow tafenoquine to be given safely to all males with G6PD activity > 30% and all females with activity > 70% of normal G6PD activity. The field performance of these newly developed sensors has not been assessed yet at scale so it is unknown how many genotypically normal females will be identified as deficient by these tests. However, even if biosensors are made widely available, and they do prove consistently accurate in operational use, substantial proportions of patients will not receive radical cure.
Because of the higher G6PD safety threshold (70%) a greater proportion of individuals, mainly females, will be excluded from tafenoquine compared to daily-administered primaquine (Fig 4). A policy of tafenoquine only in a vivax control programme could mean more than 20% of all individuals would not be eligible to receive radical cure if the G6PDd prevalence is over 10%. This underscores the need to continue provision of alternative safer regimens of primaquine for G6PD deficient patients if tafenoquine is deployed widely. Although once weekly primaquine provides a potential treatment of G6PDd patients, its safety in areas where more severe variants are prevalent has not been well established. The provision of primaquine could be greatly simplified if a regimen were developed that replaced both daily and weekly primaquine and could be used without the need to test for G6PDd [35].
These calculations are illustrative and dependent on assumptions that may not be applicable widely. Population demographics vary widely, mean duration of breast-feeding can often be longer than two years [36], the epidemiology of vivax malaria varies (vivax malaria is mostly a paediatric disease in high transmission areas like New Guinea island), and there is good evidence that severe G6PDd (Mediterranean variant) protects against symptomatic disease [37]. Thus, the proportions of vivax malaria patients excluded from radical cure treatment would be lower than predicted from the allele frequencies. This would be proportional to the severity of the prevalent G6PDd genotypes. In a very recent evaluation using G6PD prevalence data across 95 P. vivax endemic countries it was estimated that 14.3% of the population would be precluded from primaquine radical cure treatment on safety grounds [12]; in 70% because of G6PD deficiency (in this estimate all heterozygotes were considered excluded), in 12% because of infancy (<6 months), and in 12% because of either pregnancy or lactation (where breast-feeding was for 6 months). Another important consideration is genetic polymorphism in primaquine bioactivation, notably by CYP 2D6. A large number of CYP 2D6 variants have been described which vary from conferring substantial loss of function to gain in function. The *10 variant which confers moderate loss of function is the most common in East Asian populations, reaching an allele prevalence of 43%. Individuals homozygous (or mixed heterozygotes) for loss of function alleles have been reported to have reduced primaquine radical curative efficacy [38,39], and also presumably reduced risk of haemolytic toxicity. Tafenoquine may be less affected by this genetic polymorphism [40].
The current focus of vivax elimination is the administration of radical cure to patients who present with acute disease. However, there is growing evidence that asymptomatic reservoirs of vivax parasitaemia are substantial, most of which are derived from hypnozoites [2,41,42]. If vivax malaria is to be eliminated rapidly, then one approach is to provide focussed mass treatment with 8-aminoquinolines as was done extensively in the past [43]. In that context the protective effect of G6PD deficiency against vivax malaria will not affect these predictions on the proportions of patients who cannot be provided with radical cure.
Tafenoquine may provide substantial operational advantages but it will not obviate the need for primaquine. More work needs to be done to establish the safety or otherwise of alternative primaquine regimens in areas of severe and moderately severe G6PDd variants. High coverage is key to the successful elimination of Plasmodium vivax. Tafenoquine will be a significant advance in the management of vivax malaria providing single dose radical cure but a significant proportion of the population (predominantly females) will be unable to receive it. Safer primaquine regimens are needed for these patients.
|
10.1371/journal.ppat.1007980 | A single phosphoacceptor residue in BGLF3 is essential for transcription of Epstein-Barr virus late genes | Almost one third of herpesvirus proteins are expressed with late kinetics. Many of these late proteins serve crucial structural functions such as formation of virus particles, attachment to host cells and internalization. Recently, we and others identified a group of Epstein-Barr virus early proteins that form a pre-initiation complex (vPIC) dedicated to transcription of late genes. Currently, there is a fundamental gap in understanding the role of post-translational modifications in regulating assembly and function of the complex. Here, we used mass spectrometry to map potential phosphorylation sites in BGLF3, a core component of the vPIC module that connects the BcRF1 viral TATA box binding protein to other components of the complex. We identified threonine 42 (T42) in BGLF3 as a phosphoacceptor residue. T42 is conserved in BGLF3 orthologs encoded by other gamma herpesviruses. Abolishing phosphorylation at T42 markedly reduced expression of vPIC-dependent late genes and disrupted production of new virus particles, but had no effect on early gene expression, viral DNA replication, or expression of vPIC-independent late genes. We complemented failure of BGLF3(T42A) to activate late gene expression by ectopic expression of other components of vPIC. Only BFRF2 and BVLF1 were sufficient to suppress the defect in late gene expression associated with BGLF3(T42A). These results were corroborated by the ability of wild type BGLF3 but not BGLF3(T42A) to form a trimeric complex with BFRF2 and BVLF1. Our findings suggest that phosphorylation of BGLF3 at threonine 42 serves as a new checkpoint for subsequent formation of BFRF2:BGLF3:BVLF1; a trimeric subcomplex essential for transcription of late genes. Our findings provide evidence that post-translational modifications regulate the function of the vPIC nanomachine that initiates synthesis of late transcripts in herpesviruses.
| EBV is an oncogenic virus involved in the development of about 1.5% of human cancers worldwide. EBV infection has latent and lytic forms. Both forms of infection contribute to the oncogenic capacity of the virus. During the lytic cycle, a cascade of temporally regulated events takes place leading to release of new virus particles. A crucial event in the lytic cascade is expression of the class of EBV late genes, which occurs after viral genome amplification. Late genes mainly encode virus structural proteins that are essential for virus transmission. For many years, the mechanisms regulating expression of late genes remained unknown. Recently, a set of proteins that control expression of late genes was discovered. These proteins form a unique viral pre-initiation complex (vPIC), which initiates synthesis of late gene mRNAs. To this day we have yet to fully understand the process by which assembly of vPIC is synchronized to result in a functional transcription machinery. In this report, we demonstrated that BGLF3, a component of vPIC, is modified by phosphorylation during the lytic phase of the viral life cycle. Phosphorylation of BGLF3 is essential for the ability of the protein to interact with two other components of vPIC, BFRF2 and BVLF1. Our results show that formation of the BGLF3, BFRF2 and BVLF1 complex is integral for synthesis of viral structural proteins. This report establishes the importance of post-translational modifications in regulating the function of vPIC in synthesis of herpesvirus structural proteins. Our findings have the potential to promote the discovery of new anti-viral drugs that inhibit assembly and release of oncogenic herpesviruses.
| Lytic infection is intrinsic to the pathogenesis of herpesviruses. Virus particles are synthesized and assembled exclusively during the lytic phase. The lytic phase of oncogenic gamma herpesviruses contributes to tumor development by expanding the population of latently infected cells that possess the potential to become neoplastic. Lytic gene products also encode and upregulate expression of inflammatory cytokines, anti-apoptotic proteins, signaling molecules, and immunoevasins that promote cell proliferation and suppress immune recognition.
Temporal control of expression of lytic viral genes, a common theme among all herpesviruses, can be categorized into pre- and post-replication events. Mechanisms that regulate expression of these two classes of viral genes are quite distinct. Pre-replication genes, referred to as early genes, are regulated in a manner similar to that of cellular genes. Early gene promoters encompass multiple binding sites for transcription factors that facilitate recruitment of the basic transcription machinery. Post-replication genes, referred to as late genes, have unique promoter structures featuring a non-canonical TATA box element (reviewed in [1, 2]). Activation of late promoters is dependent on amplification of the viral genome. The strict dependence of late gene expression on replication of the viral genome represents one of the longstanding conundrums in the biology of DNA viruses.
Major progress in our current understanding of regulation of late gene expression resulted from identifying a group of lytic herpesvirus proteins that function as late gene transcription regulators [3–5]. This group of late gene regulators is conserved among beta and gamma herpesviruses [6–12]. We and others identified seven EBV proteins as essential for expression of late genes. These EBV late gene regulators are: BcRF1 (viral TATT box binding protein, vTBP), BDLF3.5, BDLF4, BFRF2, BGLF3, BGLF4 (viral protein kinase) and BVLF1 [13–18]. The current model suggests that late gene regulators assemble to form a viral pre-initiation complex (vPIC) on late promoters [1, 13]. Using specific siRNAs to all seven late gene regulators combined with RNA-seq of EBV gene transcripts, we demonstrated that a subgroup of late viral genes is transcribed in a manner independent of vPIC [19]. This phenomenon was confirmed by other groups using different approaches including CAGE-seq analyses [20, 21]. Two of these vPIC-independent late genes encode viral immunoevasins, BCRF1 (viral IL10) and BPLF1 (viral deubiquitinase) [19]. This new insight demonstrates the presence of distinct mechanisms for expression of EBV late structural proteins (vPIC-dependent) versus expression of late viral immunoevasins (vPIC-independent). The mechanism by which vPIC-independent late genes (viral IL10 and viral deubiquitinase) are transcribed is yet to be characterized.
While expression of all components of vPIC occurs during the early phase of the lytic cycle, transcription of late genes is nonetheless dependent on viral DNA replication. Recent reports demonstrated that late transcripts are synthesized from newly replicated viral genomes and require continuous genome amplification [17, 22]. Relatively little is known about the exact function of the various components of vPIC in transcription of late genes. Several late gene regulators have no identifiable domains or cellular homologs (e.g. BDLF4, BDLF3.5, BGLF3, BFRF2, and BVLF1). BcRF1, a viral protein predicted to have a saddle-like structure that is characteristic of the cellular TATA-box binding protein (TBP) [10], selectively recognizes late promoters by binding to a non-canonical TATA box element (TATT) [14, 23]. To understand the role of individual proteins in transcription of late genes, several protein interactions were identified between components of vPIC and subunits of RNAPII. BcRF1 and its orthologs of vTBPs in beta and gamma herpesviruses interact with several subunits of RNA polymerase II (RNAP II); a step considered necessary to recruit RNAPII complex to late promoters [4, 13, 24]. Davis et al mapped the motif interacting with RPB1, RNAPII catalytic subunit, to three leucine residues at the N-terminal domain of ORF24 (the ortholog of EBV vTBP BcRF1) [24]. HCMV UL79, the ortholog of EBV BVLF1, also interacts with multiple subunits of RNAPII. These interactions augment the transcriptional activity of RNAPII suggesting a role in transcript elongation [25]. A number of additional protein-protein interactions were mapped among the various components of vPIC that provide insight into the general organization of the whole complex. For example, BVLF1 orthologs in KSHV (ORF18) and CMV (UL79) form crucial interactions with their corresponding BDLF3.5 orthologs [6, 11]. Whether these interactions are necessary for the role of BVLF1 orthologs in promoting the elongation activity of RNAPII is yet to be determined. In addition, the KSHV ortholog of BGLF3 (ORF34) serves as a core component; the protein physically interacts with four other members of vPIC and is thought to serve as a bridge between vTBPs and the rest of the complex [12, 26]. Mutations that disrupt interaction of KSHV TBP (ORF24) with the KSHV ortholog of BGLF3 (ORF34) abolished synthesis of late transcripts [26].
Despite the significant progress made towards comprehending the organization of vPIC, the role of post-translational modifications leading to assembly, regulation, and function of vPIC need to be addressed to gain better understanding of the dynamics of the complex. Here, we asked whether phosphorylation regulates the function of vPIC in transcription of late genes. Phosphorylation regulates many primary biological processes in eukaryotic cells, such as cell division, DNA replication, transcription, differentiation, and apoptosis. To address the role of phosphorylation in regulating late gene expression, we studied the phosphorylation state of BGLF3. We found that BGLF3 is phosphorylated in vivo at threonine 42. Phosphorylation of BGLF3 is essential for transcription of vPIC-dependent late genes. Our findings indicate that phosphorylation of BGLF3 regulates the capacity of the protein to form a trimeric complex with two other late gene regulators, BFRF2 and BVLF1. Formation of this trimeric complex is crucial for expression of late genes.
BGLF3 and its herpesvirus orthologs are indispensable for transcription of late viral genes encoding structural proteins [18, 19]. The exact role of BGLF3 in the process of late gene expression remains largely unknown. The KSHV ortholog of BGLF3 was shown to interact with individual components of vPIC [12]. These interactions led to the hypothesis that BGLF3 functions as a core protein that connects various components of vPIC. How BGLF3 accommodates the formation of vPIC and whether post-translational modifications regulate the capacity of the protein to mediate one or more of these interactions are questions that were not addressed previously. In this report, we first assessed whether BGLF3 is phosphorylated in vivo during the late phase of the EBV lytic cycle. We immunoprecipitated FLAG-tagged BGLF3 from 2089 cells co-expressing the lytic cycle activator, ZEBRA. A fraction of the immunoprecipitated BGLF3 protein was resolved on SDS-PAGE and stained with colloidal Coomassie blue stain. A distinct protein band with a molecular weight equivalent to that of BGLF3 was detected (Fig 1A). The remainder of the BGLF3 eluate was subjected to trypsin digestion followed by phosphopeptide-enrichment using TiO2 resins. Bound phospho-peptides were eluted and analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS). Fig 1B shows a representative MS/MS spectrum of the single BGLF3 phospho-peptide (amino acids 39 to 45) that was reproducibly shown to be phosphorylated. Phosphorylation of this peptide was evident by the detection of fragment y5 with and without a phosphate moiety (-98 Da). Moreover, other peptide fragments in which the phosphate group was either present or lost were also detected, y6 and y4, respectively. Phosphorylation of BGLF3 was reproduced seven independent times; the 39–45 peptide was consistently identified with Mascot score greater than identity score; the expectation values of the detected 39–45 BGLF3 peptide were as follows: 0.047, 0.018, 0.017, 0.0028, 0.0028, 0.0029, and 0.0029 (Fig 1C). Sequence alignment of the motif encompassing threonine at position 42 revealed that this region is conserved in other gamma herpesviruses (Fig 1D). Based on the fragmentation pattern and the fact that threonine 42 is the only amino acid susceptible to phosphorylation in this BGLF3 peptide, we conclude that BGLF3 is phosphorylated at threonine 42 during EBV lytic infection.
We examined the functional importance of phosphorylation of BGLF3 at threonine 42 on different stages of the EBV lytic cycle, particularly late gene expression. The experiment was performed in 2089 cells induced into the lytic phase by ectopic expression of ZEBRA. In accordance with our previous data [18], knockdown of endogenous BGLF3 using specific siRNA (siBGLF3) markedly reduced expression of BFRF3 late protein, a component of the viral capsid protein (Fig 2A). To demonstrate that the effect of siBGLF3 on expression of BFRF3 was specific to silencing BGLF3 rather than an off-target activity, we inserted silent mutations to generate a form of BGLF3 that is resistant to the siRNA, referred to as rBGLF3. Ectopic expression of rBGLF3 suppressed the effect of siBGLF3 on synthesis of late products and restored expression of the late BFRF3 protein. Expressing a mutant form of rBGFL3 in which threonine 42 was mutated to alanine, rBGLF3(T42A), failed to suppress the effect of siBGLF3 on synthesis of the late BFRF3 protein. Neither knockdown of BGLF3 nor mutation of T42 had any significant effect on expression of the BMRF1 early protein, a component of the viral DNA polymerase complex (Fig 2A).
To determine whether phosphorylation of BGLF3 at threonine 42 is important for expression of late genes in physiologically relevant cell lines, we assessed the effect of the T42A mutation in HH415-16 and SNU-719 cells, which are derived from naturally EBV infected Burkitt lymphoma and gastric carcinoma, respectively. HH415-16 cells and SNU-719 cells were induced into the lytic cycle by expression of ZEBRA. Expression of endogenous BGLF3 was silenced using siBGLF3. We found that synthesis of the late BFRF3 protein was markedly reduced when threonine 42 was substituted with alanine in rBGLF3 (Fig 2B and 2C, compare lanes 4 and 5). Our results demonstrate that abolishing phosphoacceptor threonine 42 in BGLF3 is deleterious for expression of the late BFRF3 viral capsid protein in three different cell lines.
One approach that is commonly used to study constitutive protein phosphorylation is to mutate the phosphorylated site to a phosphomimetic residue, aspartate or glutamate. While phosphmimetic substitutions contributed to the understanding of the role of phosphorylation in many proteins, it often fails to mimic phosphorylation events that are regulated and not constitutive. To determine whether a phosphomimetic mutation would substitute for the presence of phospho-T42, we mutated T42 in rBGLF3 to aspartate (T42D) and glutamate (T42E) residues. As shown previously, ectopic expression of wild type rBGLF3 suppressed the effect of siBGLF3 and restored expression of the late BFRF3 protein. However, neither the aspartate nor glutamine substitutions lead to restoration of BFRF3 expression (Fig 3A).
To determine whether the presence of a phosphorylatable residues (e.g. serine) is sufficient to maintain expression of late genes, we mutated T42 to serine. We found that ectopic expression of BGLF3(T42S) suppressed the effect of siBGLF3 and restored expression of BFRF3 (Fig 3B). These results demonstrate that a phosphorylatable residue at position 42 is essential for the capacity of BGLF3 to mediate expression of late genes.
To study the effect of abolishing phosphorylation of BGLF3 at threonine 42 on synthesis of late transcripts, we used RT-qPCR to assess the level of seven EBV lytic transcripts representing three different groups of viral lytic genes: (A) early transcript: BMRF1 (polymerase associated factor); (B) vPIC-independent late transcripts: BCRF1 (viral IL10) and BPLF1(viral deubiquitinase), and (C) vPIC-dependent late transcripts: BFRF3 (capsid), BLLF1 (glycoprotein), BdRF1 (scaffold), and BLRF2 (tegument). We found that expression of all seven lytic genes was up-regulated in samples transfected with the lytic cycle activator, ZEBRA, relative to cells transfected with empty vector (CMV) (Fig 4A, 4B and 4C columns 1 and 2). Co-transfection of siBGLF3 significantly reduced the level of the four vPIC-dependent late transcripts, BFRF3, BLLF1, BdRF1, and BLRF2 but did not affect the level of early or vPIC-independent late transcripts (Fig 4A, 4B and 4C column 3). Expression of rBGLF3 suppressed the effect of siBGLF3 and restored expression of the four BGLF3-dependent late genes (Fig 4A, 4B and 4C column 4). However, alanine substitution of threonine 42 disrupted the capacity of rBGLF3 to support expression of the four BGLF3-dependent late genes (Fig 4A, 4B and 4C column 5).
To determine whether mutating threonine 42 to alanine affects the process of viral DNA amplification, we purified DNA from aliquots of the same cells that were examined in Figs 2A, 4A, 4B and 4C for protein and RNA expression, respectively. We found that expression of ZEBRA increased the level of viral DNA replication by an average of 100-fold relative to empty vector (CMV) (Fig 4D—Intracellular). Knockdown of BGLF3 or mutating the phospho-receptor threonine residue (T42) to alanine did not compromise the extent of viral genome amplification (Fig 4D—Intracellular). However, alanine substitution of phospho-T42 annihilated the capacity of the virus to produce new virus particles, as assessed by detecting the amount of the extracellular viral DNA (Fig 4D—Extracellular). These findings show that phosphorylation of BGLF3 at threonine 42 is essential for the capacity of the protein to support transcription of late genes encoding EBV structural proteins and hence production of new virions but is dispensable for viral DNA replication.
As a hub protein, the function of BGLF3 in transcription of late genes is likely to be influenced by the protein’s capacity to interact with other subunits of the viral pre-initiation complex. A conceivable explanation for the phenotype of BGLF3(T42A) is that lack of phosphorylation at T42 might compromise the ability of the mutant BGLF3 protein to interact with one or more components of vPIC. Previous work studying assembly of various protein complexes demonstrated that point mutations that reduce the affinity of a protein to a complex could be overcome by increasing the concentration of the protein’s respective interactors in the complex [27, 28].
To test the postulate that increasing the concentration of vPIC proteins might suppress the defect in BGLF3(T42A) and restore late gene expression, we eliminated expression of endogenous BGLF3 in 2089 cells using siRNA. Absence of endogenous BGLF3 was complemented by ectopic expression of wild type rBGLF3 or rBGLF3(T42A). Similar to Fig 2, expression of rBGLF3(T42A) failed to support synthesis of the late BFRF3 protein. Co-expression of four late gene regulators (BcRF1, BDLF4, BFRF2, and BVLF1) partially suppressed the phenotype of rBGLF3(T42A) and increased the expression level of BFRF3 to 54% relative to cells transfected with ZEBRA alone (Fig 5 compare lanes 2 and 6). This result suggests that increasing the protein concentration of four late gene regulators can partially suppress the defect in BGLF3(T42A) and restore expression of late genes.
To delineate the contribution of each late gene regulator in restoring expression of EBV structural proteins, we expressed BGLF3(T42A) in 2089 cells together with different mixtures of vPIC components. In each mixture, one of the four late gene regulators was omitted. Cells were harvested after 48 hours and protein lysates were prepared and analyzed by Western blotting to assess the level of the late BFRF3 protein. We found that eliminating BFRF2 or BVLF1 from the mixture of late gene regulators abolished the ability of vPIC to restore expression of the late BFRF3 protein; however, omission of BcRF1 and BDLF4 had no effect (Fig 6).
Since BFRF2 and BVLF1 are the only two proteins necessary for the ability of vPIC to restore synthesis of late products in cells expressing BGLF3(T42A), we asked whether provision of these two proteins was sufficient to suppress the defect in BGLF3(T42A). We transfected 2089 cells with ZEBRA to induce the lytic cycle. Expression of endogenous BGLF3 was knocked down using siBGLF3. Lack of BGLF3 was complemented with the mutant rBGLF3(T42A). We found that co-transfection of BFRF2 and BVLF1 suppressed the phenotype of the T42 mutation and partially restored synthesis of the late BFRF3 protein (Fig 7, lane 2). To assess the specificity of the BFRF2/BVLF1 combination, we studied the capacity of all possible combinations of the four late gene regulators to suppress the phenotype of rBGLF3(T42A). BFRF2/BVLF1 was the most competent combination to restore late gene expression in 2089 cells complemented with rBGLF3(T42A) (Fig 7). Our findings indicated a novel functional interaction between the phosphoacceptor threonine 42 of BGLF3 and the two late gene regulators, BFRF2 and BVLF1.
In summary, lack of phosphorylation at T42 has detrimental effects on synthesis of EBV structural proteins; however, this defect could be partially complemented by increasing the concentration of BFRF2 and BVLF1.
Based on previous protein interaction studies, BGLF3 serves as a bridge connecting BcRF1, the component of vPIC that binds to late promoters, to other subunits of the complex. To determine whether phosphorylation of BGLF3 at threonine 42 mediates the protein’s capacity to interact with BcRF1, we compared interaction of wild type BGLF3 or BGLF3(T42A) with BcRF1. Co-immunoprecipitation was performed using 2089 cells expressing either form of FLAG-tagged BGLF3 in the absence and presence of BcRF1. We found that both BGLF3 and BGLF3(T42A) interacted with BcRF1 to the same extent (S1 Fig lanes 3 and 5). This result suggests that wild type and mutant BGLF3 could be equally recruited to late promoters via their interaction with BcRF1. Furthermore, it corroborates our findings in Figs 6 and 7 demonstrating that ectopic expression of BFRF2 and BVLF1, but not BcRF1, was essential to partially suppress the phenotype of BGLF3(T42A) and restore expression of late genes.
To understand the nature of the functional interaction between phospho-threonine 42 in BGLF3 and the BFRF2 and BVLF1 proteins, we used co-immunoprecipitation to study the potential protein complexes formed by these three proteins. In Fig 8A, we assessed the capacity of BGLF3 and BGLF3(T42A) to interact with BFRF2 and BVLF1 when expressed individually or together in transfected 2089 cells. Neither BFRF2 nor BVLF1 were non-specifically immunoprecipitated in the absence of FLAG-tagged BGLF3 (Fig 8A lane 1). Both wild type BGLF3 and the BGLF3(T42A) mutant interacted with BFRF2 and BVLF1 when provided individually in a pairwise co-immunoprecipitation (Fig 8A lanes 2, 3, 4, and 5). Co-expression of BFRF2 and BVLF1 in the presence of wild type BGLF3 resulted in a trimeric complex (Fig 8A lane 6). However, interestingly, mutation of threonine 42 abolished the interaction between BGLF3 and BVLF1 without affecting the interaction between BGLF3 and BFRF2 (Fig 8A lane 7).
To further confirm formation of a trimeric complex that includes BGLF3, BFRF2, and BVLF1, we performed reciprocal co-immunoprecipitation using FLAG-tagged BVLF1 to pull down BFRF2 alone or together with BGLF3. In the absence of BGLF3, BVLF1 had weak affinity to the BFRF2 protein (Fig 8B). However, interaction of BVLF1 and BFRF2 increased substantially in the presence of wild type BGLF3, around 6-fold relative to no BGLF3 based on two independent experiments (Fig 8E). Mutation of threonine 42 to alanine compromised the ability of BGLF3 to form a stable complex with BFRF2 and BVLF1 (Fig 8D and 8E). Our experiments demonstrating that either BGLF3 or BVLF1 can pull down the other two components of the subcomplex suggest that all three proteins assemble into a trimeric complex. Abolishing phosphorylation of BGLF3 at T42, disrupts formation of this trimeric complex and functionally impairs expression of EBV transcripts encoding structural proteins (Fig 4).
Phosphorylation represents one of the most important post-translational modifications that regulates the activity, interaction, localization, and stability of proteins. Recently, we and others identified a viral pre-initiation complex dedicated to transcription of beta and gamma herpesvirus late genes. Previous studies mapped a number of important protein interactions that contribute to our current understanding of the overall organization of vPIC. However, none of these studies addressed the role of post-translational modifications, particularly phosphorylation, in regulating assembly or function of vPIC in transcription of late genes. Here, we studied phosphorylation of BGLF3, a protein that is indispensable for transcription of late genes and serves as a core protein connecting BcRF1, the vTBP, to other components of the vPIC module. We report the following novel findings: 1) BGLF3 is phosphorylated in vivo at threonine 42, a site conserved in other gamma herpesviruses, during the late phase of lytic infection. 2) Phosphorylation of BGLF3 at T42 is crucial for the role of the protein in transcription of late genes. 3) BFRF2 and BVLF1 suppress the defect observed in transcription of late genes due to lack of phosphorylation of BGLF3 at threonine 42. 4) As a result of phosphorylation of BGLF3 at threonine 42, a novel trimeric subcomplex forms that includes BFRF2, BGLF3 and BVLF1. 5) Phosphorylation of BGLF3 does not regulate its interaction with BcRF1 (S1 Fig) suggesting that lack of phosphorylation does not impact recruitment of BGLF3 to late promoters but abolishes binding of BVLF1 to BGLF3 and hence formation of a functional vPIC. Our findings demonstrate that phosphorylation at T42 is crucial for BGLF3 to interact functionally and physically with BFRF2 and BVLF1 (Fig 9). These findings demonstrate that post-translational modifications regulate the function of vPIC in transcription of late genes.
Much of the current understanding of the overall organization of vPIC is derived from previous studies using pairwise co-immunoprecipitation experiments. These studies suggest that components of vPIC are involved in an intricate network of protein-protein interactions that form a functional pre-initiation complex. While these studies present a plausible model for the overall structure of the complex, several questions remain unanswered. For instance, what are the dynamics of vPIC assembly? Do these mapped protein interactions take place simultaneously to generate one main complex, as previously proposed, or does assembly of vPIC occur in a dynamic stepwise manner that involves formation of various subcomplexes of late gene regulators? Formation of these subcomplexes might be strictly regulated by certain post-translational modifications to synchronize the proper assembly of vPIC and the timing at which a particular late gene regulator is added or removed from the complex. It is also conceivable that a late gene regulator, such as BGLF3, might be involved in more than one subcomplex. This interpretation might explain how BGLF3 accommodates multiple interactions previously reported using pairwise coimmunoprecipitation. Furthermore, the time at which a particular late gene regulator is added to the complex is also significant. The BVLF1 protein, for example, might be recruited to the complex at a later time point. UL79, the CMV ortholog of EBV BVLF1, was previously shown to promote the transcriptional elongation activity of RNAPII [25]. This result posits the question of whether BVLF1 is part of the viral pre-initiation complex during promoter recognition or the protein is recruited to the complex as RNAPII exits the promoter.
Our data demonstrate that BGLF3 is phosphorylated at threonine 42 (Fig 1). Mutation of this phopshoacceptor residue abolished expression of the late BFRF3 protein and markedly reduced transcription of several late transcripts (Figs 2 and 4). The effect of mutating T42 to alanine was selective to vPIC-dependent late genes encoding structural proteins; expression of early genes or vPIC-independent late genes was not affected (Fig 4). As a core protein, BGLF3 coordinates multiple interactions within vPIC. Failure to synthesize late transcripts suggested a defect in the capacity of the BGLF3 mutant to establish a specific interaction with one or more late gene regulators that form vPIC. Previous reports demonstrated that point mutations that reduce the affinity of a protein to a particular complex could be suppressed by increasing the concentration of the protein’s respective partners in the complex [27, 28]. Following a similar approach, we managed to partially suppress the phenotype of BGLF3(T42A) and restore expression of late genes by increasing the protein concentration of two specific components of vPIC, BFRF2, and BVLF1 (Figs 6 and 7). The combined effect of BFRF2 and BVLF1 was not reproduced when other combinations of late gene regulators were ectopically expressed (Fig 7). These findings led to the hypothesis that phospho-threonine 42 is likely to mediate or regulate interaction of BGLF3 with BFRF2 and/or BVLF1. Indeed, co-immunoprecipitation experiments demonstrated that BFRF2, BVLF1, and the phosphorylated form of BGLF3 interact together. The ability of BGLF3 and BVLF1 to co-precipitate the two other proteins suggests that BFRF2, BGLF3, and BVLF1 form a stable trimeric subcomplex of that is essential for the assembly of a functional vPIC (Fig 8). Phosphorylation of BGLF3 at threonine 42 is likely to augment the affinity of BVLF1 to the BFRF2:BGLF3 subcomplex. In our experiments, removal of phospho-threonine 42 weakens binding, while higher levels of the BVLF1 and BFRF2 proteins partially restore interaction with BGLF3(T42A) by slightly shifting the equilibrium towards complex formation (Fig 8C). Our data suggest that formation of this trimeric subcomplex is essential for transcription of late genes. Alanine mutation of BGLF3 at T42 disrupted complex formation (Fig 8) and markedly reduced expression of vPIC-dependent late genes (Figs 2 and 4). Furthermore, failure of BFRF2 and BVLF1 to fully restore expression of late genes (Fig 7) is corroborated by the reduced ability of the BGLF3(T42A):BFRF2 subcomplex to interact with BVLF1 (Fig 8C). Collectively, our approach involving mutation of BGLF3 at threonine 42 and suppression of the BGLF3(T42A) phenotype strongly correlates with our protein interaction studies to demonstrate the importance of the BFRF2:BGLF3:BVLF1 trimeric subcomplex in transcription of late genes (Fig 9).
In our efforts to understand the phenotype of BGLF3(T42A) we compared the ability of wild type and mutant BGLF3 proteins to interact with BVLF1. Using pairwise immunoprecipitation experiments, both wild type BGLF3 and BGLF3(T42A) were equally competent to interact with BVLF1 (Fig 8A lanes 3 and 5). Addition of BFRF2 to the complex revealed a substantial defect in the ability of BGLF3(T42A) to bind to BVLF1. One possible explanation of this outcome is that association of BFRF2 with BGLF3 results in a new interface that accommodates interaction with BVLF1 in a manner dependent on phosphorylation of BGLF3 at threonine 42. BFRF2 and BVLF1 might form a pocket that fits the motif encompassing phospho-T42 in the BGLF3 protein. Our data suggests that phosphorylation regulates formation of this trimeric complex but is not involved in recruitment of BGLF3 to late promoters; both wild type BGLF3 and BGLF3(T42A) are capable of interacting with BcRF1, the vTBP-like protein that recognizes late promoters (S1 Fig). Therefore, lack of phosphorylation at T42 in BGLF3 is likely to impede subsequent binding of other subunits to form a functional pre-initiation complex.
One approach that is frequently used to study a constitutively phosphorylated site is to mutate this phosphoacceptor residue into a phosphomimetic one. In Fig 3A, we mutated T42 to aspartate, and glutamate residues. None of the phosphomimetic mutations restored expression of late genes; however, mutating T42 to a different phosphorylatable residue, serine, maintained expression of late genes (Fig 3B). This outcome is not unexpected considering a phosphoamino acid has unique chemical characteristics relative to other amino acids including aspartate and glutamate. A phosphate group in a phosphoamino acid has a bigger hydrated shell and more negative charge relative to a carboxyl group in a phosphomimetic residue [29]. Furthermore, a phosphate group forms stronger and more stable hydrogen bonds and salt bridges in protein-protein interactions relative to a carboxyl group [30]. Signal transducing adaptor proteins, such as 14-3-3 protein and proteins containing FHA- or SH2-domains, are phospho-binding proteins that are incapable of recognizing a phosphomimetic replacement [31–33]. Studying the structure of these interactions revealed that phosphomimetic residues do not fit in the binding pocket of adaptor proteins [31, 34, 35]. Our results suggest that transcription of late genes is dependent on the presence of a phosphate group at position 42.
A phosphomimetic mutation is also less likely to substitute for the presence of a phosphate group if both phosphorylated and non-phosphorylated forms play distinct roles in the function of the protein. It is conceivable that both phosphorylated and non-phosphorylated forms of BGLF3 play separate roles in transcription of late genes. BGLF3 is capable of interacting with BVLF1 and BFRF2 individually, such interaction might occur at a specific stage of transcription of late genes that differs from that requiring formation of the trimeric complex.
An alternative interpretation of our data is that BGLF3 has two separate motifs for interaction with BVLF1. One motif binds to nonmodified BVLF1 and a second BGLF3 motif that binds to modified BVLF1. In the absence of BFRF2, BGLF3 favors interaction with the modified form of BVLF1. However, in the presence of BFRF2, BGLF3 interacts with the nonmodified form of BVLF1. Interaction of BGLF3 with the nonmodified form of BVLF1 is dependent on phosphorylation of BGLF3 at threonine 42. Changes in modification of BLVF1 and its impact on formation of various subcomplexes might represent different stages during the process of transcription of late genes. We are currently studying the possibility that BVLF1 is modified and assessing its role in transcription of late genes.
In conclusion, our results demonstrate the essential role protein phosphorylation plays in regulating the function of vPIC during transcription of late genes. Lack of a single phosphorylation site in BGLF3 abolishes expression of late structural proteins and prevents virus release. Phosphorylation of late gene regulators might serve as checkpoints to ensure the precise timing for assembly of vPIC subcomplexes during synthesis of late products (Fig 9). Identifying additional post-translational modifications that are indispensable for expression of late genes and the responsible modifying enzymes has the potential to inform the generation of a novel class of drugs against EBV and its associated diseases.
The ZEBRA protein expression vector was constructed as previously described [36]. The constructs expressing BGLF3, BcRF1, BFRF2, BVLF1, and BDLF4 were cloned into the eukaryotic pCMV6-Entry vectors using the SfgI and MluI restriction sites. The mutants BGLF3(T42A), BGLF3(T42D), BGLF3(T42E), and BGLF3(T42S) were generated by introduction of the indicated point mutations in the BGLF3 sequence using the following mutagenic primers: 5´-CAGTTTAAGCTCGTGGAGGCGCCCCTGAAGTCCTTTC-3’, 5’- CAGTTTAAGCTCGTGGAGGACCCCCTGAAGTCCTTTC-3´, 5’-CAGTTTAAGCTCGTGGAGGAGCCCCTGAAGTCCTTTC-3’, and 5’- AACAGTTTAAGCTCGTGGAGTCGCCCCTGAAGTCCTTTCTG-3’ and their complementary strands, respectively. siRNA-resistant BGLF3 (rBGLF3) was produced by inserting silent mutations in the region of the late gene regulator mRNA that is recognized by the siRNA. These silent mutations disrupt the complementarity between the siRNA and BGLF3 mRNA without affecting the amino acid sequence of the protein. Production of rBGLF3 and experiments establishing specificity of the utilized BGLF3 siRNA were described in details in our previous studies [18, 19]. The following commercial antibodies were used in Western blotting: monoclonal anti-FLAG M2 antibody (Sigma); Anti-Myc-Tag rabbit monoclonal antibody (Cell signaling); anti-β-Actin antibody (Sigma); anti-GAPDH antibody (abcam). Antibodies to BMRF1 (EAD), ZEBRA, and BFRF3 were previously described [37, 38].
2089 cells are human embryonic kidney (HEK) 293 cells stably transfected with a bacmid containing wild-type EBV B95-8 genome [39, 40]. Cells were cultured in Dulbecco’s modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum (FBS) (Gibco), and penicillin-streptomycin at 50 units/ml. Hygromycin B (Calbiochem) 100 μg/ml was added to the medium to select for 293 cells containing the EBV bacmid. The HH514-16 Burkitt lymphoma cell line is a subclone of EBV-infected P3J-HR-1 cell line [41]. SNU-719 cells is a gastric carcinoma cell line derived from a human tumor biopsy naturally infected with EBV [42]. The eukaryotic plasmids were transfected using lipofectamine 2000 (Invitrogen) following the manufacture’s protocol. Transfections were carried out in OPTI-MEM medium (Gibco). Cells were incubated at 37° C in 5% CO2 incubator and harvested 48 h after transfection.
Harvested cells were lysed in sodium dodecyl sulfate (SDS) sample buffer at 106 cell/10ul. After sonication, protein lysates were denatured at 100°C for 5 min and resolved on 10% SDS-polyacrylamide gel or 4–15% Criterion TGX Precast Protein Gel (Bio-Rad). Resolved proteins were transferred to a nitrocellulose membrane (Bio-Rad). The membrane was blocked in TBS buffer (50 mM Tris-Cl, pH 7.5 and 150 mM NaCl) supplemented with 5% non-fat milk and 0.1% Tween-20. Nitrocellulose membranes were blotted with specific primary antibodies to cellular and viral proteins. Immunocomplexes were visualized by ECL (GE) or by autoradiography using 125I-protein A (PerkinElmer).
2089 cells were harvested, washed in cold phosphate-buffered saline, and resuspended in lysis buffer (20 mM Tris-HCl, pH 7.5, 150 mM NaCl, 1 mM Na2EDTA, 1 mM EGTA, 1% Triton) containing Halt Protease and Phosphatase inhibitors(ThermoFisher). Lysates were passed through a 25-G needle 9 times then centrifuged at 21,000 x g for 10 min at 4°C. Supernatents were pre-incubated with protein A agarose beads to reduce non-specific interactions. Five percent of each supernatant was stored at -80°C as input sample. The rest of the supernatant was incubated with pre-washed anti-FLAG M2 affinity agarose beads (Sigma) for 2h at 4°C. The beads were washed four times with lysis buffer and once with elution buffer (50 mM HEPES, pH 7.4, 100mM NaCl, 1 mM DTT, 5 mM βglycerophosphate, 0.1 mM Na3VO4, 0.01% Igepal CA630, 10% glycerol). Immunoprecipitated proteins were eluted in elution buffer containing 0.5mg/ml 3X FLAG Peptide (Sigma). Input samples and immunoprecipitated proteins were detected by Western blotting using appropriate antibodies or by protein staining using colloidal Coomassie blue [43].
Sample preparation of the liquid chromatography-tandem mass spectrometry (LC MS/MS) analysis was performed following the previously described protocol [44]. Briefly, immunoprecipitated proteins were subjected to Dithiothreitol (DTT) reduction, Iodoacetamide (IAN)-mediated alkylation followed by trypsin digestion. The digested sample was desalted by Spin Desalting column (Thermo) and acidified with 0.5% Trifluoroacetic acid (TFA), 50% acetonitrile then subjected to titanium dioxide enrichment using the Top Tips system (Glygen Corp). The resulting phosphopeptide-enriched sample, dissolved in 70% formic acid and diluted with 0.1% TFA, was then subjected to LC-MS/MS analysis using the Orbitrap Fusion Mass Spectrometer that is equipped with a Waters nanoACQUITY UPLC system. A Waters Symmetry C18 180 μm x 20 mm trap column and a 1.7 μm, 75 μm x 250 mm nanoACQUITY UPLC column was utilized for online peptide separation. The acquired data was peak picked and searched using the Mascot Distiller and the Mascot search algorithm, respectively. Manual examination of the MS/MS spectra (as shown in Fig 1B) and the corresponding assigned fragment ions were conducted to verify the identified phosphopeptide.
RNA was prepared from cells using the Qia-shredder and the RNeasy Plus products from Qiagen. The concentration of RNA in each sample was determined by measuring the optical density at 260 nm. The level of viral transcripts was assessed from 100 ng of total RNA using iScript One-Step RT-PCR with SYBR Green (Bio-Rad) in a total volume of 25 μl. The level of 18S RNA was measured to normalize for the total amount of RNA. Each sample was analyzed in triplicate; the fold change in expression was calculated using the ΔΔCT formula implemented in the software of the CFX real-time PCR system (Bio-Rad). The efficiency of the primers used in RT-qPCR was determined against 10-fold increasing concentrations of viral DNA. The sequences of the primers are provided in Table 1.
The relative expression levels of target proteins were measured according to the density of bands from Western blot using densitometer machine (GS-800 Calibrated Densitometer, Bio-Rad). Expression of the late BFRF3 protein was calculated relative to the expression level of GAPDH or β-Actin. Statistical analysis for viral transcripts (Fig 4) was performed using paired t test available in GraphPad Prism software (La Jolla, CA, USA). A value of p < 0.05 was considered statistically significant.
|
10.1371/journal.pntd.0003463 | Impact of Helminth Infection during Pregnancy on Cognitive and Motor Functions of One-Year-Old Children | To determine the effect of helminth infection during pregnancy on the cognitive and motor functions of one-year-old children.
Six hundred and thirty five singletons born to pregnant women enrolled before 29 weeks of gestation in a trial comparing two intermittent preventive treatments for malaria were assessed for cognitive and motor functions using the Mullen Scales of Early Learning, in the TOVI study, at twelve months of age in the district of Allada in Benin. Stool samples of pregnant women were collected at recruitment, second antenatal care (ANC) visit (at least one month after recruitment) and just before delivery, and were tested for helminths using the Kato-Katz technique. All pregnant women were administered a total of 600 mg of mebendazole (100 mg two times daily for 3 days) to be taken after the first ANC visit. The intake was not directly observed.
Prevalence of helminth infection was 11.5%, 7.5% and 3.0% at first ANC visit, second ANC visit and at delivery, respectively. Children of mothers who were infected with hookworms at the first ANC visit had 4.9 (95% CI: 1.3–8.6) lower mean gross motor scores compared to those whose mothers were not infected with hookworms at the first ANC visit, in the adjusted model. Helminth infection at least once during pregnancy was associated with infant cognitive and gross motor functions after adjusting for maternal education, gravidity, child sex, family possessions, and quality of the home stimulation.
Helminth infection during pregnancy is associated with poor cognitive and gross motor outcomes in infants. Measures to prevent helminth infection during pregnancy should be reinforced.
| The WHO recommends anthelmintics for pregnant women after their first trimester but the benefits remain unequivocal. Although the consequences of helminth infection during pregnancy on the health of pregnant women have been well studied, the impact on the early development of offspring has been understudied. Studies suggest that helminth infection in children may be associated with poor cognitive development, but very little is known whether a similar consequence exists for offspring of women infected with helminths during pregnancy. From our study, we found that women who had intestinal worm infection at least once during pregnancy, had children with lower cognitive and motor scores. Moreover, hookworm infection in pregnant women prior to receiving anthelminthic treatment was associated with poor gross motor functions of children at one-year of age. Based on the findings of this study, measures to prevent helminth infections during pregnancy should be reinforced.
| Intestinal helminths infect more than two billion of the world’s population, with the highest prevalence in Asia and sub-Saharan Africa.[1] The burden of intestinal helminth infection is estimated to be five million disability-adjusted life years (DALYs).[2] Helminth infections are rarely directly associated with increased mortality but are related to increased morbidity arising from the chronicity and consequences of infection.[3] Although the World Health Organization (WHO) highly recommends anthelmintic therapy for pregnant women in their second trimester[4], the benefits on anemia, congenital anomalies and perinatal mortality remains unequivocal[5]. In sub-Saharan Africa, it is estimated that one-third of pregnant women are infected with soil-transmitted helminths[6] although several studies have shown wide variation in prevalence across different countries, 11.1% in Benin[7], 25.7% in Ghana[8] and 49% in Gabon[9]. In Benin, anthelminthics are a component of the routine antenatal care (ANC) package given to pregnant women after their first trimester.[10]
A recent systematic review found little evidence that deworming in children is associated with better cognitive function, though most trials included were of poor quality.[11] A cross-sectional study revealed that compared to 7 to 18 year-old-children who were not infected with Ascaris lumbricoides and Trichuris trichiura, children who were infected with either of these species of helminth performed poorly on tests of memory and verbal fluency, respectively.[12] Over the past decades, many studies have confirmed helminth infection during pregnancy as a risk factor for maternal iron deficiency (ID) and anemia[3,13,14]. However, evidence remains limited on the effects on adverse birth outcomes such as low birth weight (LBW) [15] which is known to be associated with poorer cognitive function in children.[16] Additionally, ID and anemia during pregnancy may be associated with poor cognitive function of infants as shown in a study in rural China which revealed that children of iron deficiency anemic (IDA) women performed significantly lower than those of non-IDA women in cognitive assessment tests.[17] The rapid rate of development of fetal organs makes them particularly susceptible to prenatal insults that are injurious to fetal development, and which could influence their development persisting even after birth. The early onset of delayed cognitive development could negatively influence several aspects of child development including preparedness for school.[18] Notwithstanding the evidence that helminths are associated with these indirect threats, very little is known about the impact of helminth infection during pregnancy on actual infant cognitive development. A study in Uganda concluded that Mansonella perstans and Strongyloides stercoralis infection during pregnancy may be associated with impaired executive function in children.[19]
The objective of this study was to determine whether maternal infection with helminths, both in general and with specific helminth species, during pregnancy, is associated with cognitive and gross motor functions of one-year-old children in Benin.
Our prospective cohort included singletons born to pregnant women who were enrolled before 29 weeks of gestation in the Malaria in Pregnancy Preventive Alternative Drugs (MiPPAD) clinical trial (NCT00811421) comparing sulfadoxine-pyrimethamine and mefloquine as intermittent preventive treatment of malaria in pregnancy (IPTp). The study was conducted in the district of Allada in Benin. One thousand and five HIV-negative pregnant women attending their first ANC visit in the health centers in each of the three sub districts of Allada (Sekou, Allada and Attogon) were recruited. Detailed inclusion and exclusion criteria in the MiPPAD trial are explained elsewhere.[7]
All live born children of recruited pregnant women who survived to 12 months were invited for neurocognitive assessment in the TOVI study (Fon language: Tovi means Child from the country).
We first described and compared the baseline characteristics of women with singleton live births whose children were assessed and those whose children were not assessed for cognitive function. Secondly, we performed univariate analyses to assess crude associations between the ELC and the gross motor scores with helminth infection, helminth species, helminth density, co-infection with malaria, and covariates [maternal prepregnancy body mass index (BMI), family possessions, maternal occupation, education, the RPM and HOME scores]. These covariates were considered as potential confounding factors as they are known risk factors for poor cognitive development and may share common causes with helminth infection. Next, we conducted a multiple linear regression adjusting for covariates whose p-values were less than 0.20 in the univariate analysis. Finally, we performed stepwise removal of covariates from the model if they were found not be statistically significant. From the final model, we evaluated the adjusted mean difference in ELC and gross motor scores. Infant characteristics at birth or age one-year including birth weight, preterm birth and infant helminth infection were hypothesized to be within the causal pathway (as mediators). All multivariate models were adjusted for infant sex. Although infant characteristics (preterm births, low birth weight, and weight-for-age at MSEL assessment) were hypothesized to be mediators in the association between prenatal helminth infection and infant cognitive function, we adjusted for these variables in a sensitivity analyses.
Statistical analyses were conducted using Stata IC/11.2 for Windows (StataCorp Lp, College station, TX). We used Pearson’s correlation to assess the associations between the dependent variables and other continuous variables. The student t-test, Wilcoxon rank sum test and chi-squared test were used to compare means, medians and proportions, respectively. Statistical significance was defined as p-value less than 0.05.
The study was approved by the institutional review boards of the University of Abomey-Calavi in Benin, New York University and Michigan State University in USA and the Research Institute for Development’s (IRD) Consultative Ethics Committee in France. At recruitment, we obtained written informed consent from all pregnant women and guardians of children who participated in this study in the presence of a witness. Women who could not read and write provided thumbprints to confirm their agreement to participate in the study after a nurse had explained the study.
As shown in Fig. 1, 863 live born singletons were enrolled into the birth cohort but 35 died before the age of one year leaving 828 eligible children. Of these, 635 (76.7%) were assessed for cognition using MSEL at approximately one year of age. The median age during MSEL assessments was 12.1 months (range: 11.3–15.3 months). Two children were not able to complete all of the MSEL subtests, leaving 633 children who were fully assessed. Maternal baseline characteristics were similar between women whose children were fully assessed for cognitive function and those whose children were not, as shown in Table 1. Also there was no significant difference between infant characteristics between children assessed and those not assessed.
At first ANC visit, the prevalence of helminth infection was 11.5% of which hookworm infections were the most prevalent (9.5%). Of the 52 women with hookworm infections at the second ANC visit, 12 were infected with the same species at first ANC (see Table 2 for prevalence and density of helminths). The prevalence (95% CI) of helminth infection among children by age one was 32.8% (26.0%-39.6%).
Maternal education, occupation, family possession, RPM and HOME scores and infant weight-for-age were associated with ELC and gross motor scores. Of note, maternal malarial infection was not statistically significantly associated with ELC and gross motor scores (see Table 3). Infant ELC and gross motor scores increased with increasing prepregnancy BMI class. As shown in Table 3, children born preterm and those with low birth weight had lower ELC and gross motor scores, respectively.
As shown in Table 4, maternal occupation and educational status were associated with helminth infection at second ANC visits. Family possessions score was associated with helminth infection at both ANC visits.
The difference in mean ELC scores between children whose mothers were infected with helminths at first ANC visit and those whose mothers were not infected with any helminth remained significant after adjusting for maternal education, child sex and HOME score (p = 0.013). Pregnant women who were infected with helminths at least, once during pregnancy had children with poorer ELC scores, thus-4.4 (95% CI: -7.2 to-1.5) compared to those of mothers who were never infected during pregnancy after adjustment (see Table 5).
After adjusting for gravidity, maternal education, family possession, child sex and HOME score, helminth infection at first ANC visit was negatively associated with infant gross motor function (p = 0.028). We observed that mothers who were infected with hookworms during the first ANC visit had children who scored less in the gross motor scale, -4.9 (95% CI: -8.6 to-1.1), compared to those whose mothers were never infected with hookworms at first ANC visit. With the exception of the association between gross motor scores and the occurrence of helminth infection over the course of pregnancy, sensitivity analyses performed by further adjusting for infant preterm status and weight-for-age, yielded similar results in the association between infants gross motor function and prenatal helminth infection. Helminth infection at second ANC was no longer statistically significantly associated with infant ELC scores after sensitivity analyses, p = 0.074 (see Table 5). Further adjustment for LBW (not preterm birth) and weight-for-age showed similar conclusions in the sensitivity analyses. We performed multiple regression analysis further adjusting for research nurses and found little difference in the results.
Our study has shown that intestinal helminth infection at first ANC visit is associated with poorer infant cognitive and gross motor functions at the age of one-year after adjusting for other known risk factors of cognitive and gross motor development. In our study population, prenatal hookworm infection was related to lower performance in gross motor tests. Our results also reveal that helminth infection at least once during pregnancy may have negative consequences on the cognitive and motor development of infants.
Our study is one of the few large prospective mother-child cohorts with relatively low attrition rate in francophone Africa[27] and including several assessments during pregnancy. To our knowledge, our study is the first to assess the impact of prenatal helminths on the psychomotor development of infants taking into account data from different stages of pregnancy. In addition, we used a comprehensive assessment for neurodevelopment carried out by research nurses specifically trained by an expert in cognitive assessment in African countries (co-author MJB). An additional strength of this study is the consideration of several potential confounding factors such as socio-economic status, maternal depression and RPM and HOME scores. Malaria has also been assessed several times during pregnancy allowing for the study of the impact of malaria-helminth co-infection on child development. Despite low power due to the low prevalence of co-infection, our results do not suggest a higher impact on child development of helminths when associated temporally with malaria. Maternal demographic and reproductive characteristics were also comparable between children lost to follow-up and those included in the study hence selection bias is unlikely.
Since pregnant women recruited in the trial had adequate antenatal care including at least, two ANC visits with treatment for helminth infection at first ANC visit (apart from emergency visits), our results are likely to underestimate the effect of prenatal helminth on infant cognitive function in the general population that may attend fewer ANC visits and receive fewer treatments. Also, the low sensitivity of the Kato-Katz technique for helminths[28] may have resulted in measurement error but since, in this prospective cohort, the assessment of helminth status was independent of the performance of infant in the MSEL at age 1 year, the misclassification would probably be non-differential of infant cognitive and motor scores thus the association may be biased towards the null. The low prevalence of A. lumbricoides, T. trichiura and S. Mansoni did not permit us to study their independent impact on infant cognitive function. Given that treatment was given to women after their first ANC visit, the number of chronic infections was low in our study. Therefore, the effect of chronicity of untreated prenatal helminth infection on child development could not be evaluated. By definition, helminth infection is chronic until treatment. Women testing positive for helminths may have been chronically infected prior to their first ANC visit. However, testing positive for helminth infection at second ANC visit and/or at delivery after being infected with helminths at first ANC does not specifically indicate chronicity. Instead it could indicate reinfection after being treated following mebendazole administration at first ANC visit. Our study is also limited in the inability to assess the presence of prenatal S. haematobium as urine samples were not examined for eggs of this species. Due to the low proportion of children assessed for helminth infection, we were not able to adjust for infants’ infections in models. Species of helminth in mothers and infants were largely different, yet regardless of the species there was no correlation between helminth infection in mothers and children. This therefore suggests that the association between maternal helminths and child development may be independent of infants’ helminths.
Apart from a cohort study that was nested within the Entebbe Mother and Baby Study in Uganda[19], we did not identify any published study on the impact of prenatal helminths on cognition in offspring. The negative relationship witnessed between maternal helminth infection and infant cognitive development in our study is consistent with the general conclusion in the aforementioned study. Converse to the findings of our studies, the authors found no association between maternal hookworm infection and infant neurocognitive development. One explanation may be that Nampijja et al.[19] excluded pregnant women presenting severe anemia (Hb concentration<80g/L). They also included some maternal and infant characteristics (such as maternal hemoglobin level and birth weight) in their final model. It is important to note that our study population had a low prevalence and a low intensity of helminth infection according to WHO classifications of the community endemic levels[22] and few cases of multiple infections with different helminth species.
The mechanism by which prenatal helminth infection influences infant cognitive function remains unknown. However, helminth infection especially with hookworms is a known risk factor for ID. When hookworms penetrate the intestinal mucosa of a host, they ingest the host’s blood causing intestinal blood loss and erythrocytes lysis[29]. This could result in IDA[30] which may be disadvantageous during pregnancy because of the increased physiological demand for iron. Studies have shown that in very iron deficient mothers, maternal serum ferritin concentration is correlated with that of the neonate[31] while decreased concentration is associated with a decrease in brain iron concentration[32] which could in turn alter hippocampal development of the neonate[33]. A study among one-year-old children found that, those born with inadequate brain iron stores (≤34μg/L cord ferritin) had lower psychomotor function and auditory recognition memory than those with adequate brain iron stores.[34]
Helminths may be associated with several adverse birth outcomes that could mediate the pathway between prenatal helminth infection and infant cognitive development. Although findings from clinical trials reveal no beneficial effect of anthelminthic treatment on LBW and preterm births[35], a large community study of about 5000 pregnancies in Nepal showed an increased risk of LBW and infant mortality among the children of women who did not receive antenatal anthelminthic treatment[36]. Notwithstanding the contradictory effects of prenatal helminth infection on birth outcomes, adverse birth outcomes have been confirmed by some studies to be associated with infant cognitive development.[37–39] Our results, after sensitivity analyses, however suggest that other plausible unmeasured factors could also account for the observed association between prenatal helminth infection and child development.
It is unlikely that increased susceptibility of children of infected mothers to helminth infection explains for the decreased ELC and gross motor scores, as there was no association between prenatal helminth infection and infant helminth infection by age one. Moreover, the pattern of helminth species varied in the mothers compared to the children. T. trichuris was the most prevalent species of helminths in children (20.9%) contrary to high hookworm prevalence in pregnant women.
Mebendazole is a broad-spectrum anthelminthic drug that is effective against several intestinal helminths. However, it has lower cure rates and fecal egg reduction rates for hookworms than albendazole.[40] In our study, although infection by any helminths at second ANC visits was not associated with poor cognitive or gross motor function, hookworm infection remained associated with ELC scores. This could be due to either the re-exposure of pregnant women to hookworms even after mebendazole administration or low cure rates against hookworms. Although we did not monitor the adherence to mebendazole treatment, the decline in parasite density at second ANC visit observed in the majority of pregnant women infected with the same species than at first ANC visit (see S1 Table) suggests good adherence.
This study provides evidence of an association between intestinal helminths and hookworms among pregnant women and poor cognitive and gross motor functions in their children at approximately 12 months of age. In view of these findings and as recommended by the WHO, measures to prevent helminth infections should be reinforced. Further studies are needed to corroborate our findings and explain the pathophysiological mechanisms of this relationship.
|
10.1371/journal.pntd.0004247 | Simple, Rapid Mycobacterium ulcerans Disease Diagnosis from Clinical Samples by Fluorescence of Mycolactone on Thin Layer Chromatography | Mycobacterium ulcerans infection, known as Buruli ulcer, is a disease of the skin and subcutaneous tissues which is an important but neglected tropical disease with its major impact in rural parts of West and Central Africa where facilities for diagnosis and management are poorly developed. We evaluated fluorescent thin layer chromatography (f-TLC) for detection of mycolactone in the laboratory using samples from patients with Buruli ulcer and patients with similar lesions that gave a negative result on PCR for the IS2404 repeat sequence of M. ulcerans
Mycolactone and DNA extracts from fine needle aspiration (FNA), swabs and biopsy specimen were used to determine the sensitivity and specificity of f-TLC when compared with PCR for the IS2404. For 71 IS2404 PCR positive and 28 PCR negative samples the sensitivity was 73.2% and specificity of 85.7% for f-TLC. The sensitivity was similar for swabs (73%), FNAs (75%) and biopsies (70%).
We have shown that mycolactone can be detected from M. ulcerans infected skin tissue by f-TLC technique. The technique is simple, easy to perform and read with minimal costs. In this study it was undertaken by a member of the group from each endemic country. It is a potentially implementable tool at the district level after evaluation in larger field studies.
| Mycobacterium ulcerans infection, known as Buruli ulcer, is a disease that affects the skin and underlying tissues. The organism responsible for the infection produces a potent toxin called mycolactone that causes extensive skin damage. Easy to perform and cheaper techniques are needed for diagnostic confirmation. We have evaluated fluorescent thin layer chromatography (fTLC) for detection of mycolactone in skin samples from patients with Buruli ulcer comparing them with samples from similar non-Buruli ulcer lesions that gave a negative result in the standard polymerase chain reaction (PCR) test for M. ulcerans. Fluorescent TLC had sensitivity of 73.2% and specificity of 85.7% when compared with PCR whether the skin sample was a swab, a biopsy or a fine needle aspirate. This study shows that mycolactone can be detected reliably from M. ulcerans infected skin tissue by the simple, low cost technique of fluorescent thin layer chromatography that could be developed for point of care use. It requires further evaluation in countries where Buruli ulcer disease is endemic.
| Mycobacterium ulcerans infection, known as Buruli ulcer, is a disease of the skin and subcutaneous tissues which is an important but neglected tropical disease with its major impact in rural parts of West and Central Africa where facilities for diagnosis and management are poorly developed [1]. Since prevention is not possible in the absence of either an effective vaccine or a clear understanding of the mode of transmission, a major control strategy for Buruli ulcer is early detection and treatment, hinging on effective laboratory confirmation of suspected cases.
Standard routine laboratory techniques for the confirmation of Buruli ulcer disease are M. ulcerans isolation by culture, histopathology, smear microscopy for acid-fast bacilli (AFB) and polymerase chain reaction (PCR) for detection of the M. ulcerans specific insertion sequence IS2404 [2]. Treatment decisions cannot be made on the basis of culture results because M. ulcerans grows too slowly (over 8–12 weeks) and histopathology is not available in most endemic countries. Microscopy for AFB can be done quickly at low cost but its sensitivity is only 40–60% so PCR for IS2404 which has sensitivity of 92–95% has been established as the gold standard for case confirmation [2] [3] [4]. As PCR is sophisticated and expensive its use has been restricted to a small number of reference laboratories [5] and numerous studies have been carried out to develop a simple diagnostic test that can be used at point of care facilities [6] [7]. One such technique is fluorescent thin layer chromatography (f-TLC) [8] which targets mycolactone, a polyketide-like toxin produced by M.ulcerans. The toxin, which is not produced by any other human pathogen, is responsible for the characteristic necrosis in M.ulcerans infected lesions and it is present in the skin biopsies of mice and humans infected with M. ulcerans [9] [10] [11].
Using liquid chromatography-mass spectrometry (LC-MS) mycolactone has been detected in skin lesions in 77% of patients with untreated Buruli ulcer [11]. Mycolactone can be detected in human skin samples from patients with Buruli ulcer by conventional TLC as a band at retention factor value 0.23 but a similar, weaker band was seen in normal human skin [11]. Therefore the extraction procedure has been modified and a new method has been developed to detect mycolactone using 2-naphthylboronic acid to enhance the fluorescence of the molecule [9].
In the present study the new procedure has been evaluated for accuracy in samples from patients with Buruli ulcer and patients with similar lesions that gave a negative result on PCR for the IS2404 repeat sequence of M.ulcerans
Patients were recruited from Buruli ulcer treatment centres from January 2014 to June 2014 in Benin, DR Congo, Ghana and Côte d’Ivoire if they had a skin lesion suspected to be caused by M.ulcerans infection. Samples were collected by fine needle aspiration or swab according to whether the lesion was non-ulcerated or ulcerated respectively and by biopsy if obtained at surgery (Fig 1). If PCR for the M. ulcerans repeat sequence IS2404 was positive they were included as Buruli ulcer disease patients. If the PCR was negative they were included in the control group.
Swabs, fine needle aspirates (FNA) or biopsy were put into O-ring seal plastic vials (Fisher Scientific) containing 1ml absolute ethanol. Vials were wrapped in aluminum foil and kept in the dark at room temperature. Samples from all the countries were transported to the Harvard laboratory within 3 weeks.
Samples were processed by a modification of the published method [9] (S1 FTLC procedure and S2 FTLC procedure). Ethanol containing the dissolved sample was filtered through a cotton plug into a glass vial. The sample container was further rinsed with 1 mL ethyl acetate, which was added to the glass vial through a cotton plug, and the contents were evaporated to dryness under reduced pressure of about 10-15mmHg using a rotary evaporator (Rotavapor R-210, Buchi). To separate any contaminating solid from liquid, 100 μL hexane/ether (1:1) solution was added to the glass vial, rinsed and transferred by micro-syringe into a clean glass vial which was air-dried. After evaporation, 50 μL hexane/ether (1:1) was added to the dry sample and 15 μL of the resuspended sample was spotted onto a 3×6 cm fluorescent-dye free TLC plate (TLC Silica gel 60, EMD Millipore, Darmstadt, Germany; Gibbstown, NJ, USA) alongside 40 ng synthetic mycolactone A/B standard in ethyl acetate and a co-spot of 10 μL sample with 40 ng synthetic mycolactone A/B. The plate was developed in chloroform: hexane: methanol at a ratio of 5:4:1 until the leading edge reached the top of the plate, air-dried and dipped in 0.1 M 2-naphthylboronic acid solution in acetone, then heated for 60 seconds at 100°C on a hot plate. The glass side of the plate was wiped with acetone on a paper towel. The plate was placed on a UV lamp with a 365 nm filter. The fluorescent band at retention factor 0.23 from the patient sample was compared to that of the standards to confirm the presence of mycolactone. Two readers were made to confirm the mycolactone test result before test result reporting and were blinded to the PCR test result.
Duplicate samples were taken from each patient for PCR targeting the M.ulcerans IS2404 repeat sequence. Samples were processed using the DNA extraction and PCR methods routinely used in diagnostic confirmation for patient care [12]. The amplification products were held at 4°C until they were processed further by agarose gel electrophoresis. The method of PCR included a negative extraction control and positive, negative and inhibition controls.
GraphPad Prism 5 software was used for data analysis. Descriptive statistics were used to obtain general descriptive information such as the mean and ranges from the data. One sample analysis (Fisher’s exact test) was used to compare two proportions or groups. Contingency tables were used to calculate the sensitivity, specificity and the predictive values for the various laboratory techniques employed.
Ethical approval for the study was obtained from the School of Medical Sciences, Committee on Human Research, Publication and Ethics (CHRPE/AP/229/12). Written informed consent was obtained from the patient or their parent/guardian before samples were obtained.
Table 1 shows the characteristics of the patients and the type of sample taken from each. There were 71 IS2404 PCR positive samples and 28 PCR negative samples which were used as controls. The final diagnosis in the control samples is not known. The mean age of patients in the control group was 41 years compared with 13 years in the Buruli ulcer disease group.
Out of 71 IS2404 PCR positive samples 52 were positive on fluorescent TLC giving a sensitivity of 73.2% (95% CI 61.4–83.1) (Table 2). 4 out of 28 true negatives gave a positive result on fTLC resulting in the specificity of 85.7% (95% CI 67.3–96.0). Thus 4/57 patients whose samples were positive by fTLC were false positives (7.0%). The positive predictive value (PPV) was 92.9% (95%CI 82.7–98.0) and the negative predictive value (NPV) 55.8% (95%CI 39.9–70.9). The sensitivity of fTLC was similar for swabs (73%), FNAs (75%) and biopsies (70%) (Table 3).
Of the 4 false positives, 2 were on the lower limb in patients aged 42 and 62. The clinical diagnosis of these lesions is unknown and no culture of the underlying organism was performed. The other two were on the arms in a 12 year old male and in a 13 year old female.
We have shown that mycolactone can be detected in 73% of M.ulcerans infected samples by fluorescent thin layer chromatography. The technique was easy to perform and the result could be read within 1 hour. The sensitivity was higher than that of microscopy (30–60%) or culture (35–60%) and compared favourably with that of histology (82%)[13, 14]. There was no difference in sensitivity when FNA and swabs were compared although the number of FNA samples was small. In this study samples collected in endemic countries over a 3-week period were shipped to the laboratory at Harvard and stored before testing which may have impacted on the sensitivity of the assay. This can be investigated by field evaluation of the test in endemic countries.
The specificity was 87% but the implication of four false positives is that 4 patients would have received treatment for 8 weeks with daily injections of streptomycin and oral rifampicin with the known potential for side effects associated with those drugs. Two of these patients were more than 40 years old and had lesions on their lower limbs. Confirmatory PCR would usually be necessary in this group of patients because the differential diagnosis is broad. However the other two were children with lesions on upper limb which is more worrying. The correct diagnosis in these cases is not known and it remains possible that the PCR results were false negatives. A larger study in which false positives are followed up and diagnosed specifically is required to resolve this problem. The positive predictive value was 92.9% and the negative predictive value 55.8% but this needs to be interpreted with caution as samples from different populations were analysed and the population prevalence was not taken into account.
The extraction efficiency for mycolactone from human tissue is low at 10 to 20% (Caroline Demangel, personal communication) but fTLC is sensitive enough to detect less than 10ng mycolactone extracted from a mouse footpad [9]. The number of FNA samples was low in the present study but mycolactone was detected in these in the same proportion as from swabs and biopsies. This is an important consideration since more non-ulcerated early lesions are seen when campaigns to increase awareness of Buruli ulcer are successful.
Development of a diagnostic test for M.ulcerans disease that can be carried out in local treatment centres is a high priority and at a consultative meeting organized by World Health Organisation/Neglected Tropical Diseases (WHO/NTD) and Foundation for Innovative New Diagnostics (FIND) in Geneva November 2013 [15], fTLC was identified as a promising technique with potential for implementation at the district level. Further studies are planned with the support of FIND and WHO (NTD and TDR) to address the logistics of introducing this test in endemic countries, to confirm its sensitivity and to investigate its specificity.
|
10.1371/journal.pmed.1002409 | A combination of plasma phospholipid fatty acids and its association with incidence of type 2 diabetes: The EPIC-InterAct case-cohort study | Combinations of multiple fatty acids may influence cardiometabolic risk more than single fatty acids. The association of a combination of fatty acids with incident type 2 diabetes (T2D) has not been evaluated.
We measured plasma phospholipid fatty acids by gas chromatography in 27,296 adults, including 12,132 incident cases of T2D, over the follow-up period between baseline (1991–1998) and 31 December 2007 in 8 European countries in EPIC-InterAct, a nested case-cohort study. The first principal component derived by principal component analysis of 27 individual fatty acids (mole percentage) was the main exposure (subsequently called the fatty acid pattern score [FA-pattern score]). The FA-pattern score was partly characterised by high concentrations of linoleic acid, stearic acid, odd-chain fatty acids, and very-long-chain saturated fatty acids and low concentrations of γ-linolenic acid, palmitic acid, and long-chain monounsaturated fatty acids, and it explained 16.1% of the overall variability of the 27 fatty acids. Based on country-specific Prentice-weighted Cox regression and random-effects meta-analysis, the FA-pattern score was associated with lower incident T2D. Comparing the top to the bottom fifth of the score, the hazard ratio of incident T2D was 0.23 (95% CI 0.19–0.29) adjusted for potential confounders and 0.37 (95% CI 0.27–0.50) further adjusted for metabolic risk factors. The association changed little after adjustment for individual fatty acids or fatty acid subclasses. In cross-sectional analyses relating the FA-pattern score to metabolic, genetic, and dietary factors, the FA-pattern score was inversely associated with adiposity, triglycerides, liver enzymes, C-reactive protein, a genetic score representing insulin resistance, and dietary intakes of soft drinks and alcohol and was positively associated with high-density-lipoprotein cholesterol and intakes of polyunsaturated fat, dietary fibre, and coffee (p < 0.05 each). Limitations include potential measurement error in the fatty acids and other model covariates and possible residual confounding.
A combination of individual fatty acids, characterised by high concentrations of linoleic acid, odd-chain fatty acids, and very long-chain fatty acids, was associated with lower incidence of T2D. The specific fatty acid pattern may be influenced by metabolic, genetic, and dietary factors.
| Fatty acid subclasses (e.g., saturated fatty acids or omega-6 fatty acids) and individual fatty acids in the blood have been studied to understand the aetiology of type 2 diabetes and as biomarkers of dietary intakes.
Existing studies suggest that different types of fatty acids are mutually correlated and are altered together by pharmacological intervention, dietary intervention, or both.
However, no study to our knowledge has reported whether a certain combination of different types of circulating individual fatty acids could be associated with the risk of type 2 diabetes.
We evaluated 27 individual fatty acids in the blood samples of adults within a large study from 8 countries in Europe, among a reference sub-cohort sample of 15,919 adults and among 12,132 adults who subsequently developed type 2 diabetes over a follow-up period of 12 years on average.
We identified a fatty acid combination that was partly represented by a combination of high concentrations of linoleic acid (the most abundant omega-6 polyunsaturated fatty acid), low concentrations palmitic acid (the major saturated fatty acid), and varying concentrations of the other essential and non-essential fatty acids.
We found that incidence of type 2 diabetes was lower by 63% on average when comparing the 20% of adults with fatty acid profiles most consistent with this particular combination (e.g., high linoleic acid and low palmitic acid) with the 20% of adults with fatty acid profiles least like this particular combination (e.g., low linoleic acid and high palmitic acid).
The combination of fatty acids was also linked to genes related to insulin resistance, cardiometabolic risk factors, and dietary intakes of polyunsaturated fatty acids, coffee, soft drinks, and dietary fibre.
A combination of fatty acids may be potentially important in the development of type 2 diabetes over and above individual fatty acids or fatty acid subclasses.
Dietary, pharmacological, and genetic investigations are warranted to characterise the clinical and biological implications of the combination of different types of individual fatty acids, for example, to predict the risk of type 2 diabetes, to better understand the aetiology of type 2 diabetes, and to consider interventions to favourably alter fatty acid profiles.
| Fatty acids play vital roles in metabolic homeostasis, serving as precursors of signalling molecules, energy sources, and constituents of membranes and functional lipids [1,2]. Reflecting their diverse roles, fatty acids have been evaluated as markers of physiological homeostasis, metabolic disorders, and dietary exposure in biological, clinical, and population-based research [3–5]. For example, blood or tissue levels of omega-3 polyunsaturated fatty acids (PUFAs) have been studied as a cardio-protective factor in biochemical and clinical research and as a biomarker of dietary consumption of omega-3 PUFAs in epidemiological research [1,2,4–6]. However, research to date has largely evaluated individual fatty acids or single subgroups of fatty acids, rather than combinations of fatty acids, in terms of mechanism or as potential biomarkers.
Combinations of fatty acids may have aetiological and clinical implications for metabolic diseases including type 2 diabetes (T2D). Insulin resistance and pancreatic lipotoxicity have been found to be influenced by multiple fatty acids. For example, palmitic acid (16:0) induces lipotoxicity, and unsaturated fatty acids may prevent it [7–9]. Pharmacological and nutritional research also warrant considering multiple fatty acids together. Interventions of lipid-lowering drugs or dietary carbohydrates or fats, for example, alter blood concentrations of individual PUFAs and saturated fatty acids (SFAs) jointly [3,9–11]. These findings support the notion that combinations of fatty acids are important to study in relation to the aetiology of T2D and to predict T2D risk.
A few epidemiological studies have identified combinations of circulating or tissue fatty acids associated with adiposity, hypertension, and risks of metabolic syndrome and cardiovascular diseases using a statistical pattern-recognition approach [12–15]. These studies have indicated potential biological and clinical importance of combinations over and above that of individual fatty acids. However, a combination of fatty acids has never been evaluated as a potential risk factor for incident T2D. Thus, we first aimed to identify 1 or more combinations of phospholipid fatty acids that explained variability in multiple fatty acid concentrations, using epidemiological data from the European Prospective Investigation into Cancer and Nutrition (EPIC)–InterAct study. Then, focussing on the single combination of fatty acids that explained the greatest variability, we tested the hypothesis that the combination is associated with the incidence of T2D. To provide mechanistic insights, we further examined the association of this combination of fatty acids with metabolic risk factors, genetic predisposition to obesity and insulin resistance, and dietary intakes in EPIC-InterAct. For metabolic and dietary factors, external validation was performed by evaluating data of the US National Health and Nutrition Examination Survey (NHANES).
We conducted this work as a substudy of the fatty acid project in EPIC-InterAct to explore a combination of fatty acids to add to our previous work on individual fatty acids and subclasses (S1 Protocol) [16,17]. EPIC-InterAct is a prospective study nested within 8 European countries of the EPIC study (Denmark, France, German, Italy, Netherlands, Spain, Sweden, and UK) [16,18]. In EPIC-InterAct, the case-cohort design was adopted to combine the advantages of a prospective design with the efficiency of a case-control design [19]. From the 340,234 adults with 3.99 million person-years of follow-up of the EPIC study, EPIC-InterAct (1) randomly selected 16,835 adults (‘sub-cohort’) and (2) identified 12,403 incident cases of T2D occurring by 31 December 2007; the identified cases included 778 cases in the sub-cohort by design (S1 Fig) [16,18]. All participants gave written informed consent. The study was approved by local ethics committees and the institutional review board of the International Agency for Research on Cancer [18].
The current study included 15,919 adults from the sub-cohort—after excluding 916 meeting 1 or more exclusion criteria: prevalent diabetes (n = 548), missing information on fatty acids (n = 156), missing information on incident T2D (n = 129), and post-censoring T2D (n = 4)—and included 12,132 incident T2D cases, after excluding 271 adults missing information on fatty acids (S1 Fig). In summary, we evaluated 27,296 adults in this study (12,132 cases, including 755 cases from the sub-cohort; and 15,919 adults from the sub-cohort).
Prevalent diabetes cases (excluded from the study) were identified by baseline self-report of a diagnosis, physician’s diagnosis, anti-diabetic drug use, or other evidence of T2D before the baseline date in EPIC-InterAct [18]. Incident T2D was ascertained from multiple information sources reviewed by each participating centre [18]: self-report, linkage to primary-care registers, secondary-care registers, medication use (drug registers), hospital admissions, and mortality data. Information from any follow-up visit or external evidence with a date later than the baseline visit was used. In Denmark and Sweden, incident cases were identified via local and national diabetes and pharmaceutical registers, and hence all ascertained cases were considered to be verified. Follow-up was to the date of diagnosis, 31 December 2007, or the date of death, whichever occurred earliest.
We evaluated relative concentrations of 27 individual fatty acids expressed as mole percentage of total plasma phospholipid fatty acids (Table 1), as previously described (S1 Text) [16,20]. These measurements were masked to case status. Thirty-seven fatty acids of plasma phospholipids were quantified by gas chromatography [20]. In the current analysis, 10 fatty acids were excluded because their relative concentrations were <0.05% on average. Coefficients of variation of the 27 fatty acids ranged from 1.9% to 4.6% [20].
At baseline, weight, height, and waist circumference were measured directly in every centre. Waist circumference was not measured in Umea, Sweden (n = 1,845) [18]. Sociodemographic factors, smoking status, and medical history were assessed by a questionnaire for general health. Physical activity was assessed by a questionnaire validated previously [21]. Dietary variables were derived centrally based on food frequency questionnaires or diet histories standardised in each cohort [22,23]. Using blood samples stored at −196°C (or −150°C in Denmark), biochemical assays were performed at Stichting Ingenhousz Laboratory, Etten-Leur, Netherlands, for glucose, triglycerides, high-density lipoprotein cholesterol (HDL-C), triglycerides, high-sensitivity C-reactive protein (hsCRP), and proteins related to hepatic function—alanine transaminase (ALT), γ-glutamyl transferase (GGT), and aspartate transaminase (AST)—as the liver is the major organ metabolising fatty acids.
Genetic information became available in 22,179 adults with fatty acid data, assayed with Illumina Human660W-Quad BeadChip (Illumina, Little Chesterford, UK; n = 9,166) and MetaboChip (Illumina; n = 13,013) [24]. Using these data, we conducted post hoc analyses to examine whether genetic predisposition to metabolic risk was associated with the FA-pattern score. We calculated weighted genetic risk scores for body mass index (BMI) (n loci = 97) [25] and for insulin resistance (n loci = 10) [24] using published measures of genome-wide associations (S1 Text).
Principal component analysis (PCA) was performed in the sub-cohort (n = 15,919) to combine multiple fatty acids (Table 1) together in a way to explain as much variation of those fatty acids as possible. Sampling weights were applied so that each of the 8 countries equally contributed to the PCA. Eigenvalues divided by 27 were assessed as percent of variance explained.
Principal components were inferred as representing fatty acid patterns. The pattern matrix from PCA was then used to calculate the scores, referred to as FA-pattern scores, among the rest of the study population (incident T2D cases not in the sub-cohort, Fig 1) and was also applied to quality control samples (n = 860) to assess the contribution of any batch effects [20]. We chose to focus on the first principal component for further aetiological analyses to provide potential biological implications of this single combination of fatty acids. This enabled us to conduct and report a detailed investigation into the associations of this combination with T2D incidence and metabolic, dietary, and genetic variables, and their biological implications; it also removed the need for subjective decisions about how many components to derive, which matrix rotation method to use, and how to account for multiple testing [26].
All analyses were performed using Stata (StataCorp, College Station, Texas, US), with αtwo-sided = 0.05. For descriptive purposes, a hierarchical cluster tree was generated to visually assess correlation between fatty acids [27]. Pearson correlation coefficients between fatty acids were also calculated. The strength of association of the FA-pattern score with incident T2D was evaluated by estimating hazard ratios (HRs) and 95% CIs from Prentice-weighted Cox regression, with age as the underlying timescale [19]. The estimates were obtained in each country and pooled by random-effects meta-analysis [28] for quintiles specific to the sub-cohort, for a continuous term per interdecile range (the difference between the 90th and 10th percentiles of the distribution), and for cubic-spline terms to test non-linear associations [29]. We additionally computed a 95% predicted interval for the primary results by combining random-effects variation (tau2) and variation of the main estimate [28].
The models included potential confounders, including demographics, prevalent heart disease and stroke, medication use, smoking status, physical activity, and dietary factors (consumption of alcohol, soft drinks, dietary fibre, fruits, vegetables, and processed meats), that are associated with cardiometabolic health in general. We also adjusted for BMI, waist circumference, glucose, lipids, hsCRP, and liver enzymes to examine their influence on the associations of interest. Potential confounding by genetic predisposition for greater BMI and insulin resistance was also assessed.
In pre-specified analyses, we examined whether observed associations varied by baseline age, sex, and BMI, testing an interaction term for each factor and the FA-pattern score in regression analysis. Effect modification by blood triglycerides, use of lipid-lowering drugs (yes or no), and alcohol consumption (consumer or non-consumer) was also tested post hoc because of the association of triglycerides with the FA-pattern score (r = −0.29) and possible effects of lipid-lowering drugs and alcohol on de novo lipogenesis. Missing covariates were imputed by country, using multiple imputation by chained equations with variables for the FA-pattern score, covariates, survival time, and case status [30]. We report results from single imputation, because between-imputation variability was <0.2% of total variability in multiple imputation (20 datasets); we performed sensitivity analysis using multiple imputation and complete-case analysis.
As a sensitivity analysis to assess whether HR varied over the follow-up time by reverse causation, stratified analysis was performed by splitting follow-up time at 7 years after baseline and by censoring any events occurring within the first 2 years as non-cases. We additionally evaluated the stability of our findings: examining the consistency of a main finding for the single principal component when PCA was performed after Box–Cox transformation, improving normality of distribution of all fatty acid variables. We also examined whether or not the main result was driven by single fatty acids or fatty acid subclasses through 2 approaches: adjusting models for single fatty acids and subclasses separately, and repeating the analysis after PCA of fatty acids excluding each of the 27 fatty acids or subclasses one at a time. We performed internal cross-validation [31]: First, we re-derived the FA-pattern score in a subset selected by country, age, sex, and BMI (test set); second, we applied the scoring matrix to another subset (validation set) to derive the FA-pattern score, and then we examined the associations of the independently derived score with incident T2D.
To investigate potential mechanisms for the association of the FA-pattern score with incident T2D, we estimated cross-sectional associations of the FA-pattern score with each of selected metabolic risk factors (BMI, waist circumference, lipids, glucose, hsCRP, and liver enzymes) using linear regression. Additionally, modified Poisson regression [32] was used to examine the cross-sectional association of the FA-pattern score with prevalence of hepatic steatosis defined as ALT greater than cut-points previously validated against ultrasound (30 U/l for men, 19 U/l for women) [33]. We further fitted linear regression to assess whether genetic risk scores for BMI and for insulin resistance (independent variables) could explain variability in the FA-pattern score (dependent variable). These regression models statistically adjusted for age and the other covariates used for longitudinal analyses.
We also evaluated dietary factors as potential lifestyle determinants of the FA-pattern score. Multivariable-adjusted linear models included dietary determinants as independent variables and the FA-pattern score (scaled to 1 standard deviation) as a dependent variable. This analysis evaluated major macronutrient and fibre intakes (nutrient-based analysis) and 18 selected foods or beverages (food-based analysis).
Recognising the risk of false-positive findings based on our data-driven approach, we conducted post hoc assessment of the external validity of the FA-pattern score derived in EPIC-InterAct, using cross-sectional data from NHANES 2003–2004 (n = 1,566) on total plasma fatty acids, metabolic factors, dietary factors, and potential confounders. Using the scoring matrix derived from EPIC-InterAct, we calculated the FA-pattern score in NHANES (S2 Text) [15]. Using linear regression adjusting for potential confounders, replication analyses were performed (S3 Text). In dietary analyses, 18 dietary items were first assessed in EPIC-InterAct with backward variable selection (p = 0.2 as a cutoff, additionally for the purpose of adjustment [34]) to identify which dietary variables predicted the FA-pattern score together. Then we tested selected dietary factors in NHANES for external validation (S3 Text) in linear regression adjusting for potential confounders and including the same dietary variables. The 18 food groups first tested in EPIC-InterAct were selected by possible biology of diets, fatty acid profiles, and T2D, and evaluated both individually and simultaneously.
The first component derived by PCA explained 16.1% of the variation of 27 fatty acids, and 6 to 10 components explained more variation than 1 fatty acid could explain (>3.7% of total; ‘eigenvalue’ > 1.0) (Fig 1). The first 4 components had loading values (e.g., >0.6 or <−0.6) in multiple fatty acid classes. Selected to gain insight into the biological importance of a combination of fatty acids, the first component reflected relationships between fatty acids varying in chain length and degree of unsaturation, including fatty acids that can be synthesised endogenously and those derived from dietary consumption (Table 1; Fig 1). A similar pattern was identified in cluster analysis, as fatty acids adjacent in the tree had similar loading values (Fig 1). Major contributors (correlation coefficients r > 0.5 or r < −0.5) were palmitic acid (16:0, r = −0.51), palmitoleic acid (16:1, r = −0.75), and γ-linolenic acid (18:3n-6, r = −0.51). Heptadecanoic acid (17:0) and very-long-chain SFAs (VLSFAs) with 20 or more carbons had positive contributions (r = 0.5–0.7), but their relative concentrations were low (<1% of total). While linoleic acid (18:2n-6) had a positive contribution (r = 0.45), the other PUFAs, and trans unsaturated fatty acids had lower contributions (−0.25 < r < 0.25) (Fig 1). The coefficient of variation of the FA-pattern score was 6.0% based on the quality control samples.
Adults with higher FA-pattern score were more likely to be women, non-smokers, non-users of lipid-lowering drugs, and those with generally healthier profiles of metabolic risk factors, while there was no significant relationship with age or education (S1 Table). Covariates had missing values in <5% of adults, except 49.9% for family history of diabetes, which was not assessed in 12 of the 26 study centres (S2 Table). Where it was assessed, 24.5% of participants had missing information.
In the longitudinal analysis of 12,132 cases per 190,148.9 person-years (11.9 y of follow-up per person on average), the FA-pattern score was strongly associated with incident T2D. Adjusted for sociodemographic variables, dietary factors, and medical history, the HR (95% CI) of T2D comparing the top to the bottom fifth of the FA-pattern score was 0.23 (0.19–0.29) (p trend < 0.001) (Table 2). The association persisted after adjustment for BMI (HR 0.32; 95% CI 0.25–0.40) and for triglycerides and HDL-C (0.37; 95% CI 0.27–0.50). Results changed little when additionally adjusted for concentrations of random glucose, hsCRP, hepatic enzymes, other dietary factors, family history of T2D, and genetic risk scores for obesity and insulin resistance (S3 Table). The association varied across the 8 countries (Fig 2; I2 = 88%); this variation was partly explained by country-specific mean ages and percentage of men (p < 0.05 each), although an inverse association was observed in all countries.
There was no evidence of effect modification by baseline age, sex, BMI, triglycerides, lipid-lowering drug use, or alcohol consumption (p interaction > 0.1 each). The main result was stable in sensitivity analyses that explored the influence of imputation, duration of follow-up, and normality of distribution (S3 Table). In analyses adjusting for individual fatty acid variables (S2 Fig), adjustment for 17:0 attenuated the estimates to the greatest extent among the fatty acid variables we evaluated, shifting the HR (95% CI) of 0.38 (0.30–0.47) to 0.53 (0.42–0.67), but with the association remaining significant. Cross-validation analysis confirmed the stability of the findings (S4 Table). For instance, when the FA-pattern score was derived in 7 countries, not 8, and the scoring algorithm was applied to adults in the 1 country excluded, the summary HR (95% CI) in the most adjusted model was 0.40 (0.34–0.50).
In both EPIC-InterAct and NHANES (see S5 Table for scoring coefficients), the FA-pattern score was associated with metabolic risk factors in the direction consistent with the above findings for incident T2D. Inverse associations were seen with BMI, triglycerides, glucose, hsCRP, ALT, AST, GGT, and the likelihood of having hepatic steatosis (p < 0.001 each) (Table 3). A significant positive association with HDL-C was observed in NHANES (p < 0.001), but not in EPIC-InterAct (p = 0.7).
In genetic analyses (EPIC-InterAct only), a gene score related to higher BMI was not significantly associated with the FA-pattern score: +0.1% of SD of the FA-pattern score (95% CI −0.6% to +1.7%; p = 0.3) per interdecile range of the genetic score. A gene score related to higher insulin resistance was significantly associated with lower FA-pattern score: −1.9% of SD (95% CI −3.4% to −0.4%; p = 0.02).
In dietary analyses in EPIC-InterAct and NHANES, higher intakes of PUFAs and fibre were associated with higher FA-pattern score in both cohorts (Fig 3). For example, replacing carbohydrates with PUFAs in the diet by an amount equivalent to 5% of total energy was positively associated with the FA-pattern score (0.43 SD of the score, 95% CI 0.30–0.57) in EPIC-InterAct and 0.21 (95% CI 0.11–0.32) in NHANES. In food-based analysis of EPIC-InterAct, the FA-pattern score was significantly related to higher intakes of fish, margarine, and coffee and lower intakes of soft drinks and alcoholic beverages (p < 0.05 each) when assessed individually (S3 Fig) and simultaneously (Fig 3). In NHANES, findings from EPIC-InterAct for soft drinks, coffee, and alcohol were replicated (p < 0.04) (Fig 3).
We evaluated fatty acid profiles among adults in 8 European countries and derived a FA-pattern score that represents a combination of both essential and non-essential fatty acids and that is characterised by high relative concentrations of linoleic acid (18:2n-6), stearic acid (18:0), odd-chain SFAs, and VLSFAs (≥20 carbons), and by low relative concentrations of γ-linolenic acid (18:3n-6), monounsaturated fatty acids (MUFAs), and long-chain SFAs (14:0 and 16:0). The unique combination was associated with dietary, metabolic, and genetic factors, and prospectively associated with a lower incidence of T2D. Comparing the top fifth to the bottom fifth of the FA-pattern score, T2D incidence was lower by approximately 60%. This robust association with incident T2D was independent of established risk factors and also any single fatty acids or fatty acid subclasses. These findings support the hypothesis that a combination of multiple fatty acids is an important marker for the development of T2D above and beyond the roles of single types of fatty acids. The combination of essential and non-essential fatty acids is of strong interest for further clinical or population-based investigations to predict T2D risk, identify interventional agents for T2D prevention, and better understand the aetiology of T2D.
The combination of fatty acids contributing to the identified FA-pattern score fits with known mechanisms involving the de novo lipogenesis (DNL) pathway. In DNL, fatty acids including 14:0, 16:0, 16:1n-7, and 18:1n-9 are synthesised endogenously, where stearoyl-CoA desaturase (SCD) converts 16:0 to 16:1n-7 as a rate-limiting step of fatty acid synthesis. The inverse correlation of 18:2n-6 with these fatty acids may reflect its role as a ligand of peroxisome proliferator-activated receptor α (PPARα) [2,5]. PPARα down-regulates SCD and ELOVL (elongation of very-long-chain fatty acid) enzymes, explaining the observed inverse correlation of 18:2n-6 with MUFAs, 16:0, 18:3n-6, and other n-6 PUFAs [2]. An exception of PPARα’s action is activation of ELOVL3, which leads to synthesis of VLSFAs in the adipose tissue [35] and supports the observed associations between 18:2n-6, 18:0, and VLSFAs. While our findings are in line with the benefit of dietary PUFAs (predominantly 18:2n-6), other major PUFAs (e.g., omega-3 PUFA) contributed little to the primary fatty acid combination. This could reflect their diverse roles in eicosanoid pathways and pro- and anti-inflammatory pathways, and their associations with dietary intakes (e.g., fish) independent of DNL-driving dietary factors [5].
These suggested mechanisms are linked to the development of T2D. Activation of PPARα suppresses hepatic DNL and pro-inflammatory pathways that lead to insulin resistance, dyslipidaemia, and fatty liver [2,5,36]. In an experimental setting, for example, PPARα knock-out mice developed fatty liver exhibiting overt hepatic lipogenesis [37]. Main products of DNL include 16:0 and diacylglycerols that cause a pro-inflammatory response, endoplasmic reticulum stress, and insulin resistance [1,38]. Thus, our analysis yielded a combination of multiple fatty acids that may represent biological pathways related to insulin resistance, inflammatory responses, and T2D risk. This was confirmed with the observed associations of FA-pattern score with metabolic risk factors in an expected direction and the association of FA-pattern score with genetic predisposition to insulin resistance. Associations of gene variants with fatty acids and with incident T2D cannot be confounded by long-term lifestyle characteristics. Therefore, the specific gene variants, the identified combination of fatty acids, and the risk of T2D are likely to be on the same causal pathway, warranting future research to elucidate how insulin resistance specifically alters fatty acid profiles or vice versa.
Our analysis and prior studies derived a similar combination of fatty acids [12–15], but no previous studies to our knowledge evaluated T2D incidence as an outcome. Past studies used different methods and examined varied numbers of fatty acids (10 to 42) of phospholipids [15], cholesteryl esters [12,15], plasma [13], or adipose tissue [14]. Despite the differences, all of the studies reported a combination of fatty acids partly driven by higher levels of 18:2n6 with lower 16:0, which could reflect activity of DNL [39]. These combinations were found to be associated with lower blood pressure, greater endothelial function, less weight gain over time, or lower risk of metabolic syndrome [12–14]. In contrast, an association with ischaemic heart disease or stroke was not significant [15]. The inconsistency depending on outcome is predictable as DNL could promote insulin resistance, but suppress atherosclerosis [36]. Statins and other lipid-lowering drugs also alter fatty acid profiles [10] and have divergent effects on heart disease and T2D [40]. Thus, a fatty acid pattern can be a future focus of investigations of cardiometabolic diseases and related interventions.
The association of the FA-pattern score with incident T2D was not fully explained by any single fatty acid, but was partly attenuated by adjustment for odd-chain SFAs and VLSFAs. These SFAs are associated with lower risk of cardiometabolic diseases [16,41,42], while their biological roles remains understudied. High phospholipid VLSFAs may reflect high activity of PPARα, which leads to VLSFA synthesis and less insulin resistance, apoptotic cell death, and pancreatic dysfunction [41,43]. Blood odd-chain SFAs may partly reflect dairy consumption [3,44], gut microbiota [45], or endogenous synthesis through α-oxidation [46], and thus any correlates to those factors could explain our findings. Evidence for these mechanisms and relationships with other fatty acids remains scarce and deserves further investigation.
Dietary correlates with the combination of fatty acids deserve discussion, as they were replicated in EPIC-InterAct and NHANES: alcohol and soft drinks as negative correlates, and coffee, fibre, and PUFAs as positive correlates. The finding for alcohol consumption is likely to reflect its lipogenic effect, a risk factor for liver cirrhosis and T2D [47]. Regarding coffee consumption, polyphenols may deactivate DNL [48] and lower triglyceride levels and T2D risk [49–51]. Increased PUFA intake (predominantly n-6 PUFAs) could increase insulin sensitivity as well as lower DNL [2,5,52]. Our findings for soft drinks and fibre may also reflect their glycaemic and anti-glycaemic effects, respectively, as a high glycaemic effect leads to insulin secretion and DNL [2,53].
Strengths of this work include the standardised assay of fatty acid profiles in an EPIC-InterAct population with geographic diversity; the large study size (to our knowledge by far the largest among studies of fatty acid biomarkers), allowing various sensitivity analyses; and the generalisability of our findings, strengthened by our findings across 8 European countries and the US NHANES. By focussing on a single combination of fatty acids, we were able to report details of its association with incident T2D and metabolic, genetic, and dietary factors. However, this also limited our investigation of other fatty acid patterns, in particular in relation to omega-3 fatty acids. Other limitations include possible residual confounding by factors unmeasured or measured imprecisely, although we adjusted for many covariates including major risk factors for T2D. Whether or not the combination of fatty acids itself caused the T2D onset remains unestablished. Possible exposure misclassification due to single fatty acid measurements and possible outcome misclassification were limitations, but likely to be independent of T2D case status and fatty acid profiles, respectively. We found no strong reason to think that these limitations would alter the overall conclusions. Lastly, the generalisability of our findings might be limited to high-income Western populations, and fatty acid patterns in other populations with diverse genetic backgrounds and dietary patterns are of future interest.
In conclusion, we identified a combination of plasma phospholipid fatty acids characterised by high relative concentrations of 18:2n-6, VLSFAs, and odd-chain SFAs and low relative concentrations of long-chain SFAs and MUFAs, some of which are synthesised endogenously. This particular profile was associated with a 3-fold lower relative risk of incident T2D in European populations after adjustment for confounding. While both genes and diet were linked to the FA-pattern score, association of the FA-pattern score with T2D was independent of established risk factors for T2D and not driven by individual fatty acids. These findings highlight that multiple fatty acids are jointly related to the development of T2D. The combination of fatty acids warrants further investigation of its determinants and potential application as a marker of metabolic characteristics.
|
10.1371/journal.ppat.1006747 | Modulation of host central carbon metabolism and in situ glucose uptake by intracellular Trypanosoma cruzi amastigotes | Obligate intracellular pathogens satisfy their nutrient requirements by coupling to host metabolic processes, often modulating these pathways to facilitate access to key metabolites. Such metabolic dependencies represent potential targets for pathogen control, but remain largely uncharacterized for the intracellular protozoan parasite and causative agent of Chagas disease, Trypanosoma cruzi. Perturbations in host central carbon and energy metabolism have been reported in mammalian T. cruzi infection, with no information regarding the impact of host metabolic changes on the intracellular amastigote life stage. Here, we performed cell-based studies to elucidate the interplay between infection with intracellular T. cruzi amastigotes and host cellular energy metabolism. T. cruzi infection of non-phagocytic cells was characterized by increased glucose uptake into infected cells and increased mitochondrial respiration and mitochondrial biogenesis. While intracellular amastigote growth was unaffected by decreased host respiratory capacity, restriction of extracellular glucose impaired amastigote proliferation and sensitized parasites to further growth inhibition by 2-deoxyglucose. These observations led us to consider whether intracellular T. cruzi amastigotes utilize glucose directly as a substrate to fuel metabolism. Consistent with this prediction, isolated T. cruzi amastigotes transport extracellular glucose with kinetics similar to trypomastigotes, with subsequent metabolism as demonstrated in 13C-glucose labeling and substrate utilization assays. Metabolic labeling of T. cruzi-infected cells further demonstrated the ability of intracellular parasites to access host hexose pools in situ. These findings are consistent with a model in which intracellular T. cruzi amastigotes capitalize on the host metabolic response to parasite infection, including the increase in glucose uptake, to fuel their own metabolism and replication in the host cytosol. Our findings enrich current views regarding available carbon sources for intracellular T. cruzi amastigotes and underscore the metabolic flexibility of this pathogen, a feature predicted to underlie successful colonization of tissues with distinct metabolic profiles in the mammalian host.
| The kinetoplastid protozoan, Trypanosoma cruzi, is a highly successful parasite with a broad mammalian host range and the capacity to colonize a variety of tissues within a given host to establish life-long infection. T. cruzi infection causes Chagas disease in humans, characterized by severe cardiomyopathy and gastrointestinal motility disorders, with limited treatment options. Despite the critical role of T. cruzi amastigotes in sustaining mammalian infection, little is known about their metabolic requirements or the range of nutrients available to these parasites in the host cell cytoplasm. Here, we demonstrate that T. cruzi infection triggers a host response in infected cells that includes increased mitochondrial respiration and biogenesis and increased glucose uptake into infected cells. We show that exogenous glucose supports optimal intracellular parasite replication and that cytosolic T. cruzi amastigotes access glucose in situ, potentially via a facilitated transport process characterized here. These findings expand our view of the range of carbons available to the intracellular parasite and suggest even greater metabolic flexibility of the tissue-infective stages of T. cruzi than previously appreciated, a capability predicted to contribute to successful host colonization.
| Chagas disease is a vector-borne parasitic disease caused by the kinetoplastid protozoan parasite Trypanosoma cruzi. Acute T. cruzi infection is most often asymptomatic or characterized by flu-like symptoms, but can cause severe and fatal myocarditis in the first weeks following infection [1]. More typically, parasites establish chronic infection that is controlled, but not eliminated, by host immune mechanisms [2]. A subset of chronically infected individuals develop progressive disease characterized by serious cardiac and gastrointestinal disturbances [3], for which treatment options are limited [4]. T. cruzi exhibits a broad mammalian host range where it can colonize diverse tissue types [5, 6]. In the chronic stage of infection, when parasites are maintained at very low densities, persistence has been reported most often in cardiac muscle, gastrointestinal smooth muscle and adipose tissue [6–9]. Knowledge of the molecular mechanisms governing successful intracellular colonization, replication and long-term persistence by T. cruzi are currently lacking but represent potentially exploitable processes for the development of new therapeutics.
In mammalian hosts, T. cruzi transitions between two main developmental forms. The non-dividing, motile trypomastigote can actively penetrate most nucleated cell types by exploiting the host cell plasma membrane repair process [10]. Once inside the host cell, the parasite sheds its temporary vacuole [11] and progresses through a developmental program that culminates in the formation of the morphologically and biochemically distinct amastigote form that replicates in the host cell cytosol. Transcriptomic profiling of this developmental transition revealed strong signatures of global metabolic reprogramming in the parasite as it transforms from the trypomastigote to the amastigote stage [12]. As an obligate intracellular parasite, T. cruzi amastigotes are forced to draw from host nutrient pools to fuel their growth and survival, although nutrient uptake by amastigotes in situ has not been directly demonstrated. On the basis of expression data [12–14], functional studies [15], and metabolic assays conducted with isolated amastigotes [16, 17], it has been proposed that amino acids and fatty acids are the most likely sources of carbon for T. cruzi amastigotes to fuel their metabolism. Hexose sugars have largely been discounted as a potential carbon source for this cytosolic pathogen [13, 15, 18], due to the perception that glucose is a negligible commodity in the interior of a mammalian host cell [18] and the failure to demonstrate hexose transporter expression or uptake in isolated T. cruzi amastigotes [13, 15]. Nevertheless, the ability of T. cruzi to colonize a wide range of cell and tissue types predicts a degree of metabolic flexibility and/or the potential for the parasite to reprogram host metabolic pathways to suit its specific metabolic requirements as reported with some viral and bacterial pathogens [19–21].
Metabolic abnormalities have been reported in chronic Chagas patients [22–24] and in animal models of acute and chronic T. cruzi infection [25–29]. These include dysregulation of glucose [23, 28] and lipid metabolism [29] as well as mitochondrial electron transport chain dysfunction [26, 27, 30] in T. cruzi-infected skeletal [31] and cardiac muscle [25]. Recently, metabolite profiling studies have revealed increased uptake and metabolism of glucose in T. cruzi-infected cardiac muscle [32]. At the cellular level, transcriptomic analyses reveal modulation of host metabolic pathway expression including upregulation of metabolite transporters in T. cruzi-infected fibroblasts [12, 33]. How such metabolic changes in the host impact the intracellular T. cruzi amastigote life cycle have not been determined. However, results of genome-scale functional studies predict that the immediate metabolic environment can influence intracellular parasite growth [34]. In the present study, we sought to determine how T. cruzi infection impacts host glucose metabolism and mitochondrial respiration at the cellular level and how parasite-triggered changes in host cellular metabolism influence the intracellular infection cycle.
Metabolite profiling of T. cruzi-infected hearts has provided evidence of increased glucose uptake and metabolism at the whole organ level [32]. Here, we examined the impact of T. cruzi infection on glucose metabolism at the cellular level using low passage normal human dermal fibroblasts (NHDF), which have been shown to increase host glucose transporter expression during T. cruzi infection [12, 33]. Following infection (48 hpi) we observed a significant increase in 2-deoxyglucose ([3H]-2-DG) uptake into infected fibroblast monolayers in a manner that correlated with increasing parasite load (Fig 1A) under conditions where host cell abundance remained unchanged (S1A Fig). [3H]-2-DG uptake by both uninfected and infected NHDF was blocked by cytochalasin B (Fig 1B), consistent with a role for host plasma membrane glucose transporters [35] in mediating this host cell response to T. cruzi infection. Glucose uptake assays performed in parallel with fibroblasts (Fig 1C) and mouse skeletal myoblasts (Fig 1D) following infection with one of three T. cruzi strains: Tulahuén, CL Brener and CL-14, revealed comparable results with increased glucose uptake occurring in parasite-infected host cells as a generalized response among the isolates and mammalian cells tested here.
Glucose transport by mammalian cells is a highly regulated process [36] that is responsive to acute changes in the environment, including glucose restriction [37], intracellular pathogen infection [19–21], and acute exposure to PAMPs or cytokines [38, 39]. Physiologic triggers leading to increased glucose uptake, including pathogen infection, frequently promote increased glycolytic rates and lactate production from pyruvate, as well as decreased flux through the TCA cycle with reduced respiratory rates [40–42]. Unlike these examples, we find no increase in lactate secretion to accompany increased glucose uptake into T. cruzi-infected host cells (Fig 2A), but evidence of increased mitochondrial respiration in parasite-infected fibroblasts (Fig 2B) as determined by monitoring the oxygen consumption rate (OCR) in cell monolayers using a Seahorse extracellular flux analyzer. While the OCR measured in T. cruzi-infected monolayers was consistently greater than that measured in uninfected cell monolayers (Fig 2B), the potential for parasite respiration to contribute to the total OCR signal complicated immediate interpretation of this result. To examine this further, we sought a method to specifically inhibit T. cruzi amastigote respiration in situ in order to reveal the host and parasite contributions to the total OCR signal. For this, we utilized the endochin-like quinolone ELQ300, which targets cytochrome bc1 in the mitochondrial electron transport chain of apicomplexan parasites without affecting mammalian respiratory complexes [43, 44]. To validate the utility of ELQ300 for our purpose, we performed preliminary experiments to demonstrate that respiration in isolated T. cruzi amastigotes was inhibited by ELQ300 in a dose-dependent manner. Treatment with 1 μM ELQ300 resulted in maximal inhibition (90%) of basal OCR in isolated amastigotes (S1B Fig) with no impact on host mitochondrial respiration (S1C Fig). To confirm that ELQ300 was effective in blocking T. cruzi amastigote respiration in situ, we exploited a mitochondrial complex III-deficient human fibroblast cell line (CIII mutant) [45, 46] with significantly lower respiratory rates than normal human fibroblast control lines (S1D Fig). Treatment of infected CIII mutant fibroblasts with 1 μM ELQ300 almost completely abrogated the increase in OCR due to infection (S1E Fig), consistent with the compound inhibiting >90% of amastigote respiration in situ. In contrast, experiments performed in parallel with respiration-competent, control human fibroblasts (S1F Fig) show that following treatment with 1 μM ELQ300 to inhibit parasite respiration, a significant residual OCR signal remained, attributable to increased host cell mitochondrial respiration associated with T. cruzi infection. Similar results were obtained with T. cruzi-infected NHDF treated with 1 μM ELQ300 (Fig 2C) or with 1 μM GNF7686 (S1G–S1I Fig), a compound for which T. cruzi cytochrome b is the validated target [47]. Therefore, through differential targeting of T. cruzi mitochondrial complex III using small molecule inhibitors, we were able to measure intracellular amastigote respiration in situ and to determine that host mitochondrial respiration increases as a result of T. cruzi infection. We further demonstrate that increases in host mitochondrial respiration are accompanied by increased host mitochondrial content specifically within the parasitized subpopulation of the infected cell monolayers (Fig 2D and 2E (GFP+); S2 Fig).
To assess the potential for altered host glucose and mitochondrial metabolism to impact intracellular amastigote replication, flow cytometry-based proliferation assays were performed that enabled determination of the number of divisions that an individual amastigote has undergone in infected host cells within a set time frame, following exposure to different conditions. Examination of T. cruzi amastigote proliferation in complex III mutant fibroblasts, which display a significantly reduced mitochondrial respiratory capacity as compared to normal fibroblasts (S1D Fig), reveals nearly identical proliferation profiles for amastigotes in CIII mutant or normal control fibroblasts (Fig 3A). In contrast, amastigote proliferation was substantially reduced when glucose was removed from the extracellular medium, with the majority of amastigotes completing only 3 divisions within 48 hours rather than 4 divisions as when glucose was present (Fig 3B). The inhibitory effect of glucose restriction on T. cruzi amastigote growth was greatly enhanced by the addition of the glucose analogue 2-deoxyglucose (2-DG) (Fig 3C), which inhibits glycolysis. Notably, the fibroblast host monolayer was not measurably impacted even in the absence of exogenous glucose, until concentrations >2 mM 2-DG were reached (S3A Fig). In the absence of exogenous glucose, amastigote proliferation was arrested by the addition of 2 mM 2-DG (Fig 3D), where the median number of parasites/infected host cell was 1 (S3B Fig). However, the continued presence of viable amastigotes in infected monolayers at 66 hpi in cultures treated with 2 mM 2-DG in the absence of glucose (Fig 3E) suggests that this effect is cytostatic rather than lethal for the parasite. The near complete arrest of intracellular T. cruzi growth in the presence of 2-DG, as opposed to the more modest effect of glucose restriction alone (Fig 3B and 3E), suggests that 2-DG may directly inhibit parasite glucose metabolism in addition to its inhibitory effect on host glycolysis. This implies that the parasite can access and internalize this glucose analog in situ, which counters the hypothesis that glucose is not accessible to the amastigote stage of T. cruzi [15, 18].
To explore the relationship between T. cruzi amastigotes and exogenous glucose more closely, we first examined the capacity of amastigotes, isolated from NHDF monolayers at 48 hpi, to utilize exogenous glucose to drive glycolysis and mitochondrial respiration, employing glutamine as a positive control to fuel respiration [17]. Extracellular flux analysis revealed that isolated T. cruzi amastigotes respond to exogenous glucose with significant increases in OCR (Fig 4A) and extracellular acidification rate (ECAR), which correlates with glycolytic activity, (Fig 4B) that were quenched by the injection of 2-DG. A similar increase in amastigote OCR was observed in response to glutamine, with little change in ECAR as expected (Fig 4A and 4B). To ensure that exogenous glucose is metabolized by the isolated amastigotes, rather than triggering metabolic changes in the parasite through an independent mechanism, metabolite profiling was performed following incubation of isolated amastigotes in medium containing [13C]-U-glucose for 3 hours. As shown in Table 1, 13C incorporation was detected in glycolytic and TCA cycle intermediates as well as pentose phosphate pathway intermediates and several amino acids, providing direct confirmation that T. cruzi amastigotes are capable of internalizing and metabolizing exogenous glucose in catabolic and anabolic processes.
Next, we performed transport assays to measure the kinetics of [3H]-2-DG uptake by freshly isolated intracellular amastigotes and extracellular trypomastigotes, a life cycle stage of T. cruzi for which hexose transporter expression is abundant [12, 15]. The initial rates of hexose transport (V0) measured for isolated amastigotes and trypomastigotes were found to be comparable, with a similar KM (87.0 ± 21.7 vs. 81.2 ± 3.7 μM) and Vmax (857.0 ± 76 vs. 666.5 ± 36.1 pmol 2-DG/mg protein/min) (Fig 4C). We then sought to determine whether the capacity for glucose uptake and metabolism by T. cruzi amastigotes is relevant in the context of an intracellular infection of mammalian cells. Infected NHDF monolayers (48 hpi) were pulsed with [3H]-2-DG for 20 minutes in the presence and absence of cytochalasin B, which significantly impairs glucose transport in mammalian cells but not in T. cruzi [35, 48]. Intracellular amastigotes purified from host cells showed incorporation of [3H] when mammalian glucose uptake was not inhibited with cytochalasin B (Fig 4D). To establish that [3H]-2-DG was internalized by the intracellular parasites and not non-specifically bound to amastigote surfaces following disruption of infected cells, isolated amastigotes were treated with the pore-forming peptide antibiotic, alamethicin [49] to permeabilize the parasite membrane (S4A Fig), which resulted in the release of >50% of the amastigote-associated label on average (Fig 4E). Similar evidence of in situ [3H]-2-DG uptake by intracellular T. cruzi was observed when parasites were resident in mouse skeletal myoblasts (S4B Fig). Additional examination of CL Brener (S4C Fig) and CL-14 (S4D Fig) strain amastigotes in fibroblasts provided further evidence for the internalization of [3H]-2-DG by intracellular T. cruzi amastigotes in situ. We further demonstrate that glucose, as the sole exogenous carbon source available to isolated amastigotes, is capable of sustaining ATP pools in the parasite over a 24-hour period, at levels similar to a mixture of glucose, glutamine and pyruvate (Fig 4F). Combined, these data demonstrate the potential for glucose to be utilized by intracellular T. cruzi amastigotes as a fuel for parasite metabolism in situ.
We have identified changes in host cellular metabolism associated with intracellular T. cruzi infection which include increased glucose uptake by infected host cell monolayers, increased mitochondrial respiration, and evidence of increased mitochondrial content specific to parasitized host cells. While these metabolic perturbations likely reflect multiple complex origins including compensatory changes triggered by increased metabolic demands, our data indicate that the parasite benefits from the increased glucose transport observed by infected cells, where exogenous glucose levels impact the proliferation rate of intracellular T. cruzi amastigotes. The growth-promoting effect of extracellular glucose could arise if cytosolically-localized T. cruzi amastigotes access intracellular glucose pools directly or if host cell metabolic processes fueled by extracellular glucose indirectly modulate the ability of amastigotes to efficiently replicate. While we cannot rule out the latter possibility, we provide evidence that intracellular amastigotes from different T. cruzi strains are capable of internalizing and retaining the radiolabeled glucose analog, [3H]-2-DG, during their intracellular replicative cycle in the mammalian cell cytoplasm. By studying amastigotes in isolation we also definitively demonstrate their capacity to take up exogenous glucose and metabolize this carbon to fuel glycolysis and mitochondrial respiration. 13C-glucose tracer studies confirm results of bioenergetics studies and further demonstrate that exogenous glucose is also shuttled into anabolic pathways including the pentose phosphate pathway. Combined with the finding that in the absence of other exogenous carbon sources, glucose is capable of sustaining ATP levels in isolated T. cruzi amastigotes to a similar degree as a mixture of substrates, our data indicate the potential for glucose to serve as an important substrate for the intracellular life stages of T. cruzi.
These data counter the view that glucose is unlikely to be utilized by intracellular T. cruzi parasites [18]. While glucose concentrations in the mammalian cell cytoplasm have previously been considered insufficient to support the growth of cytosolic pathogens, this argument is weakened with studies using fluorescent glucose sensors that demonstrate the existence of a significant and dynamic pool of cytosolic glucose in multiple human cell lines [50, 51]. A more direct argument against glucose as a potential substrate for intracellular T. cruzi amastigotes is the report that hexose transporter expression, as well as the ability to transport exogenous glucose, is negligible in this life cycle stage of the parasite [15]. However, our demonstration that isolated T. cruzi amastigotes transport glucose with similar kinetics as the trypomastigote stage of the parasite suggest that amastigotes utilize a facilitated hexose transporter with similar properties, if not identical, to the transporters expressed by extracellular T. cruzi life stages [52]. We further confirmed hexose transporter expression in amastigotes from three different T. cruzi strains by quantitative RT-PCR, albeit at much lower levels than in trypomastigotes. Notably, CL-14 amastigotes had the lowest expression (S5 Fig), but the capacity to take up glucose from the host cytosol was exhibited by each parasite strain. However, without targeted molecular studies, the role of hexose transporters in glucose acquisition by intracellular T. cruzi amastigotes versus possible alternative mechanisms such as fluid-phase endocytosis through the amastigote cytostome [53] remains unresolved.
Consistent with the potential for mammalian cells to sense fuel imbalances that may be incurred with the acquisition of glucose and other carbons by resident intracellular parasites and to mount a compensatory response [37, 41], we find that parasitized cells have more mitochondria and increased basal respiration. Selective inhibition of T. cruzi amastigote respiration with small molecule inhibitors of parasite cytochrome bc1, ELQ300 [43] and GNF7686 [47], enabled us to distinguish between parasite and host respiration and demonstrate elevated respiratory rates of mammalian cells during infection. However, unlike the impact of glucose restriction and/or 2-DG on amastigote proliferation, reduced host cell respiration, as seen in the mitochondrial complex III-deficient lines, did not impact T. cruzi replication. Infection outcomes are also anticipated to be host cell type–and perhaps parasite strain–dependent, as increased mitochondrial respiration in T. cruzi-infected macrophages was previously shown to be associated with increased nitric oxide production and parasite clearance [54], while our results show no association in non-phagocytic cells. Additional studies are needed to better understand the complex interplay between T. cruzi and host metabolism at the cellular, organ and organismal levels. How metabolic changes incurred at the cellular level impact regional and global metabolism in infected mammalian hosts [26, 27, 30–32, 55] and vice versa and how these changes impact the pathophysiology of disease are critical questions for future investigation.
In summary, we demonstrate that T. cruzi infection modulates host cell metabolism, stimulating glucose uptake into infected monolayers, which can be scavenged directly by intracellular amastigotes for utilization in energy generating and biosynthetic processes. Thus, in addition to amino acids and fatty acids predicted to constitute the main intracellular source of carbon for T. cruzi amastigotes [13, 18], we propose that glucose offers additional flexibility with respect to fuel utilization by these intracellular parasites. While the exact degree of T. cruzi amastigote metabolic plasticity has yet to be determined, a greater number of nutrient options is predicted to enhance the chances of parasite survival in different host tissues and under varying environmental conditions, including pharmacological inhibition of specific metabolic pathways.
Mammalian cell lines: mouse skeletal muscle myoblast (C2C12; ATCC #CRL-1772), African green monkey kidney epithelial (LLcMK2; ATCC #CCL-7) and human dermal fibroblasts (NHDF; ATCC #CRL-2522 and NHDF-Neo; Lonza, #CC-2509) were propagated in Dulbecco’s Modified Eagle Medium (DMEM; Hyclone) supplemented with 1 mM pyruvate, 25 mM glucose, 2 mM glutamine, 100 U/ml penicillin, 10 μg/ml streptomycin and 10% fetal bovine serum (FBS) (D-10) at 37°C and 5% CO2. Human patient dermal fibroblast lines were purchased from the Cell line and DNA Bank of Genetic Movement Disorders and Mitochondrial Diseases (GMD-MDbank): Complex III mutant fibroblast harbor a mutation in subunit BCS1L of ETC complex III [45] (CIII mutant; GMD-MDbank #F-MT2614), Normal 1 fibroblast (GMD-MDbank #F-CR2631; Normal 2 fibroblast (GMD-MDbank #F-CR2571) were propagated in D-10 medium containing 50 μg/mL uridine (Sigma-Aldrich) at 37°C and 5% CO2.
Mammalian cell lines expressing mCherry targeted to the mitochondrial matrix were generated by retroviral transduction of NHDF and C2C12 with a construct containing the sequence encoding the first 25 amino acids of the mouse Cox8a protein fused to mCherry [58, 59]. Briefly, 5 x 105 Phoenix-AMPHO packaging cells (ATCC #CRL-3213) were plated in a 100 mm tissue culture dish and transfected the following day with 10 μg of the plasmid pLNCX2 containing the chimeric sequence (kindly provided by C-H Lee, HSPH) using TransIT-LT1 (Mirus Bio) per manufacturer’s protocol. Virion-containing medium obtained from Phoenix cell cultures 2 days post-transfection was passed through a 0.45 μm filter, and stored at -80°C. Mammalian cells were seeded 1.5 x 105 per well in a 6 well plate, and viron-containing medium with 4 μg/mL polybrene was added the following day. Transgenic fibroblasts were selected with 400 μg/mL G418 (Sigma-Aldrich) starting two days post-transduction and confirmed by microscopy and flow cytometry after 2 weeks.
Trypanosoma cruzi Tulahuén strain parasites (from M. Perrin, Tufts University) were transfected with pROCK-GFP-NEO for constitutive expression of GFP from the tubulin locus [60]. Axenically-grown T. cruzi epimastigotes were transfected as described [61] with minor alterations. Briefly, 10 μg of linearized plasmid was transfected into 4 x 107 epimastigotes in 100 μL of Tb BSF-buffer using the U-33 program of an Amaxa Nucleofector II (Lonza). Cells were subsequently transferred to 5 mL LIT and incubated at 27°C overnight, then cloned in 96-well plates. After 3 weeks, clones were screened by flow cytometry, and GFP expressing parasites were confirmed by microscopy and PCR amplification of a 655 bp GFP sequence (forward: 5’-TTCACTGGAGTTGTCC-3’; reverse: 5’-AGTTCATCCATGCCAT-3’) and 772 bp Neor sequence (forward: 5’-ATGGGATCGGCCATT-3’; reverse: 5’-TCAGAAGAACTCGTCAAG-3’) using Taq DNA polymerase (GenScript) per manufacturer protocol. To generate trypomastigotes, 1 mL of stationary phase epimastigotes was added to a T75 flask of confluent LLcMK2 cells. The medium was changed to fresh D2 every day, and after 1 week newly emerging trypomastigotes were collected and used to start new mammalian stage cultures. Constitutive GFP expression in the parasite population was confirmed by routine fluorescence microscopy (Nikon TE-300).
NHDF were plated at 1.5 x 106 in T75 flasks and infected with multiplicity of infection (MOI) of 10 for 18–24 h. At 48 hpi, monolayers were scraped and amastigotes were released from disrupted host cells by syringe passage (281/2G needle; BD) into the indicated ice-cold buffer. Amastigotes were purified from debris by passage through a PD-10 desalting column (GE Healthcare Life Sciences), and fractions containing clean amastigotes were centrifuged at 4000 g for 10 minutes at 4°C to pellet amastigotes, which were resuspended in warm (37°C) buffer as needed for indicated applications.
NHDF were plated and infected as for mammalian glucose uptake assays (described below). At 48 hpi genomic DNA was isolated using the DNeasy kit (Qiagen) and eluted in water. 1 μL of sample was combined with 10 μL of iTaq Universal SYBR Green Supermix (Bio-Rad) and 5 μM of each human TNF primer (forward: 5’-TAAGATCCCTCGGACCCAGT-3’; reverse: 5’-GCAACAGCCGGAAATCTCAC-3’) in a 20 μL reaction, run as above, and analyzed using the default Standard Curve (absolute quantitation) settings of a StepOnePlus.
All transport assays were performed using 1,2-3H(N)-2-deoxyglucose, ([3H]-2-DG; PerkinElmer).
The wells of an XFe24 cell culture microplate (Agilent Technologies) were coated with 0.1% gelatin and incubated at 37°C for 1 hour before gelatin was aspirated and NHDF were plated at 1.5 x 104 in 250 μL. Cells were infected with MOI of 50 for 1 hour then placed in 250 μL D-2. At 48 hpi media was changed to Seahorse XF base medium without phenol red (DMEM-based medium; Agilent Technologies), supplemented with 2 mM glucose and 10 mM glutamine, by rinsing cells twice with 1 mL medium, then replacing it with 500 μL medium. After 1 hour of incubation at 37°C, the supernatant was collected and 2 μL of sample was assayed by lactate assay kit I (Biovision) per manufacturer protocol.
NHDF or C2C12 stably expressing mCherry in the mitochondria were plated at 1.5 x 105 per well and infected with MOI of 20–40 GFP-expressing parasites for 18 h or treated with the indicated concentration of valproic acid starting from the time of infection. All centrifugation steps were carried out at room temperature at 350 g for 5 minutes. At 48 or 66 hpi, infected cells were trypsinized and centrifuged, then resuspended in 4% paraformaldehyde in PBS to fix on ice for 20 minutes. Cells were recentrifuged and the pellet was permeabilized in PBS + 0.1% Triton X-100 (PBST) with 0.5 μg/mL DAPI. Samples were run on a MACSQuant VYB and data was analyzed using FlowJo 7.6. Cells were discriminated based on size and DAPI staining. Infected cells were identified based on GFP expression (S2D Fig), and mCherry expression was examined for all populations. At least 10,000 events were collected per condition.
Flow cytometric analysis of an endogenous mitochondrial marker, was carried out using antibodies to ATP5B (Abcam #3D5) to stain infected and uninfected NHDF (without mCherry) as above. Following permeabilization for 20 min at room temperature, cells were centrifuged and resuspended in blocking solution (1% bovine serum albumin in PBS) and incubated for 30 minutes at room temperature. After centrifugation, cells were incubated for 30 minutes at room temperature with mouse anti-ATP5B (Abcam #3D5) at 1:2,000 in blocking solution then washed twice in PBST and incubated in 1:2,500 AlexaFluor 594 Goat anti-Mouse IgG (Thermo Fisher Scientific) in blocking solution for 30 minutes, washed as with primary antibody and resuspended in PBST with 0.5 μg/mL DAPI for flow cytometry analysis as above.
NHDF were seeded at 1.5 x 106 in T75 flasks and infected with a MOI of 10 for 2 hours. At 48 hpi, cells were scraped, pelleted by centrifugation and resuspended in Laemmli SDS-PAGE sample buffer at 2 x 104 cells/μL. Cell lysates were treated with 0.5 μL benzonase to digest DNA, incubated on ice 30 minutes, then at 95°C for 10 minutes before clarification by centrifugation at 15,000 g for 10 minutes. Due to the presence of parasites in infected cells compared to mock-infected cells, we could not load gels based on protein quantification. Instead, 2 x 105 cells (~30 μg of uninfected cells) were loaded in wells of a 4–15% Mini-Protean TGX Precast Gel (Bio-Rad) and the Western gel and wet transfer to a PVDF membrane was done per manufacturer protocol. All incubation and wash steps were done while shaking. The membrane was blocked in 1:1 SeaBlock:PBS (Thermo Fisher Scientific) for 1 hour at room temperature, incubated overnight at 4°C with the corresponding antibodies in 1:1 SeaBlock:PBS-0.1% Tween-20, then washed three times in PBS-0.1% Tween-20 prior to incubation with secondary antibodies in 1:1 SeaBlock:PBS-0.1% Tween-20 for 30 min at room temperature. The membrane was washed 3 times briefly in PBS-0.1% Tween-20 and 3 times in PBS for 10 minutes before imaging using an Odyssey Imaging System (Li-cor). The intensity of the signal for each antibody was assessed by Image Studio software (version 5.2). Lack of cross-reactivity of the antibodies with amastigote proteins was verified using amastigotes lysate following amastigote purification. Vimentin was used as a loading control. Primary antibodies: TOMM20 (F-10, Santa Cruz; 1:1,000), ATP5B (3D5, Abcam; 1:1,000), vimentin (5741, Cell Signaling; 1:2,000). Secondary antibodies: Dylight 800 anti-rabbit (Invitrogen, 1:10,000), Dylight 800 anti-mouse (Invitrogen, 1:5,000), Dylight 700 anti-rabbit (Invitrogen, 1:20,000).
To determine the number of divisions that an intracellular T. cruzi amastigote has undergone in a defined period of time, a modified flow cytometry protocol, based on [34] was performed. Briefly, NHDF were plated at 1.5 x 105 per well in 6 well plates and infected with MOI of 15 for 2 hours using CFSE-stained trypomastigotes. For staining, 5 x 106 trypomastigotes/mL were stained with 1 μM CFSE (Thermo Fisher Scientific) in PBS by incubating at 37°C for 15 minutes. Extra dye was quenched by addition of D-10, and trypomastigotes were pelleted by centrifugation at 2100 g for 10 minutes and incubated in fresh D-10 at 37°C for 30 minutes before infection. At 18 (pre-replication) and 48 (replicative phase) hpi, infected monolayers were trypsinized, washed once in PBS, and cells were lysed to release amastigotes by passing the supernatant 10 times through a 281/2G needle. Lysate was fixed by adding paraformaldehyde (Electron Microscopy Sciences) to a final concentration of 1% and incubating 20 minutes on ice. Samples were centrifuged at 300 g for 5 minutes to pellet away host debris, and the supernatant centrifuged at 4000 g for 10 minutes to pellet amastigotes. Pellets were resuspended in PBS with 0.1% Triton X-100 and 0.01 μg/mL DAPI for analysis by flow cytometry. Amastigotes were run on a MACSQuant VYB (Miltenyi Biotec) or LSRII (BD Biosciences) and at least 10,000 events were collected per condition. Data was analyzed using FlowJo 7.6, and amastigotes were discriminated based on size and DAPI staining. Proliferation was modeled using FlowJo 7.6, and CFSE intensity at 18 hpi was set as peak 0 for all samples.
Host cell and intracellular T. cruzi amastigote numbers were assessed as described with minor modifications [34]. Briefly, NHDF were plated at 1.5 x 103 per well in 384 well plates and infected with MOI 1.25 for 2 hours before incubation in phenol-free media at the indicated glucose concentration. At 18 hpi, the indicated concentration of 2-DG was added, and at 66 hpi media was removed and host cell number and parasite number were assessed using 10 μL of CellTiter-Fluor (Promega) and 10 μL of Beta-Glo (Promega) per well, respectively.
Isolated amastigotes were resuspended in cytobuffer at 2 x 107 parasites/mL and incubated for 3 hours at 37°C with 6 mM U-13C-glucose (Cambridge Isotope Laboratories) or unlabeled glucose (Sigma Aldrich). For metabolite extraction, amastigotes were rapidly cooled to 4°C in a dry-ice ethanol bath with gentle agitation as described [62], then centrifuged at 3200 g for 10 minutes at 4°C and resuspended in 80% (v/v) methanol:water to extract metabolites from the pellet as described [63]. Samples were run in technical triplicate and metabolites detected by the Beth Israel Deaconess Medical Center Mass Spectrometry Facility as described [64–66]. The percent of label incorporation was calculated for each replicate as the peak area of all 13C-labeled variants of a metabolite divided by the sum of both the labeled- and unlabeled-metabolite peak areas. Background was subtracted by averaging the percent label incorporation of unlabeled-replicates and subtracting that value from each labeled replicate.
Amastigotes were isolated in KHB and incubated at 4 x 105 amastigotes/mL in 5 mM glucose, 5mM glutamine, or 1 mM pyruvate as indicated at 37°C. Total ATP content was measured in isolated T. cruzi amastigotes under each condition at 0 hr and 24 hr using the ATPlite assay (PerkinElmer) following the manufacturer’s protocol.
RNA was extracted from either trypomastigotes or purified amastigotes (48 hpi) using the RNeasy purification kit (Qiagen). Using 500ng RNA, cDNA was generated through the iScript cDNA Synthesis Kit (Bio-Rad). Primer sets used for amplification of TcHT were TcHT-F: 5’-TGATGTACCATGTGTCCTCGGCAACG-3’ and TcHT-R 5’-ATGGCACTGCGCTGGACCCGA-3’ [15] or TcHT-F2: 5’-TCCTTCGTGCTCCTGACGAATT-3’ and TcHT-R2: 5’-AAAAGATGAACGCGACTGCCTG-3’. The primer set for amplification of a parasite housekeeping, ribosomal RNA large subunit gamma M1 was Ribo-F: 5’ -TGTGGAAATGCGAAACAC-3’ and Ribo-R: 5’-CCCAGGTTTTTGCTTTAATG-3’ [12]. Thermal cycling proceeded at 95°C for 10 minutes followed by forty cycles of 95°C 15 seconds and 60°C 1 minute using a StepOnePlus Real-Time PCR System (Applied Biosystems). Relative TcHT abundance was measured using SYBR green iTaq Universal Mix (Bio-Rad) and calculated using the ribosomal control and ΔΔCt method.
NHDF were seeded at 5 x 104 per well in 24 well plates on 12 mm round German glass coverslips (Electron Microscopy Services) and infected with MOI of 10 for 1 hour on glass coverslips. At 48 and 66 hpi, coverslips were fixed in 4% paraformaldehyde in PBS overnight at 4°C, then stained with 2.5 μg/mL DAPI and mounted on slides in Mowiol mounting medium. Slides were imaged using a Nikon TE300 and the number of amastigotes per cell was counted for 100 cells per condition.
Figures presented show mean values with standard deviation of biological replicates or medians (non-parametric data). Independent experiments were compared where indicated. Comparison of more than two groups was performed using a One-way ANOVA for single factor experiments and Two-way ANOVA for comparisons with two independent variables. The non-parametric Kruskal-Wallis test was used for comparisons of more than two groups that did not have normal distributions. If significant, post hoc tests were used (p values indicated) to compare specific groups and correct for multiple comparisons between groups. Statistical analysis was performed using Prism 7 (GraphPad).
|
10.1371/journal.ppat.1007265 | Exclusive dependence of IL-10Rα signalling on intestinal microbiota homeostasis and control of whipworm infection | The whipworm Trichuris trichiura is a soil-transmitted helminth that dwells in the epithelium of the caecum and proximal colon of their hosts causing the human disease, trichuriasis. Trichuriasis is characterized by colitis attributed to the inflammatory response elicited by the parasite while tunnelling through intestinal epithelial cells (IECs). The IL-10 family of receptors, comprising combinations of subunits IL-10Rα, IL-10Rβ, IL-22Rα and IL-28Rα, modulates intestinal inflammatory responses. Here we carefully dissected the role of these subunits in the resistance of mice to infection with T. muris, a mouse model of the human whipworm T. trichiura. Our findings demonstrate that whilst IL-22Rα and IL-28Rα are dispensable in the host response to whipworms, IL-10 signalling through IL-10Rα and IL-10Rβ is essential to control caecal pathology, worm expulsion and survival during T. muris infections. We show that deficiency of IL-10, IL-10Rα and IL-10Rβ results in dysbiosis of the caecal microbiota characterised by expanded populations of opportunistic bacteria of the families Enterococcaceae and Enterobacteriaceae. Moreover, breakdown of the epithelial barrier after whipworm infection in IL-10, IL-10Rα and IL-10Rβ-deficient mice, allows the translocation of these opportunistic pathogens or their excretory products to the liver causing organ failure and lethal disease. Importantly, bone marrow chimera experiments indicate that signalling through IL-10Rα and IL-10Rβ in haematopoietic cells, but not IECs, is crucial to control worm expulsion and immunopathology. These findings are supported by worm expulsion upon infection of conditional mutant mice for the IL-10Rα on IECs. Our findings emphasize the pivotal and complex role of systemic IL-10Rα signalling on immune cells in promoting microbiota homeostasis and maintaining the intestinal epithelial barrier, thus preventing immunopathology during whipworm infections.
| The human gut is home to millions of bacteria, collectively called the microbiota, and also to parasites that include whipworms. The interactions between gut cells, the microbiota and whipworms define conditions for balanced parasitism. Cells lining the gut host whipworms but also interact with gut immune cells to deploy measures that control or expel whipworms whilst maintaining a barrier to prevent microbial translocation. Whipworms affect the composition of the microbiota, which in turn impacts the condition of the gut lining and the way in which immune cells are activated. In order to avoid tissue damage and disease, these interactions are tightly regulated. Here we show that signalling through a member of the IL-10 receptor family, IL-10Rα, in gut immune cells is critical for regulating of these interactions. Lack of this receptor on gut immune cells results in persistence of whipworms in the gut accompanied by an uncontrolled inflammation that destroys the gut lining. This tissue damage is accompanied by the overgrowth of members of the microbiota that act as opportunistic pathogens. Furthermore, the destruction of the gut barrier allows these bacteria to reach the liver where they cause organ failure and fatal disease.
| A single layer of intestinal epithelial cells (IECs) in conjunction with the overlaying mucus acts as a primary barrier to viruses, bacteria and parasites entering the body via the gastrointestinal tract [1]. Paradoxically, the intestinal epithelium is also the host tissue for diverse pathogens including intestinal parasitic worms [2, 3]. Amongst the intestinal worms, whipworms (Trichuris trichiura) infect hundreds of millions of people and cause trichuriasis, a major Neglected Tropical Disease [4, 5].
Whipworms live preferentially in the caecum of their host, where they tunnel through IECs and cause inflammation that potentially results in colitis [6, 7]. It has been proposed that IEC activation, resulting from the initial recognition or physical contact with whipworms, influences the immunological response that ultimately determines whether parasites are expelled from the intestine or persist embedded in the intestinal epithelium causing a chronic disease [2, 4]. Most of our understanding of the host response to whipworms comes from studies of the natural whipworm infection of mice with T. muris, which closely mirrors that of humans [3, 7]. Resistance to infection is recapitulated by infecting mice with a high dose (200–400) of T. muris eggs and is mediated by a type-2 immune response that includes increased production of interleukin 4 (IL-4), IL-13, IL-25, IL-33, IL-9 and antibody isotypes IgG1 and IgE and results in worm expulsion [3, 7]. Conversely, chronic disease is modelled by infecting mice with a low dose (20–25) of T. muris eggs that results in the development of a type-1 immune response characterised by production of inflammatory cytokines, mainly IFN-γ, and the antibody isotype IgG2a/c [3, 7]. Type-1 immunity promotes intestinal inflammation that when severe can cause colitis [3, 7]. However, in chronic infections such responses are modulated by the parasite to optimize their residence and reproduction and ensure host survival, thus achieving a balanced parasitism [4, 7]. This immunomodulation is partly mediated by transforming growth factor (TGF)-β, IL-35 and IL-10 production by macrophages and T cells in response to excretory-secretory (ES) parasite antigens [3, 4, 7]. Besides this immunomodulatory role of IL-10 in chronic infections, IL-10 is important in the induction of host resistance (through type-2 response) during acute (high dose) T. muris infections [3, 8].
Intestinal mucosal homeostasis is regulated principally through IL-10 receptor signalling [9]. The IL-10 receptor is a heterotetramer complex composed of two alpha and two beta subunits, IL-10Rα and IL-10Rβ, respectively [9, 10]. While the IL-10Rα subunit is unique to IL-10, the IL-10Rβ chain is shared by receptors for other members of the IL-10 superfamily of cytokines [9–12]. Specifically, a single IL-10Rβ subunit pairs with IL-22Rα, IL-20Rα, or IL-28Rα subunits to form the receptors for IL-22, IL-26 and the interferon λ species (IL-28α, IL-28β and IL-29), respectively [10–12] (S1 Fig).
IL-10 is a key anti-inflammatory cytokine that limits innate and adaptive immune responses [9, 10]. The development of spontaneous enterocolitis in mice deficient for IL-10 and IL-10Rβ has demonstrated the crucial role of IL-10 in maintaining the integrity of the intestinal epithelium [13, 14]. Similarly, IL-22 contributes to the homeostasis of mucosal barriers by directly mediating epithelial defence mechanisms that include inducing the production of antimicrobial peptides, selected chemokines and mucus. IL-22 is also involved in tissue protection and regeneration [10, 12, 15]. The IL-22 receptor is exclusively expressed on non-haematopoietic cells, such as IECs [10, 12, 15]. Likewise, IL-28 receptor expression is largely restricted to cells of epithelial origin, although also expressed in B cells, macrophages and plasmacytoid DCs, where it mediates the antiviral, antitumor and potentially antibacterial functions of the interferon λ species [10, 16–18]. IL-26 is also reported to promote defence mechanisms against viruses and bacteria at mucosal surfaces in humans, however, the IL-26 receptor in the mouse is an orphan receptor because the Il-26 gene locus is not present in mice [10, 11, 19].
Previous studies indicate the importance of the IL-10 receptor signalling in responses to whipworms. Specifically, IL-10 promotes host resistance and survival to whipworm infection, with IL-10 deficiency leading to morbidity and mortality that may be due to a breakdown of the epithelial barrier and the outgrowth of opportunistic bacteria [8, 20]. Mice lacking the IL-10Rα chain develop a chronic T. muris infection accompanied by intestinal inflammation [21]. Furthermore, in IL-22-deficient mice whipworm expulsion is delayed, correlating with reduced goblet cell hyperplasia [22]. However, the specific role that each subunit (IL-10Rα, IL-10Rβ, IL-22Rα and IL-28Rα) plays on the intestinal epithelia barrier maintenance, mucosal homeostasis and broader host response to this parasite remains unclear. There is also little understanding on how these receptors can promote resistance to colonisation by opportunistic members of the microbiota that potentially drive the pathology observed in the absence of IL-10 during whipworm infection.
In the present study, we use mutant mice to dissect the role of IL-10Rα, IL-10Rβ, IL-22Rα and IL-28Rα in host resistance to T. muris infections. We demonstrate that IL-10 signalling, exclusively through IL-10Rα and IL-10Rβ, promotes resistance to colonization by intestinal opportunistic bacterial pathogens and maintenance of the intestinal epithelial barrier, thus preventing the development of systemic immunopathology during whipworm infection.
The care and use of mice were in accordance with the UK Home Office regulations (UK Animals Scientific Procedures Act 1986) under the Project licenses 80/2596 and P77E8A062 and were approved by the institutional Animal Welfare and Ethical Review Body. All efforts were made to minimize suffering by considerate housing and husbandry. Animal welfare was assessed routinely for all mice involved. Mice were naïve prior the studies here described.
Il10-/- and Il10ra-/- mice in a C57BL/6J background were obtained by treatment of Il10 fl/fl and Il10ra fl/fl [21] embryos with cre recombinase. Il10rafl/fl Vilcre/+ mice were obtained by crossing of Il10rafl/fl with Vilcre/+ mice. Il22-/- mice, as previously described [23], were received from Prof. Fiona Powrie (University of Oxford).
Il10rbtm1b/tm1b, Il22ra1tm1a/tm1a, Ifnlr1tm1a/tm1a, Rag1tm1Mom and wild-type (WT) C57BL/6N mice were maintained and phenotyped by the Sanger Mouse Genetics Programme [24]. For experiments with Il10-/-, Il10ra-/-, Il10rbtm1b/tm1b and Il10rafl/fl Vilcre/+ colonies, both WT and mutant mice littermates were derived from heterozygous breeding pairs. All animals were kept under specific pathogen-free conditions, and colony sentinels tested negative for Helicobacter spp. Mice were fed a regular autoclaved chow diet (LabDiet) and had ad libitum access to food and water.
Recipient mice were irradiated with two 5-Gy doses, 4 h apart, and injected intravenously with bone marrow harvested from donor mice at 2 million cells per 200 μl sterile phosphate-buffered saline. The mice were transiently maintained on neomycin sulfate (100mg/L, Cayman Chemical) in their drinking water for 2 weeks (wk). Bone marrow was allowed to reconstitute for 4 wk before mice were infected with T. muris.
Infection and maintenance of T. muris was conducted as described [25]. Age and sex matched female and male WT and mutant mice (6–10 wk old) were orally infected under anaesthesia with isoflurane with a high (400) or low (20–25) dose of embryonated eggs from T. muris E-isolate. Mice were randomised into uninfected and infected groups using the Graph Pad Prism randomization tool. Uninfected and infected mice were co-housed. Mice were monitored daily for general condition and weight loss. Mice were culled including concomitant controls (uninfected and WT mice) at different time points or when their condition deteriorated (observation of hunching, piloerection, reduced activity or weight loss from body weight at the beginning of infection reaching 20%). Mice were killed by terminal anesthesia followed by exsanguination and cervical dislocation. The worm burden was blindly assessed by counting larvae that were present in the caecum. Blinding at the point of measurement was achieved by the use of barcodes. During sample collection, group membership could be seen, however this stage was completed by technician staff with no knowledge of the experiment objectives.
Every other day, from day 35 to day 45 p.i., whipworm-infected Rag1tm1Mom mice were intraperitoneally injected with an antibody blocking the IL-10Rα (BioXcell, clone 1B1.3A) or an isotype control (BioXcell, clone HRPN) for a total delivery of 2mg per mouse.
Adult worms were cultured in RPMI 1640 (Sigma-Aldrich) and ES products were collected after 4 h and following overnight culture. The ES were prepared as described [26].
To evaluate disease pathology, caecal and liver segments were fixed in 4% paraformaldehyde and 2–5 μm paraffin sections were stained in haematoxylin and eosin (H&E) or Periodic Acid-Schiff (PAS) according to standard protocol. Slides were scanned using a Hamamatsu NanoZoomer 2.0HT digital slide scanner (Meyer Instruments, Inc) and images were analysed using the NDP View2 software. From blinded and randomised histological slides, intestinal inflammation was scored by two research assistants as follows: submucosal and mucosal oedema (0, absent; 1, mild; 2, moderate; or 3, severe); submucosal and mucosal inflammation (0, absent; 1, mild; 2, moderate; or 3, severe); percentage of area involved (0, 0–5%; 1, mild, 10–25%; 2, moderate, 30–60%; or 3, severe, >70%). Crypt lengths were measured and goblet cells counted. Liver pathology was documented, including presence of immune infiltrate, granulomas and necrosis.
For immunofluorescence, 5 μm sections of frozen caecal and liver tissues were stained with α-Enterococcus spp. antibody (1/1000, LSBio), α-Escherichia coli spp. antibody (1/1000, LSBio) or α-ZO-1 (1/200 ThermoScientific). Sections were mounted using ProLong Gold anti-fade reagent (Molecular Probes) containing 4’,6’-diamidino-2-phenylindole (DAPI) for nuclear staining. Confocal microscopy images were taken with a Leica SP8 confocal microscope. From each mouse, a slide was examined to determine the presence of bacterial infection in the liver. Although only low sensitivity is possible from a single slide per mouse, bacteria should not be present in the livers of healthy animals, and any detected therefore indicate bacterial translocation.
For transmission electron microscopy, tissues were fixed in 2.5% glutaraldehyde/2% paraformaldehyde, post-fixed with 1% osmium tetroxide in 0.1M sodium cacodylate buffer and mordanted with 1% tannic acid followed by dehydration through an ethanol series (contrasting with uranyl acetate at the 30% stage) and embedding with an Epoxy Resin Kit (Sigma-Aldrich). Ultrathin sections cut on a Leica UC6 ultramicrotome were contrasted with uranyl acetate and lead nitrate, and images recorded on a FEI 120 kV Spirit Biotwin microscope on a F415 Tietz CCD camera.
Levels of parasite-specific immunoglobulins IgG1 and IgG2a/c were determined by ELISA in serum as described [27]. Briefly, ELISA plates (Nunc Maxisorp, Thermo Scientific) were coated with 5 μg/ml T. muris overnight-ES. Serum was diluted from 1/20 to 1/2560, and parasite-specific IgG1 and IgG2a/c were detected with biotinylated anti-mouse IgG1 (Biorad) and biotinylated anti-mouse IgG2a/c (BD PharMingen), respectively.
Serum IL-6 and TNF-α were determined with the Mouse IL-6 and TNF-α ReadySet-Go! ELISA kits (eBioscience).
The presence of lipopolysaccharide (LPS) in serum was determined with the LAL assay kit (Hycult Biotech).
Blood was collected under terminal anaesthesia into heparinized tubes for plasma preparation. Within 1 hour of collection, blood samples were centrifuged and plasma recovered and stored at -20°C until analysis. Clinical chemistry analysis of plasma was performed using the Olympus AU400 analyzer (Beckman Coulter Ltd) and was blinded to the operator via barcodes.
The majority of parameters were measured using kits and controls supplied by Beckman Coulter. Samples were also tested for haemolysis. Four parameters were measured by kits not supplied by Beckman Coulter: transferrin, ferritin (Randox Laboratories Ltd), fructosamine (Roche Diagnostic) and thyroxine (Thermo Fisher).
To identify microbial species from the livers of mice, mouse tissues were homogenized aseptically under laminar flow. Organ lysates were immediately cultured in nonselective Luria-Bertani (LB) and Brain Heart Infusion (BHI) media under aerobic and anaerobic conditions for 36–48 h. All colonies from each plate, or within a defined section, were picked in an unbiased manner for DNA extraction and 16S rRNA gene sequencing using the universal primers: 7F, 50-AGAGTTTGATYMTGGCTCAG-30; 926R, 50-ACTCCTACGGGAGGCAGCAG-30. Bacterial identifications were performed using the 16S rRNA NCBI Database for Bacteria and Archaea.
To study the caecal and liver microbiota composition of uninfected and T. muris-infected mice, luminal contents of the caecum were collected by manual extrusion and a piece of liver was taken upon culling of mice. Bacterial DNA was obtained using the FastDNA Spin Kit for Soil (MBio) and FastPrep Instrument (MP Biomedicals). V5-V3 regions of bacterial 16S rRNA genes were PCR amplified with high-fidelity AccuPrime Taq Polymerase (Invitrogen) and primers: 338F, 50-CCGTCAATTCMTTTRAGT-30; 926R, 50-ACTCCTACGGGAGGCAGCAG-30. Libraries were sequenced on an Illumina MiSeq platform according to the standard protocols. Analyses were performed with the Quantitative Insights Into Microbial Ecology 2 (QIIME2-2018.4; https://qiime2.org) software suite [28], using quality filtering and analysis parameters as described in the Supplemental Experimental Procedures.
For all analyses, the individual mouse was considered the experimental unit within the studies. Experimental design was planned using the Experimental Design Assistant [29]. A multi-replica design was used, where each replica was run completely independently. Within each replica there were concurrent controls of infected and non-infected animals. The number of animals for each genotype within a replica varies as it was constrained by the outcome of breeding.
The effect of genotype on worm burden within infected mice was assessed across multiple replicas using an Integrative Data Analysis (IDA) [30] treating each replica as a fixed effect utilising the generalised least square regression function within the nlme version 3.1 package of R (version 3.3.1). A likelihood ratio test was used to test for the role of genotype by comparing a test model (Eq 1) against a null model (Eq 2). As genotype was found to be highly significant in explaining variation, a F ratio test for Eq 1 was used to explore the role of genotype as a main effect and whether it interacted with sex. The effect of genotype was not found to interact with sex (p>0.05).
The effect of gene knockout on worm burden was assessed for each sex separately using a Mann Whitney U test from the Prism 7.0 software (GraphPad). This analysis pools data across replicas as the IDA analysis found that this was not a significant source of variation. A non-parametric test was used as the data is bound and has some non-normal distribution characteristics. Similarly, cytokine levels between infected WT and mutant mice and plasma chemistry parameters between infected isotype and IL-10Rα -treated Rag1tm1Mom mice were evaluated using a Mann Whitney U test from the Prism 7.0 software (GraphPad).
The survival data, pooled across replicas, was tested for a significant effect of gene knockout for each sex independently using Log-rank Mantel-Cox tests from the Prism 7.0 software (GraphPad).
To evaluate the effect of the degree of colonization by pathobionts on the survival of infected IL-10 signalling-deficient mice, the mice were classified based on the histological assessment of the liver and survival time (time after infection that mice succumbed), into two groups: severe (presence of granulomas necrosis and foamy macrophages in the liver, weight loss and poor survival) and mild (minor liver infiltration, no weight loss and extended survival). Next, the degree of colonization by Enterococcus, Enterobacteriaceae and Escherichia-Shigella (combined percentage of abundance of these three groups from total microbiota) among both groups was compared using a Mann Whitney U test from the Prism 7.0 software (GraphPad). In addition, a correlation analysis of the degree of colonization of the pathobionts and the plasma levels of the liver enzyme aspartate aminotransferase was performed using a two-tailed Spearman correlation test.
A similar IDA analysis was used to study the effect of genotype on infection, for each plasma chemistry variable across multiple replicas. In this IDA, a likelihood ratio test was used to test for an interaction between genotype and infection by comparing a test model (Eq 3) against a null model (Eq 4). This regression model was fitted to separate the various sources of variation allowing the impact of genotype in the presence of infection to be estimated.
P-values were adjusted for multiple testing using the Benjamini and Hochberg method [31] with a false discovery rate of 5%. Percentage change was calculated to allow comparison of the effect across variables by taking the estimated coefficient from the regression analysis and dividing it by the average signal seen for that variable. The model estimates without normalisation are presented in S4 Table.
The effect of genotype and infection on caecum score and goblet cells per crypt was assessed across the multiple replicas using an IDA as described for the plasma chemistry variables.
For all IDAs, the model fit was assessed by visual exploration of the residuals with quantile-quantile and residual-predicted plots for each genotype group.
To dissect the role of the members of the IL-10 family of receptors during whipworm infection, mouse lines with knockout mutations for the following loci were challenged with T. muris: Il10, Il10ra, Il10rb, Il22, Il22ra and Il28ra (S1 Fig). The influence of these mutations on anti-parasite immunity and worm expulsion was evaluated. Like WT mice, a high dose infection with T. muris did not result in chronic infection of IL-22, IL-22Rα and IL-28Rα mutant mice; after 32 days of infection, the mice had expelled all worms and had high levels of parasite specific IgG1 in their sera that indicated the development of a type-2 response (S2A, S2B and S2C Fig). Moreover, worm expulsion occurred before day 21 post infection (p.i.), accompanied by production of T. muris specific IgG1 (S3A, S3B and S3C Fig). These results are contrary to previous reports describing delayed worm expulsion in IL-22 mutant mice at day 21p.i. [22].
Using low dose infections, at day 35 p.i., there were also no differences between WT and IL-22, IL-22Rα and IL-28Rα mutant mice in the establishment of a chronic infection that is characterized by high levels of parasite specific IgG2a/c in serum (S4A, S4B and S4C Fig). These findings indicated that IL-22 and IL-28 signalling are dispensable for the host to mount a response to T. muris infection. When taken together with previous data, these results suggest that the IL-10 receptor is the only member from this family of receptors responsible for the control of host resistance and survival to whipworm infection.
We then examined the contribution of IL-10 signalling to the responses to T. muris infection. IL-10, IL-10Rα and IL-10Rβ mutant mice were infected with a high dose of eggs and survival, tissue histopathology and worm burdens throughout infection up to day 28 p.i. were evaluated. We used WT littermate controls that were co-housed with the mutant mice throughout the experiments. Moreover, we included uninfected WT and mutant mice as additional controls in the cages. IL-10, IL-10Rα and IL-10Rβ mutant mice did not develop spontaneous colitis in our mouse facility.
As previously reported [8], female and male IL-10 mutant mice succumbed to whipworm infection between day 19 and 24 p.i., showing a dramatic weight loss and high numbers of worms in the caecum when compared with WT mice (Fig 1A). Similarly, female and male IL-10Rα mutant mice displayed weight loss and all required euthanasia by day 28p.i. concomitant with high worm burdens in the caecum (Fig 1B). Although the defects in the expulsion of worms in IL-10Rα mutant mice have been described [21], this is the first report of reduced survival of these mice upon whipworm infection. Likewise, high numbers of worms were present in the caecum of IL-10Rβ mutant mice and survival was reduced by 60% and 75% in females and males, respectively (Fig 1C). Defective worm expulsion and survival in T. muris-infected IL-10, IL-10Rα and IL-10Rβ mutant mice correlated with increased inflammation in the caecum (Fig 2). Specifically, while infected WT mice presented mild inflammation (Fig 2A and 2B) and goblet cell hyperplasia (Figs 2C and S5), a characteristic response to T. muris, infected IL-10 signalling-deficient mice displayed submucosal oedema, large inflammatory infiltrates in the mucosa with villous hyperplasia, distortion of the epithelial architecture (Fig 2A and 2B) and loss of goblet cells (Figs 2C and S5). Together, these results indicate that during T. muris infections, IL-10 signalling is crucial for controlling worm expulsion and caecal mucosal and submucosal inflammation leading to unsustainable pathology.
Reduced survival of whipworm-infected IL-10 signalling-deficient mice correlated with liver pathology, while we did not observe pathology in other systemic organs such as the spleen. Specifically, upon culling and dissection of T. muris-infected IL-10, IL-10Rα and IL-10Rβ mutant mice, we observed granulomatous lesions in their livers including necrotic areas and lymphocytic and phagocytic infiltrates (Fig 3). Moreover, 50, 25 and 12.5% of T. muris-infected IL-10, IL-10Rα and IL-10Rβ mutant mice, respectively, showed extensive numbers of foamy (lipid-loaded) macrophages in their livers (S6 Fig). Because survival upon whipworm infection is similarly reduced among IL-10, IL-10Rα and IL-10Rβ mutant mice but IL-10Rα is the only subunit that is exclusively used for IL-10 signalling, we focused subsequent experiments on IL-10Rα-deficient mice.
Liver disease was reflected in changes to plasma chemistry markers of liver damage. Compared to WT mice, T. muris-infected mice with defects in IL-10 signalling presented significantly dysregulated plasma levels of liver enzymes (decreased concentrations of alkaline phosphatase and increased concentrations of aspartate and alanine aminotransferase), accompanied by reduced concentrations of glucose, fructosamine, albumin and thyroxine (Figs 4A and S7). Upon whipworm infection, we also observed augmented levels of ferritin and transferrin, which are indicators of systemic infection, in IL-10 signalling-deficient, but not in WT mice (Figs 4A and S7). We found no or minimal differences in plasma chemistry between uninfected WT and mutant mice (S1, S2 and S3 Tables). The changes in plasma chemistry parameters between infected WT and mutant mice were accompanied by increased circulating concentrations of the inflammatory cytokines IL-6 and TNF-α (Figs 4B and 4C and S8).
Liver pathology appears to be caused by dissemination of gut bacteria or their products to the liver, upon breakdown of the caecal epithelial barrier due to whipworm infection and IL-10 signalling defects.
Outgrowth of opportunistic bacteria contributes to the mortality of IL-10 mutant mice during whipworm infection [8, 20]. Furthermore, intestinal inflammation can promote microbial dysbiosis and impair resistance to colonization by opportunistic pathogens [32, 33]. We hypothesised that infection of IL-10 signalling-deficient mice with whipworms caused caecal dysbiosis and the overgrowth of opportunistic bacteria from the microbiota. Thus, we analysed the microbiota composition of uninfected and T. muris-infected WT and IL-10, IL-10Rα and IL-10Rβ mutant mice using high-throughput 16S rRNA sequencing. No significant differences in overall gut microbial profiles and alpha/beta diversity were detected between uninfected IL-10 signalling-deficient and WT mice (S9, S10 and S11 Figs), thus indicating that IL-10 signalling did not impact caecal microbial community structure, an observation that is consistent with the lack of spontaneous inflammation in these mice in our mouse facility. Similarly, whipworm infection of WT mice did not lead to changes in overall microbial community structure and alpha/beta diversity, as shown by the lack of significant differences between the gut microbial profiles of infected and uninfected WT mice (S9, S10 and S11 Figs). Conversely, whipworm infection of IL-10Rα mutant mice resulted in a caecal microbial profile distinct from that of infected WT mice (p = 0.001, CCA, Fig 5A), but also of uninfected WT and mutant mice, as shown by both PCoA and CCA (S10A Fig). The observed changes in the caecal microbial community structure were associated with a significant increase in microbial beta diversity (i.e. differences in species composition between groups; p = 0.001, ANOSIM; Fig 5B) and a decrease in alpha diversity (i.e. species diversity within a group) (measured through Shannon diversity, p = 0.01, ANOVA; Fig 5C) in T. muris-infected IL-10Rα mutant mice when compared to WT and uninfected mice (S10B and S10C Fig). In particular, the observed decrease in alpha diversity of the caecal microbiota in T. muris-infected IL-10Rα mutant mice was associated with significant reductions of both microbial richness (i.e. the number of species composing the microbial community; p < 0.001, ANOVA; Figs 5C and S10C) and evenness (i.e. the relative abundance of each microbial species in the community; p < 0.001, ANOVA; Figs 5C and S10C). Network analysis identified a positive correlation between the presence and relative abundance of several opportunistic pathogens (i.e. Enterobacteriaceae, Escherichia/Shigella, Enterococcus, and Clostridium difficile), as well as lactic acid-producing bacteria (i.e. Lactobacillus), and the microbial profiles of T. muris-infected IL-10Rα mutant mice (Fig 5D). Moreover, analysis of differential abundance of individual bacterial taxa via Linear Discriminant Analysis Effect Size (LEfSe) revealed that Enterobacteriaceae, Enterococcaceae and Lactobacillaceae were significantly more abundant in infected mutant mice, than in any of the other mouse groups (LDA Score (log10) of 4.78, 4.77, and 4.44 respectively; Figs 5E and S10E). The increase in abundance of these groups in the T. muris-infected IL-10Rα mutant mice was also observed when comparing the relative OTU abundances (Fig 5F).
Similarly, T. muris-infected IL-10 and IL-10Rβ mutant mice presented a clear and consistent overgrowth of Enterobacteriaceae, Escherichia/Shigella and Enterococcus (S9 and S11 Figs). The degree of colonization by these pathobionts correlated with the reduced survival (time after infection that mice succumbed) and extent of liver disease observed (S12 Fig). Co-housing of the uninfected and T. muris-infected WT and mutant mice did not result in microbiota transfer by coprophagia as the dysbiosis and presence of pathobionts in the infected mutant mice was not observed in any other groups (S9, S10 and S11 Figs).
Altogether these results indicate that absence of IL-10 signalling during whipworm infection causes intestinal dysbiosis due to the overgrowth of facultative anaerobes, members of the microbiota that have been previously described as opportunistic pathogens [34–36]. Moreover, the presence of the parasite is critical to the development of the observed dysbiotic state.
We hypothesized that the opportunistic pathogens (or their products) present in the dysbiotic microbiota of the whipworm-infected IL-10 signalling-deficient mice were disseminating to the liver, thus causing lethal disease. To test this hypothesis, we examined whether a breakage of the epithelial barrier allowed bacteria from the Escherichia and the Enterococcus genera to translocate intracellularly through the caecal epithelia. Immunofluorescence labelling for the tight junction protein ZO-1, showed that in contrast to WT mice that have expelled the worms, and in which tight junctions are intact (Fig 6A), whipworm-infected IL-10Rα-deficient mice lacked tight junctions in regions of the caecal epithelia replaced with high inflammatory infiltrate (Fig 6B and 6D). Transmission electron microscopy of those regions clearly showed a neutrophilic infiltrate and the denuded epithelium (Fig 6E and 6F). Tight junctions of cells in close proximity to the worm are also lost (Fig 6C). Moreover, using transmission electron microscopy, we observed the presence of intracellular cocci and bacilli in the caecal epithelia of whipworm-infected IL-10 signalling-deficient, but not WT mice (Fig 7A). Immunofluorescence staining with antibodies against Escherichia spp. and Enterococcus spp. further indicated the intracellular translocation of these opportunistic pathogens through the enterocytes of the caecum of whipworm-infected IL-10 signalling-deficient mice (Fig 7B and 7C). These observations indicate that both Escherichia spp. and Enterococcus spp. invade the caecal epithelium of IL-10 signalling-deficient mice upon whipworm infection. Furthermore, they suggest that translocation of these opportunistic pathogens or their products to the liver are the potential cause of lethal liver disease observed in these mice.
Livers of T. muris-infected WT and IL-10 signalling-deficient mice were cultured under aerobic and anaerobic conditions to identify bacterial isolates using 16S rRNA sequencing. We found E. coli, E. faecalis and E. gallinarum in the livers of some whipworm-infected IL-10 signalling-deficient but not in WT mice (S5, S6 and S7 Tables). Moreover, using 16S rRNA PCR and sequencing, we found Escherichia/Shigella and Enterococcus were more abundant in whipworm infected IL-10Rα–deficient mice than in uninfected WT and IL-10Rα–deficient mice and for Enterococcus this was also significant when comparing with infected WT mice (Fig 8A–8C). Although systematically examining entire livers was not feasible, we did observe bacteria staining with the Escherichia spp. and Enterococcus spp. antibodies in the sections from livers of some whipworm-infected IL-10 signalling-deficient mice but not in whipworm-infected WT mice (Fig 8D and 8E).
In summary, our findings indicate that IL10 signalling, via the IL-10Rα and IL-10Rβ, promotes resistance to colonization by opportunistic pathogens and controls immunopathology preventing microbial translocation and lethal disease upon whipworm infection.
Conditional knockout mice lacking IL-10Rα on T cells and monocytes/macrophages/neutrophils did not recapitulate the phenotype of the complete mutant, thus suggesting that these cell types alone are not the main responders to IL-10 during whipworm infection [21]. This indicates that expression of IL-10Rα on other immune cells or IECs or in a combination of effector cells may be responsible for the IL-10 effects on worm expulsion and inflammatory control. To identify whether the main target cells of IL-10 were of haematopoietic or non-haematopoietic (epithelial) origin, we generated bone marrow chimeric mice by transferring either WT or IL-10Rα and IL-10Rβ–deficient bone marrow into lethally irradiated WT or IL-10Rα and IL-10Rβ–deficient mice and infected them with a high dose of T. muris. We observed decreased survival around day 20p.i. of 100% of irradiated WT mice reconstituted with bone marrow of IL-10Rα and IL-10Rβ mutant donors (Figs 9A and S13A), which was accompanied by caecal and liver pathology (Figs 9B and S13B). By contrast, WT mice receiving bone marrow cells from WT donors did not show any morbidity signs or caecal inflammation, even when worm expulsion was not always observed (Figs 9A and S13A). Conversely, reconstitution of irradiated IL-10Rα and IL-10Rβ mutant mice with WT donor bone marrow protected them from the unsustainable pathology caused by whipworm infection (Figs 9C and 9D and S13C and S13D). These results suggest that the main target cells responding to IL-10 are of haematopoietic origin.
To support these findings and overcome the limitations of bone marrow chimera mice that include incomplete immune system reconstitution and microbiota dysregulation, we generated conditional mutant mice for the IL-10Rα on IECs (Il10rafl/fl Vilcre/+) and infected them with a high dose of T. muris. Similar to WT controls (Il10ra+l+ Vilcre/+ and Il10rafl/fl Vil+/+), Il10rafl/fl Vilcre/+ mice expelled the worms as early as day 20 p.i. and developed a type 2 response indicated by the presence of specific parasite IgG1 antibodies in the serum (S14 Fig).
To further investigate which cells of haematopoietic origin are mediating the regulatory functions of IL-10 during whipworm infections, we infected RAG1-deficient mice, which lack the adaptive immune compartment (T and B cells), with a high dose of T. muris. These mice developed a chronic infection with no symptoms of immunopathology (Fig 10A and 10C). From day 35 to day 45 p.i., we treated the mice with an antibody blocking the IL-10Rα or an isotype as control. We observed no differences in survival, worm burdens, plasma chemistry parameters, caecal and liver pathology and bacterial translocation between both groups (Fig 10) indicating that cells of the innate immune compartment alone are not the drivers of immunopathology and in this setting IL-10 signalling plays no role.
Together, these findings suggest that expression of the IL-10 receptor on several immune cells types, is crucial in controlling the development of lethal liver disease due to dysbiosis and microbial translocation upon whipworm infection.
We have shown that upon infection with whipworms, signalling by IL-10, but not IL-22 or IL-28, is crucial for the resistance to colonization by opportunistic pathogens, control of host inflammation, intestinal barrier maintenance and worm expulsion. We dissected the contribution of the IL-10 cytokine and the subunits of its cognate receptor and observed that lack of any of the components resulted in the development of a chronic whipworm infection that led to unsustainable pathology, confirming previous reports [8, 20, 21] and extending the observations to deficiency of the IL-10Rβ chain.
During whipworm infection IL-10 signalling on cells of haematopoietic origin is critical for both the development of a type-2 response resulting in worm expulsion, and the control of type-1 immunity-driven inflammation and pathology. Specifically, IL-10 promotes type-2 responses [8, 37, 38] that are indispensable for IEC turnover to maintain epithelial integrity and goblet cell hyperplasia to increase the mucus barrier. Several important roles are played by this barrier: maintaining bacterial communities that compete against and prevent colonisation by inflammatory pathobionts [39, 40]; separating IECs from luminal bacteria; and expelling the worm through the direct action of mucins [3, 41]. In contrast, the absence of IL-10 signalling results in a type-1 inflammatory response [8, 37, 38] that fails to induce the mechanisms for worm expulsion and causes intestinal epithelium damage. Inflammation and worm persistence disrupts the intestinal microbiota, affecting colonization resistance and promoting the overgrowth of opportunistic pathogens. The disruption of the epithelial barrier allows these pathobionts or their products to translocate and reach the liver, where they cause inflammation and necrosis resulting in liver failure and leading to lethal disease (Fig 11).
The actions of IL-10 signalling on the control of type-2 and -1 responses during whipworm infections may depend on the timing, cell type and organ where IL-10 is produced and the receptor is expressed. Early in an infection (before day 15 p.i.), IL-10 signalling-deficient mice, infected with T. muris, lacked the type-2 response and goblet cell hyperplasia observed in WT mice [37, 38]. IL-10 signalling therefore contributes to worm resistance via development of type-2 responses in the caecum and mesenteric lymph nodes, ultimately resulting in worm expulsion in WT mice. At later stages of infection (day 21–28 p.i.), IL-10 signalling controls the type-1 driven pathology both in the caecum and the liver leading to reduced survival. At this time point, infected IL-10 signalling-deficient mice displayed higher levels of IFN-γ, IL-12, TNF-α and IL-17 and severe caecal and liver inflammation when compared with WT mice [8, 37]. Also treatment of chronically infected (low dose) WT mice after day 30 p.i. with a monoclonal antibody against IL-10R resulted in increased pathology and weight loss accompanied with increased production of type-1 cytokines [6].
Our results clearly demonstrate the haematopoietic origin of the cells that respond to IL-10 upon T. muris infection. In previous studies, IL-10Rα conditionally knocked out in mouse T cells, monocytes, macrophages and neutrophils did not result in inflammation or defects in worm expulsion [21]. These immune cell types alone are clearly not the main responders to IL-10. Our findings in RAG1-deficient mice indicate that cells of the innate immune compartment alone are not the drivers of immunopathology and there is no immunoregulation in these mice by IL-10. A remarkable observation is that extensive mechanical damage caused by the worm in the absence of adaptive immunity is insufficient to cause liver disease. This suggest that the immunopathology only develops in the presence of an adaptive immune system. Our findings highlight the complexity of the IL-10 immunoregulatory response and suggest that more than one cell type may be required to respond to IL-10 at different stages during infection.
Future studies including rescue experiments transferring defined populations of IL-10R competent cells in IL-10 signalling-deficient mice or using inducible conditional knockout mice at different times during infection will help to identify the critical IL-10 responsive cell(s). The IL-10-responding cells may be stimulated directly by the microbiota or whipworms, through pattern recognition receptors such as MyD88 [42], Nod2 [39] and Nlrp6 [43] or indirectly, by limiting the inflammatory responses of other cells.
Our findings are in agreement with the multi-hit model of inflammatory gut disease [44]: infection with whipworms is a colitogenic trigger that initiates the inflammatory process; lack of IL-10 signalling causes an inflammatory type-1 response that determines the dysregulation of the mucosal immune response; and the microbiota impacts the susceptibility and responses to infection. The dysbiosis that we observed during T. muris infections of mice lacking IL-10 or its receptor was characterized by an increase in the abundance of opportunistic pathogens from the Enterobacteriaceae family (Escherichia/Shigella) and Enterococcus genus. These facultative anaerobes occur in much lower levels in the microbiota than obligate anaerobes [45]. However, host-mediated inflammation resulting from an infection or genetic predisposition, such as mutations in IL-10 [32, 36, 46, 47], increases available oxygen. The higher oxygen tension benefits the growth of aerotolerant bacteria [35, 36], disrupting the intestinal microbiota and colonization resistance [32, 36, 46, 47]. Mice deficient in IL-10 signalling do not develop spontaneous inflammation and dysbiosis in our facility. Therefore, changes to the microbiota are directly attributable to the colonization of the intestine by whipworms. We did not observe transfer of microbiota by coprophagy (in particularly, members of the Enterobacteriaceae family and the Enterococcus genus) and subsequent colitis susceptibility in co-housed uninfected and infected mice of both WT and mutant strains. Similarly, no transfer of microbiota was observed in IL-10 mutant mice co-housed with Il10-/-Nlrp6-/- mice harbouring an expanded population of the pathobiont Akkermansia muciniphila [43]. Together, these results suggest that deficiency in IL-10 signalling alone is insufficient to trigger dysbiosis; whipworm infection is required to reach this disbalanced state.
We did not observe major changes to the microbiota in WT mice that cleared whipworm infections before d15 p.i. [48]. Nevertheless, the microbial alterations detected in IL-10 signalling-deficient mice, which develop chronic infections from a high-dose inoculum, were similar to those of chronically infected (low-dose) WT mice. These changes included decreased alpha diversity of the microbiota concomitantly with an increase in the abundances of Lactobacillus and Enterobacteriaceae (Escherichia/Shigella) [49, 50] and Enterococcus [49]. The changes in the microbiota seen during whipworm chronic infection are therefore conserved and occur more rapidly and drastically when type-1 immune responses are not regulated.
Increased abundance of Lactobacillus and Enterobacteriaceae has been also observed in the intestinal microbiota of Heligmosomoides polygyrus-infected susceptible mice [51, 52], and may indicate that helminth infections favour the establishment of certain bacterial groups and vice versa [49, 51, 53]. The significant reduction of bacteria of the genus Mucispirillum (family Deferribacteraceae) in the microbiota of whipworm-infected IL-10 signalling-deficient mice, is likely a consequence of the goblet cell loss, as these bacteria colonise the mucin layer of the gut [54]; indeed, Mucispirillum abundance increases during Trichuris infection of both pigs and mice [49, 50, 55], where goblet cell hyperplasia occurs.
Both Enterobacteriaceae (Escherichia/Shigella) and members of the Enterococcus genus such as E. faecalis are pathobionts that can cause sepsis-like disease when intestinal homeostasis is disrupted [32, 34, 35]. In whipworm-infected IL-10 signalling-deficient mice, we observed infiltration of neutrophils and macrophages in the intestinal epithelia and neutrophilic exudates in the lumen, potentially as a mechanism of clearance of these bacteria. Nevertheless, this inflammatory response results in tissue damage and bacteriolysis that induce immunopathology [56]. Tissue damage caused by the worm further increases inflammation and opens a door for opportunistic pathogens and their products to translocate through the intestinal epithelia. When immune cells (neutrophils and macrophages) fail to control the bacteria or their products in the intestine, these are drained by the portal vein into the liver [57–59]. Liver Kupffer cells located in the periportal area phagocytise antigens and microorganisms within the portal venous circulation [57–59] and promote anti-inflammatory responses mediated in part by IL-10 [57]. Lack of IL-10 signalling and translocation of opportunistic pathogens and their products to the liver may contribute to granulomatous inflammation and production of proinflammatory cytokines by Kupffer cells and infiltrating bone-marrow-derived-monocytes/macrophages resulting in failure of microbial clearance, tissue damage with consequent liver failure [58, 59] and lethal disease. We were able to isolate E. coli, E. faecalis and E. gallinarum from the livers of some mutant mice and also observed increased abundances of Enterococcus and Escherichia/Shigella 16S rRNA sequences in the livers of infected IL-10Rα mutant mice. Besides bacterial growth, liver pathology and disease could also be caused by bacterial metabolites and products such as LPS of Gram-negative bacteria and lipoteichoic acid (LTA) of Gram-positive bacteria, which are known triggers of sepsis [60].
Similarly, microbial translocation has been described during hookworm [61] and HIV [62] infections that result in intestinal epithelial damage and permeability. Moreover, microbial translocation also occurs during inflammatory bowel disease (IBD) [63–65], where intestinal inflammation and damaged barrier function results from a combination of factors, including dysbiosis and mutations in genes encoding proteins involved in the immune response, such as IL-10 [57]. We did not detect LPS in serum of whipworm-infected IL-10 signalling-deficient mice (with values below the sensitivity threshold of the assay), suggesting that either the pathobionts mediating the disease are Gram-positive and therefore, other microbial products, such as LTA and peptidoglycan, may be the cause of systemic immunopathology or that opportunistic pathogens and their products were confined to the liver where they cause liver failure and disease. Previous publications showing prolonged survival of whipworm infected IL-10-deficient mice conferred by antibiotic treatment clearly support the role of microbial translocation on liver pathology and lethal disease observed in these mice [8,20]. Nevertheless, our findings on RAG1-deficient mice showing bacterial translocation in the majority of the mice in the absence of pathology suggest that microbial translocation alone is not the cause of immunopathology, but requires the presence of the adaptive immune system to trigger tissue damage.
Liver damage was reflected in changes in plasma chemistry parameters in whipworm-infected IL-10 signalling-deficient mice. Specifically, decreased hepatic synthetic function (lower plasma albumin, hypoglycaemia) and release of liver aminotransferases into the circulation are the result of hepatocyte damage and liver necrosis [66, 67]. Low albumin and enhanced cellular uptake of thyroxine by phagocytic cells results in hypothyroidism [68–70]. Low circulating levels thyroxine are related to decreased alkaline phosphatase [71] and augmented low density lipoprotein (LDL) [69]. In phagocytic cells, thyroxine increases phagocytosis, bacterial killing and TNF-α and IL-6 production [72]. Furthermore, TNF-α and IL-6 impact redistribution of iron from plasma into the liver and mononuclear phagocyte system, resulting in low concentration of plasma iron (hypoferremia) [73]. During infection, hypoferremia limits iron availability to pathogenic microorganisms and reduces the potential pro-oxidant properties of iron, which may exacerbate tissue damage [73, 74]. These changes were reflected by increased levels of the iron binding and transport proteins, ferritin and transferrin, which are indicators of liver disease, inflammation and infection [74].
While IL-10 signalling is critical in controlling microbiota homeostasis and gut and liver immunopathology during whipworm infections, our data indicated that IL-22 is dispensable in the responses to T. muris. Interestingly, in our facility IL-22Rα-deficient mice infected with C. rodentium presented similar dysbiosis and sepsis-like pathology (caused by E. faecalis) to the one observed in whipworm-infected IL-10 signalling-deficient mice [32]. This may indicate that the intestinal inflammation elicited by C. rodentium infection of the epithelium is enough to trigger dysbiosis upon genetic predisposition by the lack of the IL22rα, while the colonization of the intestinal epithelium of these mice by whipworms is not sufficient to trigger the inflammatory responses that cause breakage of the microbiota homeostasis. In addition, the damage of the epithelium upon whipworm infection is restricted to specific areas where the worm is invading unlike C. rodentium infection which tends to occur more extensively across the epithelium. Moreover, the effect of IL-22 on anti-microbial production may be more relevant in responses to prokaryotic infections, such as those by C. rodentium.
Our results on the role of IL-22 signalling during T. muris infection are contrary to a previous report describing a delay in worm expulsion in IL-22 mutant mice due to reduced goblet cell hyperplasia [22]. We hypothesize this difference is due to differences in the kinetics of infection and the microbiota between mouse facilities that clearly affect the epithelial and immune intestinal responses responsible for the expulsion of the worms. Moreover, the microbiota composition of IL-22 mutant mice of each facility is directly influenced by the lack of IL-22 through its effects on antimicrobial production and mucus barrier function and this in turn affects the development of the intestinal immune system [75]. Although a role of IL-22 in inducing goblet cell hyperplasia and promoting microbiota homeostasis during whipworm infections cannot be excluded [76, 77], the induction of this mechanism of worm expulsion in the T. muris model is strongly dependent on the actions of IL-13 [3, 7] and regulated by IL-10 [38]. Similar observations have been made for other helminth infections in rodents including, Nippostrongylus brasiliensis [78] and Hymenolepis diminuta infections [79]. Taken together these observations suggest that in helminth infections IL-22 signalling plays a relatively minor role in worm expulsion.
Recent work has suggested that IL-28 plays a protective role in both dextran sulphate sodium and oxazalone-induced colitis in mice [80]. Our data, however, indicates that this cytokine is dispensable in responses to whipworm and consolidates the view that regulation of damage to intestinal tissue is context dependent reflecting extent of epithelial disruption. For whipworm, the data suggests that the focal damage generated by infection only becomes a significant problem in the absence of IL-10 signalling and/or following very heavy infections. Indeed, opportunistic bacteria-driven disease can occur upon heavy T. suis infection of weaning pigs. The resulting necrotic proliferative colitis involves crypt destruction, with inflammatory cells in the lamina propria and loss of goblet cells, and was reduced by antibiotic treatment, implicating enteric bacteria in the disease etiology [81]. Similar to our findings, accumulation of bacteria invading the mucosa was observed at the site of worm attachment and opportunistic members of the Enterobacteriaceae family that included Campylobacter jejuni and E. coli were isolated from these pigs and potentially contributed to the development of severe intestinal pathology [81]. Moreover, heavy T. trichiura infections in children cause Trichuris dysentery syndrome that is accompanied by a chronic inflammatory response, evidenced by high circulating levels of TNF-α [82, 83], which can potentially be driven by the overgrowth of opportunistic pathogens of the microbiota. Dysfunction of IL-10 signalling may trigger the development of dysbiosis and pathology during whipworm infection of weaning pigs and children as polymorphisms in the IL-10 gene in humans have been associated with T. trichiura infection [84]. Here, the IL-10 signalling deficient mice serve as a model to understand how polymorphisms in either the cytokine or the receptor impact the responses to whipworm infections.
In summary, our data provide critical insights into how IL-10 signalling, but not IL-22 or IL-28, orchestrates protective immune responses that result in whipworm expulsion while maintaining intestinal microbial homeostasis and barrier integrity. These findings contribute to the understanding on how IL-10 signalling controls colitis during trichuriasis and on the actions of Trichuris ova-based therapies for diseases such as IBD. Further studies will shed light into specific immune populations driving this process through IL-10 production and exerting effector functions in response to its signalling.
|
10.1371/journal.ppat.1006539 | Schistosome egg antigens, including the glycoprotein IPSE/alpha-1, trigger the development of regulatory B cells | Infection with the helminth Schistosoma (S.) mansoni drives the development of interleukin (IL)-10-producing regulatory B (Breg) cells in mice and man, which have the capacity to reduce experimental allergic airway inflammation and are thus of high therapeutic interest. However, both the involved antigen and cellular mechanisms that drive Breg cell development remain to be elucidated. Therefore, we investigated whether S. mansoni soluble egg antigens (SEA) directly interact with B cells to enhance their regulatory potential, or act indirectly on B cells via SEA-modulated macrophage subsets. Intraperitoneal injections of S. mansoni eggs or SEA significantly upregulated IL-10 and CD86 expression by marginal zone B cells. Both B cells as well as macrophages of the splenic marginal zone efficiently bound SEA in vivo, but macrophages were dispensable for Breg cell induction as shown by macrophage depletion with clodronate liposomes. SEA was internalized into acidic cell compartments of B cells and induced a 3-fold increase of IL-10, which was dependent on endosomal acidification and was further enhanced by CD40 ligation. IPSE/alpha-1, one of the major antigens in SEA, was also capable of inducing IL-10 in naïve B cells, which was reproduced by tobacco plant-derived recombinant IPSE. Other major schistosomal antigens, omega-1 and kappa-5, had no effect. SEA depleted of IPSE/alpha-1 was still able to induce Breg cells indicating that SEA contains more Breg cell-inducing components. Importantly, SEA- and IPSE-induced Breg cells triggered regulatory T cell development in vitro. SEA and recombinant IPSE/alpha-1 also induced IL-10 production in human CD1d+ B cells. In conclusion, the mechanism of S. mansoni-induced Breg cell development involves a direct targeting of B cells by SEA components such as the secretory glycoprotein IPSE/alpha-1.
| Infection with helminth parasites is known to be inversely associated with hyper-inflammatory disorders. While Schistosoma (S.) mansoni has been described to exert its down-modulatory effects on inflammation by inducing a network of regulatory immune cells such as regulatory B (Breg), the mechanisms of Breg cell induction remain unclear. Here, we use in vivo and in vitro approaches to show that antigens from S. mansoni eggs, among which the major glycoprotein IPSE/alpha-1, directly interact with splenic marginal zone B cells of mice which triggers them to produce the anti-inflammatory cytokine IL-10 and their capacity to induce regulatory T (Treg) cells. We also found that IPSE/alpha-1 induces IL-10 in human CD1d+ B cells, and that both natural and recombinant IPSE/alpha-1 are equally effective in driving murine and human Breg cells. Our study thus provides insight into the mechanisms of Breg cell induction by schistosomes, and an important step towards the development of helminth-based treatment strategies against hyper-inflammatory diseases.
| Helminths can persist for up to decades in the human host. This is hypothesized to be, at least in part, because of their evolutionarily adapted relationship with the host [1]. Helminths are well-known for their strong capacity to promote the regulatory arm of the hosts immune system, thereby prolonging their survival within the host [2]. As a bystander effect, helminths can also suppress immune responses to other antigens, such as allergens and auto-antigens, and other pathogens. This bystander effect seems to be so pronounced that it may prevent the development of inflammatory diseases. Indeed, both epidemiological studies and mouse models show a clear protective role of helminths against various forms of auto-immunity, allergic airway inflammation, colitis etc. [3,4,5,6,7]. The formation of a network of regulatory immune cells plays a crucial role for the protective effect. Helminth infection, and in particular infection with schistosomes such as Schistosoma (S.) mansoni are well-known to induce regulatory B (Breg) cells [8–15], a relatively new member in the network of regulatory immune cells. Breg cells have gained considerable attention due to their ability to down-modulate inflammation in a variety of conditions ranging from autoimmune disorders such as experimental autoimmune encephalomyelitis (EAE), collagen-induced arthritis (CIA), lupus and chronic colitis to anaphylactic and allergic airway inflammation [8,10–12,16–23]. Regulatory B cells suppress pro-inflammatory immune responses via several mechanisms, of which the ones best described are the expression of the regulatory cytokine interleukin-10 (IL-10) and induction of regulatory T (Treg) cells [24].
We previously reported the induction of Breg cells by schistosome infection in both mouse and human, and found the most potent IL-10-producing Breg cells within the human CD1d+ B cell subset. This corresponds to the CD1d+CD23lowCD21+ marginal zone (MZ) B cell subset in mice, which efficiently reduced experimental allergic airway inflammation in our model [12]. The cellular mechanisms to achieve Breg cell induction as well as the nature of the B cell-activating S. mansoni antigens however remain largely unknown. The identification of relevant stimulatory molecules and optimal Breg cell-inducing conditions is a critical step in enhancing the activity of Breg cells for use as a new therapeutic tool against inflammatory diseases.
Both an indirect induction of a regulatory phenotype in B cells by activation of accessory cell types, as well as a direct binding and interaction between S. mansoni antigens and B cells via pattern recognition receptors (PRRs) such as Toll-like receptors (TLRs) expressed on B cells [25] are plausible options. In the splenic MZ, located at the border of white and red pulp, MZ B cells are nested between SIGN-R1+ MZ macrophages and Siglec-1+ metallophilic macrophages [26]. MZ macrophages not only fulfill a main function in sensing blood-borne pathogens, but also perform functional interactions with MZ B cells. These interactions have important implications for the maintenance of the MZ itself and the function of MZ B cells and macrophages [27–29]. Hence, it is therefore tempting to speculate that MZ macrophages are a prime candidate as Breg cell induction partner. On the other hand, the direct ligation of various TLRs on B cells, including TLR2, TLR4, TLR7 and TLR9, has been described to induce IL-10 production [30,31]. In addition, BCR and CD40 engagement were described to be involved in IL-10-dependent regulatory B cell function in models of EAE, CIA, and contact hypersensitivity [17–19,32].
In the current study, we tested the hypothesis that eggs and/or egg-derived excretory-secretory molecules from S. mansoni, without the context of natural infection, are sufficient to drive Breg cell development by activating splenic MZ B cells. In addition, we investigated whether Breg cells are induced indirectly by activation of accessory cell types in the MZ, or by direct binding and interaction via PRRs on B cells. We found that egg antigens drive Breg cell development in vivo and in vitro by direct interaction with splenic B cells, which after binding and internalization of egg antigens secrete elevated levels of IL-10 and are capable of driving Treg cell development. The egg antigen-induced Breg cell development was independent of macrophages of the marginal zone but was enhanced by CD40 ligation. Most importantly, we identified the egg glycoprotein IPSE/alpha-1 as a single molecule from S. mansoni that is capable to induce Breg cells both in mice and man. This knowledge will assist to further define helminth-specific conditions for the generation of Breg cells to be used in therapeutic approaches.
To elucidate the mechanism by which S. mansoni can drive the development of Breg cells, we first investigated whether schistosome eggs or their soluble antigens were sufficient to drive Breg cell development in vivo, without the context of natural infection. Intraperitoneal treatment of C57BL/6 mice with two doses of 5000 S. mansoni eggs or 100 μg of soluble egg antigens (SEA) one week apart efficiently induced IL-10 protein expression in splenic CD19+ B cells one week after the last injection. IL-10 protein secreted during 2 days ex vivo restimulation with SEA was 3 to 4-fold increased compared to the amount secreted by restimulated B cells from control-treated animals (Fig 1A), while IL-6 was unchanged. This indicates a typical cytokine expression pattern characteristic for Breg cells. The frequencies of B cells expressing intracellular IL-10 protein were likewise significantly 2-fold increased (Fig 1B and 1C). Also the surface activation marker CD86, often upregulated on activated B cells and Breg cells [33–35], was increased on splenic B cells by egg or SEA treatment, while CD40 expression was not significantly changed (Fig 1D). To verify that the observed effects are specific and exclude a general influence of protein solutions on B cell IL-10 production and activation, we treated mice with human serum albumin (HSA) as infection-unrelated control protein. We did not observe an increased IL-10 secretion (S1A Fig) or CD86 expression (S1B Fig) by B cells compared to the PBS group, and concluded that PBS is a suitable control for subsequent experiments. Furthermore, injection of eggs or egg antigens was as efficient in increasing the frequency of IL-10-expressing B cells as was chronic infection with S. mansoni (S1C Fig), and egg-injected mice continued to have an elevated B cell IL-10 production until at least 4 weeks after the last egg injection (S1D Fig), indicating that this phenotype is persisting over longer periods. SEA was purified from liver eggs and can contain LPS to variable extent. Since the TLR4 ligand LPS is known for its capacity to drive B cell IL-10 expression and Breg cell development [33,36,37], it is crucial to exclude that the Breg driving capacity by SEA in vivo was due to LPS contamination of the schistosome antigen preparations. Therefore, the same experiment was repeated in TLR4-deficient animals and compared to wild-type. Upon SEA treatment, IL-10 secretion of B cells as well as intracellular IL-10 and surface CD86 expression was overall comparable in both groups (S2 Fig), indicating that the Breg-inducing capacity by SEA was largely not attributable to a putative LPS contamination. To confirm the regulatory function, splenic B cells from the various groups were tested for their capacity to drive Treg cell development, an acquired phenotype previously described for splenic B cells during natural schistosome infections [12]. Indeed, splenic B cells from egg- or SEA-injected, but not control-treated, mice induced the development of CD25+Foxp3+ Treg cells during 4 day co-culture with CD25-depleted CD4 T cells (Fig 1E and 1F), which confirms the regulatory capacity of egg antigen-activated B cells ex vivo. As expected, IL-10 protein concentration in co-culture supernatants was only increased in presence of egg antigen-activated but not control B cells (Fig 1G). Collectively, these data show that schistosome eggs or their soluble antigens on their own are sufficient to induce IL-10-producing Breg cells in vivo, without the context of natural infection, and that these B cells are bona fide Breg cells that can drive Treg cell development.
Different B cell subsets have been described to give rise to Breg cells, especially in spleen where subsets differ e.g. in tissue localization and pathogen recognition receptor expression [25,38]. We and others had previously identified CD23lowCD21+ marginal zone B cells as the major IL-10-producing splenic B cell subset during chronic Schistosoma infection and mediating protection in a mouse model of airway inflammation [8,12]. To test whether soluble egg antigens act on the same splenic subset, we sorted splenic CD23lowCD21+ marginal zone B cells from egg-, SEA-treated, or control mice for subsequent ex vivo restimulation and cytokine analysis, and compared this with the major splenic B cell subset, CD23hiCD21- follicular B cells (Fig 2A). Only marginal zone B cells but not follicular B cells showed significantly increased IL-10 secretion as was measured in culture supernatants after 2 day restimulation with SEA (Fig 2B). As for total B cells (Fig 1A), also for the individual subsets IL-6 expression was not increased (Fig 2C). Intracellular IL-10 expression and CD86 expression was likewise significantly upregulated in marginal zone B cells of egg- or SEA-treated mice compared to control-treated mice. SEA seemed to be more potent than egg injection in activating follicular B cells, as SEA-injection also significantly increased intracellular IL-10 and CD86 expression in this subset, although expression levels remained significantly lower compared to marginal zone B cells (Fig 2D and 2E). For analysis of B cell activation we generally restimulated cells ex vivo with SEA. This increased the baseline expression of CD86 in all groups compared to medium (average MFI of follicular B cells: 1013–1632±62.2; for marginal zone B cells: 2640–4118±176.9, for all groups and without significant differences between groups), but was required for detection of B cell cytokines as a result of the in vivo antigen exposure. Without SEA restimulation, we found a trend of increased IL-10 production by B cells which only reached significance upon additional restimulation (S1E Fig), indicating that renewed exposure to antigen is required to achieve detectable B cell activity and cytokine production. This is also supported by experiments in IL-10 GFP reporter mice, in which IL-10 (GFP) accumulates in B cells during the entire in vivo treatment period. Here, increased IL-10 (GFP) expression, without ex vivo restimulation, was only detectable in B cells of 14 weeks chronically infected mice, but not in B cells of mice that received two injections of eggs within a relatively short period of 2 weeks (S1F and S1G Fig). Altogether, these data indicate that the result of in vivo development of IL-10 producing B cells is in principle detectable without restimulation (S1E and S1G Fig), but the data from egg-injected IL-10 reporter mice also suggest that an ex vivo SEA restimulation is required to visualize its full IL-10 potential, something which will happen in vivo during a natural infection due to the constant production of eggs and the high levels of circulating antigens. Taken together, these data support the notion that B cells, and in particular marginal zone B cells, are responsive to in vivo schistosome antigen stimulation thus supporting the findings in natural schistosome infections.
Molecules secreted by schistosome eggs are highly glycosylated and known to bind to C-type lectin receptors [39,40]. Since B cells show a very restricted expression of those receptors [41], we hypothesized that other C-type lectin receptor-expressing cell types in the splenic marginal zone, such as macrophages or dendritic cells, bind SEA and provide additional signals to the marginal zone B cells to support Breg cell development. Among the accessory cell types, macrophages of the splenic marginal zone were of particular interest because of their known interactions with marginal zone B cells [28,42] and schistosome antigens [43,44]. However, it was unknown whether marginal zone macrophages can capture SEA in vivo and are important for B cell IL-10 expression. To evaluate this, fluorescently labeled SEA was injected i.v. and 30 minutes to 24 hours later various splenic cell types were analyzed for bound SEA using fluorescence microscopy and flow cytometry. Already after 30 minutes of injection, SEA clustered along the marginal zone area of the spleen as detected by fluorescence microscopy of splenic tissue sections (Fig 3A). SEA localized predominantly within two specialized macrophage subsets of the marginal zone: SIGN-R1-expressing MZ macrophages and Siglec-1-expressing marginal metallophilic macrophages (Fig 3B). In contrast, after injection of labeled ovalbumin (OVA) as non-schistosomal control protein no fluorescence signal was detected in the spleen (S3 Fig). Flow cytometry confirmed that 83% of metallophilic macrophages were positive for SEA only 30 minutes after injection, and still 53% of cells after 24 hours (Fig 3C). In addition, F4/80-expressing red pulp macrophages and Ly6Chi monocytes efficiently bound SEA (81.7 and 82.6%, respectively), while only a small fraction of dendritic cells and neutrophils did bind SEA (Fig 3C, S4A–S4C Fig for gating scheme of cell types). Metallophilic macrophages not only bound SEA abundantly, they also significantly upregulated typical surface activation markers such as CD11c and CD86, but not MHCII, at 30 minutes and 2 hours after SEA injection, respectively (Fig 3D). Because macrophages of the marginal zone were most potent in binding SEA, we next addressed their role for SEA-mediated marginal zone Breg cell induction. To this end, macrophages were depleted in vivo by i.p. injection of clodronate-containing liposomes [45] prior to injection of schistosome eggs. Eggs were injected 3 and 4 weeks after clodronate treatment, at time-points when only macrophages, including metallophilic and MZ subsets, were significantly reduced in spleens, but no other cell types (S4D Fig and [46]). Successful and specific depletion of splenic macrophages was also confirmed by fluorescence microscopy of tissue sections (Fig 3E) and flow cytometry (S4E Fig) at 7 days after the last egg injection, when B cell activity was analyzed. Unexpectedly, IL-10 secretion of splenic B cells from macrophage-depleted mice was equal to that from control liposome-treated mice (Fig 3F). Also the upregulation of intracellular IL-10 (Fig 3G) and CD86 expression in B cells, as well as the induction of Foxp3+CD25+ and IL-10+CD25+ CD4 T cells (S4F and S4G Fig) following SEA injection was not affected by the absence of macrophages. In conclusion, splenic macrophages are not essential for schistosome antigen-induced Breg cell development, despite the high binding of SEA by different macrophage subsets.
To test whether B cells directly bind and interact with schistosome antigens without the help of surrounding accessory cell types, fluorescently labeled SEA was injected i.v. and its co-localization with B cells analyzed by fluorescence microscopy of splenic tissue sections. Indeed, egg antigens were found to co-localize with some splenic B220+ B cells (Fig 4A). Flow cytometry, a more sensitive method compared to fluorescence microscopy, showed that only MZ B cells but not follicular B cells did bind SEA in vivo, with a maximum of 56.4% of cells being positive at 2 hours after SEA injection. SEA was still detectable on 11.5% of marginal zone B cells at 24 hours (Fig 4B). Marginal zone B cells also showed an increased CD86 surface expression following SEA injection, which was significant at 6 hours after SEA injection (Fig 4C). Interestingly, marginal zone B cells that bound SEA showed a higher CD86 expression compared to cells that were found to be negative for SEA, i.e. approximately 3-fold (SEA positive) versus only 1.5-fold (SEA negative) compared to B cells from untreated animals (Fig 4C). This suggests that marginal zone B cells not only efficiently bind egg antigens in vivo, but also become (more) activated as a consequence of this interaction. Binding of SEA to B cells was confirmed in vitro by culturing splenic B cells with fluorescently labeled SEA for 60 minutes, after which up to 16% were positive for SEA as measured by flow cytometry (Fig 4D), which is a similar percentage as found for total B cells after in vivo application of labeled SEA (Fig 4B). By using SEA labeled with the pH-sensitive dye pHrodo, it was shown that egg antigens were not only bound to the surface but were also internalized by B cells into acidic cellular compartments (Fig 4D). As in vivo, also in vitro the marginal zone B cell subset bound SEA more efficiently than the follicular B cell subset, with in average 15.9% versus 6.9% of cells being positive for SEA (Fig 4E). Collectively, these data show that B cells, and in particular MZ B cells, are capable of directly interacting with schistosome antigens by binding and internalization of those antigens, both in vivo and in vitro.
Next, we investigated whether the observed direct interaction of B cells with SEA can drive IL-10 expression and induction of regulatory B cell function in vitro. To this end, CD19+ splenic B cells from naïve mice isolated using MicroBeads were cultured for 3 days with SEA. SEA-stimulated B cells secreted significantly more IL-10, but not IL-6, compared to non-stimulated B cells (Fig 5A), showing a typical cytokine pattern characteristic for schistosome-induced Breg cells. Separate cultures of sorted MZ and follicular B cells showed once more that the MZ B cell subset reacted more potently to SEA, e.g. with a significantly higher IL-10 secretion compared to the follicular subset (Fig 5B). In addition, frequencies of B cells expressing intracellular IL-10 were likewise increased (Fig 5C). To ensure that the IL-10 phenotype is not reliant on artificially high levels of stimulation with PMA/ionomycin which is added to facilitate detection of intracellular IL-10, we repeated the assay using cells from IL-10-GFP reporter (TIGER) mice with similar results (S5 Fig). CD40 and CD86 expression were upregulated compared to cultures in medium alone (Fig 5D). Importantly, in vitro SEA-activated B cells were also capable of driving Treg cell development during a 4 day co-culture with CD25-depleted CD4 T cells (Fig 5E), thus providing further evidence for a bona fide regulatory function of the in vitro induced Breg cells. Because we had seen internalization of egg antigens into acidic compartments (Fig 4D), we wondered whether lysosomal processing is necessary for induction of IL-10 expression. Addition of chloroquine, an inhibitor of endosomal acidification [47], significantly reduced the IL-10 secretion and frequency of IL-10+ B cells induced by SEA and CpG (ligand for endosomal TLR9), but not by Pam3Cys (ligand for surface TLR2) (Fig 5F and 5G). This suggests that internalization and endosomal processing of SEA is required for B cell IL-10 induction. The type of receptor involved in direct activation of Breg cells by SEA remains unknown. Egg antigens are abundantly glycosylated and known to bind to C-type lectin receptors [39,40]. Lex-motifs, one of the most abundant glycan structures present in SEA, bind to the C-type lectin receptor SIGN-R1. However, when treating SIGN-R1-deficient mice with SEA, IL-10 expression was equally well induced compared to wild-type mice (S6A Fig), suggesting no involvement of the Lex-motifs. Furthermore, stimulation of various TLRs on B cells, including TLR2, TLR4, TLR7 and TLR9, has been described to induce IL-10 production [30,31]. Because SEA has been reported to contain TLR2 activity [48,49], we compared SEA-induced Breg cell responses in wild-type and TLR2-deficient B cells. We did however not observe any difference in IL-10 secretion between the two strains (S6B Fig), excluding a role of TLR2-triggering SEA components in SEA-induced B cell IL-10 production. This is further supported by the fact that TLR2-mediated B cell activation was independent of endosomal processing while SEA-mediated activation was dependent on it (Fig 5F and 5G). Collectively, these data demonstrate that Breg cells can be generated in vitro by culture with schistosome antigens, that endosomal processing is involved in this process, and that these Breg cells are functional in the sense that they support Treg cell development.
Previous studies highlighted a role for CD40 ligation during in vitro Breg cell induction [50,51]. We therefore tested whether CD40 ligation could increase the SEA-mediated effect on Breg cell development. Addition of anti-CD40 stimulatory antibody (Ab) to the 3 day SEA culture increased IL-10 secretion of splenic B cells by 1.7-fold compared to SEA alone. A similar enhancing effect was observed for IL-6 secretion (Fig 6A) and CD86 expression (Fig 6B). In contrast to anti-CD40 Ab, addition of anti-IgM Ab did not significantly enhance the SEA-mediated B cell activation (S7 Fig). To exclude stimulatory effects from LPS in vitro, B cells from TLR4-deficient mice were stimulated with SEA with or without addition of anti-CD40 Ab. The fold increase of IL-10 secretion compared to B cells cultured in medium was similar for wild-type and TLR4-deficient cells, thus excluding a major effect of the TLR4 ligand LPS (S6C Fig). Finally, co-culture of CD25-depleted CD4 T cells with anti-CD40-stimulated B cells increased the frequency of CD25+Foxp3+ Treg cells, which was further increased if B cells had been stimulated with SEA plus anti-CD40 Ab (Fig 6C). Thus, SEA stimulation plus CD40 ligation of B cells further enhanced the capacity to drive the development of IL-10-producing Breg cells in vitro.
Previous reports only addressed the role of CD40 ligation in vitro, but the relevance for in vivo Breg cell induction was not investigated. We therefore blocked CD40 ligand in vivo by i.p. injection of a hamster anti-mouse CD40 ligand blocking mAb (200 μg; every 4 days starting at day -1 prior to 1st SEA injection) during SEA treatment of mice and analyzed the effect on Breg cell activation. In hamster IgG-injected control mice, SEA treatment increased the amount of B cell-derived IL-10 secretion by 3.3-fold compared to PBS treatment. Upon anti-CD40 ligand administration this increase was only 1.6-fold and thereby significantly lower (Fig 6D). Importantly, the SEA-mediated upregulation of intracellular IL-10 and CD86 expression by B cells was even fully abolished when CD40 ligand was blocked (Fig 6E and 6F). Taken together, CD40 ligation enhances the Breg cell-inducing effect of SEA both in vitro and in vivo.
SEA is a complex mixture of several different antigens. In the next step, we therefore aimed to identify specific antigens in SEA that are relevant for Breg cell induction. We focused on three major antigens that provoke an antibody response in nearly all infected patients: omega-1, kappa-5 and IPSE/alpha-1 [52,53]. B cells were able to bind fluorescently labelled natural IPSE/alpha-1 (nIPSE), a secreted egg antigen we purified from egg extracts, in a dose-dependent manner during 60 minutes in vitro culture (Fig 7A). During 3 days culture however, nIPSE induced significantly elevated IL-10 but not IL-6 secretion by B cells in a concentration dependent manner (Fig 7B and S8A Fig). Importantly, recombinant IPSE/alpha-1 expressed in tobacco plants (pIPSE), which behaves as nIPSE in terms of protein dimerization and human basophil activation (S9 Fig), had similar effects to the natural molecule on B cell IL-10 and IL-6 secretion (Fig 7C). Both nIPSE and pIPSE-stimulated B cells were capable of driving Treg cell development during B cell-T cell co-culture (Fig 7D). Interestingly, SEA depleted of IPSE/alpha-1 (SEAΔIPSE) was as efficient as total SEA in inducing IL-10 secretion and CD86 expression by B cells, which suggests that also SEA antigens other than IPSE/alpha-1 can activate B cells (Fig 7B). However, as opposed to IPSE/alpha-1, other major components in SEA, such as omega-1 and kappa-5, did not increase IL-10 in any of the tested concentrations (1–20 μg/ml) (S8B Fig), thus excluding a role for these antigens in SEA-mediated Breg cell induction. Notably, omega-1 is toxic to B cells at concentrations of 5 μg/ml and above, and was therefore only tested at 1 μg/ml. For better comparability, we determined the following average relative amounts of IPSE/alpha-1, omega-1 and kappa-5 within SEA, based on the yields of several purifications of these molecules from SEA: 1.2% IPSE/alpha-1, 0.6% omega-1 and 1.8% kappa-5. Furthermore, to proof that the Breg cell-inducing effect is specific for molecules in SEA, we stimulated B cells in vitro with adult worm antigen (AWA) as control of an S. mansoni-derived antigen mixture not containing IPSE. AWA was unable to induce IL-10 secretion when tested in the same concentration as used for SEA (S8C Fig). For a possible future therapeutic application of antigen-activated Breg cells against e.g. allergic diseases, it is crucial to confirm the IL-10-inducing effect in human B cells. After 3 days in vitro stimulation with SEA and anti-CD40, we found a significant increase in the percentage of total human IL-10+ CD19+ B cells compared to cells cultured with anti-CD40 alone. Comparing different B cell subsets, which have previously been attributed with regulatory properties [54,55], we found the increase in IL-10+ B cells after SEA stimulation to be most pronounced among CD1d+ B cells rather than CD24+CD27+ and CD24+CD38+ B cells. Both nIPSE and pIPSE significantly increased the fraction of IL-10-expressing cells among CD1d+ B cells, whereas neither had an effect on the other two subsets investigated. As CD1d+ B cells only comprise a very small fraction of all B cells (CD24+CD27+ B cells >> CD24+CD38+ B cells > CD1d+ B cells), the effect of nIPSE and pIPSE does not translate into an increase in the percentage of IL-10+ cell in the total B cell pool in contrast to SEA (Fig 8). This also suggests that additional molecules in SEA may have an IL-10-inducing effect. Collectively, we demonstrated that SEA was bound to and internalized by B cells, and that this direct interaction drives the development of Breg cells. Furthermore, we identified IPSE/alpha-1 as a single molecule of SEA that induces Breg cells in mice and humans, both as a natural and a recombinant molecule.
In this study, we sought to identify the molecules and mechanisms involved in the induction of IL-10-producing B cells by the helminth S. mansoni. We found that soluble antigens derived from schistosome eggs, amongst which the secretory antigen IPSE/alpha-1, directly interacted with B cells. This led to the development of Breg cells characterized by IL-10 secretion and Treg cell-inducing capacity. Next to the potentially therapeutic relevance of how to generate regulatory, anti-inflammatory cells, this study also provides mechanistic insight into how schistosomes interact with the host immune system, expanding the regulatory arm of immunity and thereby prolonging its survival in the host.
While we and others have shown IL-10 expression in splenic Breg cells during natural infection with S. mansoni [8,11,12], the contribution of S. mansoni-derived egg antigens was not yet studied. SEA is highly immunostimulatory and well-known to promote Th2 as well as Treg cell responses in the host [56]. Here, we found that S. mansoni egg antigens are also able to induce IL-10-producing B cells in vivo, without the context of natural infection. Because SEA is a complex mixture of several different antigens, it is not unexpected that different types of immune cells and different qualities of immune responses are induced. The use of Breg cells as a therapeutic tool against inflammatory diseases is especially attractive because Breg cells in turn can induce Treg cell development [57], which would thus amplify the beneficial regulatory effect. It was therefore important to investigate whether SEA-induced IL-10-producing B cells have the capacity to trigger Treg cell development. This was indeed the case as we could show by in vitro co-cultures of T cells with egg antigen- or IPSE-activated B cells. Similarly, Breg cells isolated from naturally schistosome-infected mice and humans were previously shown to drive Treg cell development in vitro [8,12,58], thus indicating a common feature of schistosome-induced Breg cells.
Our finding that egg antigens could induce splenic Breg cell development in vivo raised the question whether those antigens can directly interact with B cells in the spleen. In in vivo binding studies using fluorescently labeled SEA and fluorescence microscopy, egg antigens were indeed found to directly bind to splenic B cells. Although various egg antigens are abundantly glycosylated, the restricted and low expression of C-type lectin receptors by B cells [41] argues rather for the involvement of non-C-type lectin PRRs expressed by B cells in the direct binding of SEA components. Indeed, preliminary experiments with B cells of SIGN-R1-deficient mice, showed an equal IL-10 expression in response to SEA compared to wild-type littermates, suggesting no involvement of the Lex-motifs, one of the most abundant glycan structures present in SEA and known to bind SIGN-R1. Instead, we found that within a mixture of schistosome antigens, at least the egg glycoprotein IPSE/alpha-1 was capable of driving Breg cell development in vitro by directly interacting with B cells, equipping them with Treg cell-inducing capacity. Two independent experimental approaches suggested that egg antigens are taken up and processed in acidic lysosomes. Previous reports on in vitro induction of Breg cells by helminth antigens did not use highly sort-purified B cells as in our study, but total splenocyte preparations [8,10] or merely B cell-enriched cultures [13]. Hence, it was impossible to exclude indirect, accessory cell type-mediated B cell stimulation or IL-10 production by other cell types [59,60]. Other reports addressed B cell activation by IL-10 production, but did not study the regulatory activity of schistosome antigen-exposed B cells compared to unstimulated B cells [14]. It must be emphasized that the sole demonstration of upregulated IL-10 expression is not sufficient to characterize B cells as Breg cells, as IL-10 can fulfill other roles in B cell biology independent from a regulatory function. We thus present the first report on direct induction of functional Breg cells with in vitro regulatory activity by helminth antigens.
With respect to the development of therapeutic applications, it would be interesting to see whether SEA-induced Breg cells are more potent than Breg cells induced by other compounds, like TLR7 or TLR9 ligands. Opposite to SEA, stimulants like R848 and CpG also induce substantial amounts of B cell proliferation and pro-inflammatory cytokines like IL-6 in addition to high levels of IL-10. It is currently unknown which side-effects would result in a therapeutic application, but it is tempting to speculate that compounds that selectively induce IL-10 are more preferable. We tried to compare IL-10-producing B cells induced in vitro by different agents, including SEA, for their capacity to inhibit allergic airway inflammation in vivo. We were however not able to confirm a suppressive capacity for any of the conditions despite the usage of a published model [31]. In the past, we have successfully applied adoptive transfers of in vivo, schistosome-induced Breg cells (generated during a natural infection) in allergic airway inflammation models [12]. Therefore, we assume that underlying differences between in vivo and in vitro stimulation of Breg cells may be crucial for the activity in a disease model. This may be related to issues like a differential homing or to the strength and kinetics of activation and cytokine production which determine the suppressive capacity of Breg cells on bystander immune activation in the host. Knowing that schistosome antigens can directly induce Breg cell development, we next addressed signals that regulate or enhance antigen-induced B cell IL-10 expression. In previous studies, CD40 engagement was described to induce B cell IL-10 expression [18,19,50,51,59]. This is in line with our results showing that SEA-induced IL-10 expression was significantly increased by addition of agonistic CD40 Ab. This points to the potential of a combined therapy for inflammatory diseases using helminth antigens together with anti-CD40 Ab treatment. Indeed, a report by the group of Mauri et al. provided evidence that experimental therapy with an agonistic Ab against CD40 can ameliorate autoimmune disease [61], as did cellular therapy with Breg cells [50], although a combined treatment was not yet tested. The groups of Fillatreau and Mauri proposed a two-step model for the acquisition of regulatory properties by B cells, with exposure to innate stimuli—such as TLR ligands—as one step and CD40 or BCR engagement as second step to establish Breg cell function [36,50]. We found a similar dependency for in vivo Breg induction by helminth antigen, for which CD40 ligation was crucial. Our data also show that, although Breg cell induction in vitro can be achieved alone without additional stimuli, engagement of CD40 further enhances this effect. Several cell types including T cells, B cells, DCs, basophils, NK cells, mast cells and macrophages express CD40 ligand (CD40L, CD154) [62] and could in principle serve as interaction partner ligating CD40 on B cells. It is however tempting to speculate that neutrophils play a role as they have been reported to express CD40L and activate MZ B cells for immunoglobulin production in a contact-dependent manner [63].
As we found MZ B cells to be the main IL-10-producing B cell subset, it was tempting to speculate that accessory cell types of the splenic marginal zone interact with schistosome antigens, and subsequently drive the development of MZ Breg cells. Macrophages of the splenic MZ were of particular interest because of their known interactions with MZ B cells during steady state [28,42] and their expression of SIGN-R1. This C-type lectin receptor was found to bind schistosome antigens in vitro by using SIGN-R1-overexpressing fibroblasts [43,44]. However, it was unknown whether MZ macrophages can capture SEA in vivo and are important for B cell IL-10 expression. As hypothesized, we found macrophages of the MZ to efficiently bind SEA upon in vivo administration. However, Breg cell induction was not affected upon in vivo depletion of macrophages, thus excluding a major role of macrophages in this process. Indirectly, also Mangan et al. [10] showed that macrophages were dispensable for Breg cell induction during schistosomiasis, as macrophage depletion did not affect the B cell-mediated control of anaphylaxis. The immunological role of SEA-binding MZ macrophage subsets and the identity of the binding receptor remain to be determined.
A limited number of reports is available that used specific helminth antigens to induce B cell IL-10 expression, namely the filarial antigen ES-62 [64] and the oligosaccharide lacto-N-fucopentaose III (LNFP III) that contains the LeX trisaccharide antigen present on various schistosome glycoproteins [13]. Both antigens were either applied in vivo or used in vitro for stimulation of B cell-enriched cultures that still contained other cells, which means it remains unclear whether those antigens can directly bind to and interact with B cells, without indirect support from other cell types. Our study therefore identified IPSE/alpha-1 as the first helminth molecule with direct Breg cell-inducing capacity in mice and humans. Importantly, this capacity was resembled by recombinant IPSE, which is an important prerequisite for a possible therapeutic use. The use of helminth molecules for therapeutic purposes has gained renewed interest as controlled human infections with helminths showed disappointing effects in recent phase II and phase III trials (reviewed in [65]). More studies are required to define the optimal antigen, dose, time point and length of treatment as well as the suitability to treat specific inflammatory diseases [66,67]. Therefore, the identification of helminth-specific Breg-inducing antigens is warranted, even more so as the availability of active recombinant forms will ultimately allow its production under GMP conditions.
IPSE/alpha-1 was originally described as basophil IL-4-inducing principle of Schistosoma eggs [53] and was shown to induce a mixed Th1/Th2 type of immune response in spleen upon in vivo administration [68]. In our assays, IPSE/alpha-1 directly interacted with murine B cells via a still unknown receptor, which led to activation, induction of B cell IL-10 secretion and Treg cell induction in vitro. Although IPSE/alpha-1 is a highly glycosylated protein, we consider a role of IPSE-related glycans as unlikely because both, pIPSE and nIPSE were capable to induce B cell IL-10 expression despite differences in glycosylation (native IPSE contains Lex motifs [69], while plants per definition cannot make Lex motifs [70]. In addition, natural omega-1 and IPSE share a similar glycosylation (Lex related) but have opposing activities both in B cells (here) and on DCs [71]. IPSE/alpha-1 has been shown to not only bind to IgE but also to IgG, both to Fc and Fab fragments [53]. We therefore hypothesize that IPSE could bind to B cells via the B cell receptor or surface-exposed IgG. Particularly important for a possible therapeutic use is our finding that both natural and plant-derived IPSE/alpha-1 induced IL-10 expression also in human CD1d+ Breg cells. This is even more intriguing as the CD1d+ B cell subset has been previously described to be increased in number and activity both in experimental infections in mice and in people living in endemic areas [12,58]. We propose a mechanism of schistosome-induced Breg cell induction in which B cells directly interact with schistosome egg antigens by binding and internalizing antigen, lysosomal processing and subsequent up-regulation of CD86 and IL-10 expression. The MZ B cell subset appeared to be particularly responsive, and CD40 engagement further enhanced Breg cell activity. Furthermore, we have successfully identified the secreted egg antigen IPSE/alpha-1 as one of the Breg-inducing antigens. These egg antigen-induced Breg cells were potent in driving Treg cell development, allowing for induction of two potent regulatory responses by the same antigen. To our knowledge, our study provides the first description of a helminth-specific molecule that interacts with and induces Breg cells, and a mechanistic insight into how schistosomes interact with their host, influence its regulatory immunity and thereby promoting their prolonged survival in the host.
Female C57BL/6OlaHsd mice from Harlan, TLR4-deficient mice (on C57BL/6 genetic background, kindly provided by Dr. S. Akira, Osaka, Japan), TLR2-deficient mice (on C57BL/6 genetic background, kindly provided by the group of Dr. K. Willems van Dijk), SIGN-R1-deficient mice (on C57BL/6 genetic background, kindly provided by the group of Dr. W. Unger), DEREG (DEpletion of REGulatory T cells) mice (on C57BL/6 genetic background, kindly provided by Dr. T. Sparwasser) and IL-10-GFP reporter (TIGER) mice (on C57BL/6 genetic background, kindly provided by Dr. R.A. Flavell) were housed under SPF conditions in the animal facilities of the Leiden University Medical Center in Leiden, The Netherlands, and used for experiments at 8–14 weeks of age. Percutaneous infection of mice with S. mansoni was performed as described elsewhere [12], and mice sacrificed in the chronic phase (14–15 weeks) post infection.
Freshly isolated S. mansoni eggs from trypsinized livers of hamsters infected for 50 days were washed in RPMI medium with 300 U/ml penicillin, 300 μg/ml streptomycin (both Sigma-Aldrich, Zwijndrecht, The Netherlands) and 500 μg/ml amphotericin B (Thermo Fisher Scientific, Breda, The Netherlands) and then kept at -80°C. SEA, AWA, omega-1, kappa-5 and IPSE/alpha-1 were prepared and isolated as described previously [53,71,72,73]. The purity of the antigen preparations was checked by SDS-PAGE and silver staining, and protein concentrations determined using the BCA procedure. The antigen preparations had an endotoxin content of less than 150 ng/mg protein (SEA) or 3 ng/mg protein (purified molecules) as tested by Limulus Amoebocyte Lysate (LAL) test and TLR4-transfected HEK-reporter cell lines (kindly provided by Prof. Golenbock, University of Massachusetts Medical School, Boston, USA).
Recombinant IPSE was produced by transient expression in Nicotiana benthamiana and purified according to the methods described in [70]. In short, the complete sequence encoding the 134 AA mature Schistosoma mansoni IPSE (Smp_112110) was codon optimized and preceded by a signal peptide from the Arabidopsis thaliana chitinase gene (cSP) and a N-terminal 6x histidine-FLAG tag (H6F) was included. The full sequence was synthetically constructed at GeneArt and cloned into a pHYG expression vector. In all experiments the silencing suppressor p19 from tomato bushy stunt virus in pBIN61 was co-infiltrated to enhance expression. For gene expression the two youngest fully expanded leaves of 5–6 weeks old N. benthamiana plants were infiltrated by injecting Agrobacterium tumefaciens containing the IPSE expression plasmid. N. benthamiana plants were maintained in a controlled greenhouse compartment (UNIFARM, Wageningen) and infiltrated leaves were harvested at 5–6 days post infiltration. Plant-produced recombinant IPSE was obtained by applying leaf apoplast fluid containing IPSE to Ni-NTA Sepharose (IBA Life Sciences) in 50 mM phosphate buffered saline (pH 8) containing 100 mM NaCl. Bound IPSE was eluted with phosphate buffered saline (pH 8) containing 0.5M imidazole. Total soluble apoplast proteins and purified IPSE were separated under reducing/non-reducing conditions by SDS-PAGE on a 12% Bis-Tris gel (Invitrogen) and subsequently stained with Coomassie brilliant blue staining.
Single cell suspensions of murine spleens were prepared by dispersion through a 70 μm cell strainer (BD Biosciences, Breda, The Netherlands), and erythrocytes depleted by lysis. For analysis of splenic myeloid cell populations, spleens were digested for 1 hour at 37°C by incubation with collagenase D (2 mg/ml; Roche, Woerden, The Netherlands) and DNase I (2000 U/ml; Sigma-Aldrich) before dispersion. B cells were purified from spleens by using anti-CD19 MicroBeads (Miltenyi Biotec, Leiden, The Netherlands) following the manufacturer’s protocol. Purity was routinely ~95–98%. After a typical CD19 MACS sort, circa 83% of all contaminating cells were CD3+ T cells, 7% CD11b+CD11c+ cells, 2% CD11b+ CD11c- cells, and 8% other cells. To determine cytokine secretion of splenic B cell subsets, CD19+ B cells were subsequently sorted by flow cytometry for follicular B cells (CD23+CD21low) and marginal zone B cells (CD23-CD21hi) which resulted in purities of > 98%. CD4+ T cells were purified from spleens by negative selection and depleted of CD25-expressing cells using anti-CD25 MicroBeads (Miltenyi Biotec).
Peripheral blood mononuclear cells (PBMCs) were isolated from heparinized blood of healthy volunteers by Ficoll gradient centrifugation, and B cells were purified from PBMCs by using anti-CD19 MicroBeads (Miltenyi Biotec, Leiden, The Netherlands) following the manufacturer’s protocol.
Mouse splenic CD19+ B cells (1.5x106/ml) were cultured in medium (RPMI 1640 glutamax; Thermo Fisher Scientific), containing 5% heat-inactivated Fetal Bovine Serum (FBS; Greiner Bio-One, Alphen aan den Rijn, The Netherlands), 5 × 10−5 M 2-Mercaptoethanol (Sigma-Aldrich) and antibiotics (100 U/mL penicillin and 100 μg/mL streptomycin; Sigma-Aldrich). Human B cells (1.5 x 106/ml) were cultured in medium (RPMI 1640; Thermo Fisher Scientific), containing 10% heat-inactivated Fetal Bovine Serum, pyruvate (1 mM), glutamate (2 mM) and antibiotics (100 U/mL penicillin and 100 μg/mL streptomycin; all Sigma-Aldrich). The following stimuli were added as indicated in the figure legends: SEA (20 μg/ml), SEA depleted of IPSE (SEAΔIPSE, 20 μg/ml), natural (1, 5, 10, 20 μg/ml) or plant-derived IPSE (10 μg/ml), omega-1 (1 μg/ml), kappa-5 (1, 5, 10, 20 μg/ml), AWA (5, 10, 20 μg/ml). For some conditions, co-stimulatory rat anti-mouse CD40 antibody (2 μg/ml; clone 1C10; BioLegend, Uithoorn, The Netherlands) or goat anti-mouse IgM (0.5, 1, 2 μg/ml; Jackson ImmunoResearch, Suffolk, UK) was added to the culture. After 3 days culture at 37°C, supernatants were collected for cytokine analysis by ELISA. Cells were restimulated with PMA (100 ng/ml) and ionomycin (1 μg/ml) for 4 hours in the presence of Brefeldin A (10 μg/ml; all Sigma-Aldrich) for flow cytometric analysis of intracellular IL-10. In experiments addressing involvement of lysosomal acidification, the inhibitor chloroquine (5 μM; Sigma) was added at the start of a two days culture and refreshed after 24 hours, and the TLR ligands CpG ODN 1826 (5 μg/ml; Invivogen) or Pam3Cys (10 μg/ml; Invivogen) used as control stimuli next to SEA.
For in vivo stimulation of B cells, mice were i.p. injected with two doses of 5000 eggs or 100 μg SEA in PBS, determined as optimal doses where B cell IL-10 production plateaued in prior dose-titration experiments, and PBS or 100 μg human serum albumin (HSA) in PBS as control 7 days apart. At day 14 after the first injection, splenic B cells were harvested and cultured in medium at 1.5 x 106 cells/ml or restimulated for 2 days with SEA (20 μg/ml) to allow detection of cytokines, as established for in vivo schistosome-exposed B cells before [12]. Supernatants were collected to determine cytokine concentration by ELISA. Cells were cultured for additional 4 hours with Brefeldin A (10 μg/ml; Sigma-Aldrich) to detect intracellular IL-10 by flow cytometry. In some experiments, mice were treated i.p. with 200 μg of hamster anti-mouse CD40 ligand blocking antibody (clone: MR1) or 200 μg hamster IgG as control (Jackson ImmunoResearch) 4 times, every 4 days starting one day before SEA or PBS treatment. For in vivo depletion of macrophages, mice were i.p. injected with 200 μl clodronate-containing liposomes and control mice with 200 μl of PBS liposomes (ClodronateLiposomes.com, Amsterdam, The Netherlands) [45] three weeks prior to egg antigen treatment. Successful and specific depletion of splenic macrophage subsets was confirmed by fluorescence microscopy and flow cytometry.
In vitro or in vivo SEA-stimulated CD19+ B cells were co-cultured with MACS-sorted CD4+CD25- T cells at 1:1 ratio (each 1 x 106/ml) to test for in vitro Treg cell induction. After 4 days, Treg cell frequencies were determined by flow cytometry by gating for Foxp3+CD25+ cells in the CD3+CD4+ T cell population, and culture supernatants collected for subsequent ELISA.
SEA, IPSE/alpha-1 and ovalbumin (OVA) were fluorescently labeled with PF-488 or PF-647 using the PromoFluor labeling kits (PromoCell, Heidelberg, Germany) according to the manufacturer’s protocol. For some experiments, SEA was co-labeled with the pH-sensitive pHrodo Red dye (Thermo Fisher Scientific). After protein labeling, non-reacted dye was removed using Zeba desalt spin columns (Thermo Fisher Scientific). For analysis of binding in vitro, CD19+ splenic B cells were cultured for 60 minutes at 37°C with 20 μg/ml fluorescently labeled SEA or 1–10 μg/ml of IPSE antigen, then washed in ice-cold PBS before analysis by flow cytometry. For analysis of in vivo binding, mice were i.v. injected with 200 μg of fluorescently labeled SEA or OVA as non-schistosomal control protein and spleens snap-frozen 30 minutes to 24 hours later. Binding of SEA to B cells was analyzed by confocal fluorescence microscopy of tissue sections and by flow cytometry.
Basophils were purified from 250 ml of peripheral blood of healthy human donors to a mean purity of 99% by a three-step protocol consisting of a density gradient centrifugation via Ficoll/Percoll (100/6, density 1.080 g/l), followed by enrichment of the basophils via counter flow elutriation and final purification by magnetic cell sorting using the basophil isolation kit II for negative selection of basophils (Miltenyi-Biotech). Purified basophils were cultured in Iscove's Modified Dulbecco's Media (IMDM; PAA) containing 2 mM glutamine (PAA), 5 μg/ml insulin (Gibco), 50 μg/ml apo-transferrin (Sigma-Aldrich), 100 μg/ml Pen/Strep (PAA), 10% heat-inactivated Fetal Calf Serum (FCS-Gold; PAA) and 2.5 ng/ml IL-3 (kind gift of Kirin Brewery, Japan). Basophils were pre-incubated for regeneration for 30 min at 37°C, 6% CO2, and then stimulated at a concentration of 0.025 x 106 basophils /ml in 96well flat-bottom culture plates in 100 μl at 37°C, 6% CO2. Concentration of stimuli was as indicated. Culture supernatants were collected after 18h and stored at -20°C.
Flow cytometric analysis of murine B cells was performed by staining with fluorochrome-labeled antibodies against CD19, CD21 (both BD Biosciences), CD23, CD40, CD86 or IL-10 (all eBioscience) after fixation with 1.9% paraformaldehyde and permeabilization with 0.5% saponin (Sigma-Aldrich). Human B cells were stained for CD19, CD38 (both BD Biosciences), CD24, CD27 (both eBioscience), CD1d, IL-10, TNF (all Biolegend), and CD39 (Sony Biotechnology, San Jose, USA). Splenic myeloid cell subsets were discriminated using fluorochrome-labeled antibodies against CD11b, CD11c (both eBioscience), CD8, Ly6C (both Biolegend), F4/80 (AbD Serotec, Puchheim, Germany), Gr-1 (BD Biosciences), and Siglec-1 (Dr. J. den Haan, VUMC, Amsterdam, The Netherlands). Treg cells were fixed and permeabilized with the eBioscience Foxp3 fixation/permeabilization kit and stained using fluorochrome-labeled antibodies against CD3, CD4, Foxp3 (all eBioscience) and CD25 (BD Biosciences). All cells were stained with Aqua dye (Thermo Fisher Scientific) prior to fixation to discriminate dead cells. For all flow cytometric stainings, FcγR-binding inhibitor (2.4G2) was added and FMOs were used for gate setting. Flow cytometry was performed using a FACSCanto or Fortessa (BD Biosciences).
The concentration of murine IL-6 and IL-10 as well as human IL-4 present in culture supernatants was quantified by commercial ELISA kits according to the manufacturer’s instructions (BD Biosciences or Eli-Pair, Diaclone).
Spleens were snap-frozen in O.C.T. medium (Tissue-Tek; Sakura, Alphen aan den Rijn, The Netherlands). Cryosections (10 μm) were fixed in ice cold acetone for 10 minutes, air-dried, and blocked in 1% BSA plus 20% FBS in PBS before staining with Abs at room temperature. Cryosections were incubated with rat anti-mouse Siglec-1 (clone SER-4; provided by Dr. J. den Haan, VUMC, Amsterdam, The Netherlands) followed by Alexa555-conjugated goat anti-rat IgG (Invitrogen), anti-SIGN-R1 Alexa647 (clone 22D1; Dr. J. den Haan) and anti-B220 eFluor450 (eBioscience). Images were acquired using a Zeiss LSM 710 confocal laser scanning microscope with Zen software (Carl Zeiss Microimaging, Jena, Germany).
All data are presented as mean ± standard error of the mean (SEM). Statistical analysis was performed with GraphPad Prism version 6.00 for Windows (GraphPad Software, La Jolla, CA, USA) using nonparametric Mann-Whitney U test to compare different groups and Wilcoxon paired test to compare B cell subsets. One-sample t-test of log-transformed data was applied to calculate significant changes for data which are expressed as fold increase. All p-values < 0.05 were considered significant.
All animal studies were performed in accordance with the Animal Experiments Ethical Committee of the Leiden University Medical Center (DEC-12204). The Dutch Experiments on Animals Act is established under European Guidelines (EU directive no. 86/609/EEC regarding the Protection of Animals used for Experimental and Other Scientific Purposes). For the isolation of B cells from PBMCs, human subjects were recruited within the framework of the study P09.170, which was approved by the Medical Ethical Committee of the Leiden University Medical Center. For the isolation of basophils, donors were recruited under approval by the Ethics Committee of the University of Luebeck (AZ-12-202A). Studies were performed according to the declaration of Helsinki and all participants were adults and have given written informed consent.
|
10.1371/journal.pgen.1003559 | Negative Regulation of the Novel norpAP24 Suppressor, diehard4, in the Endo-lysosomal Trafficking Underlies Photoreceptor Cell Degeneration | Rhodopsin has been used as a prototype system to investigate G protein-coupled receptor (GPCR) internalization and endocytic sorting mechanisms. Failure of rhodopsin recycling upon light activation results in various degenerative retinal diseases. Accumulation of internalized rhodopsin in late endosomes and the impairment of its lysosomal degradation are associated with unregulated cell death that occurs in dystrophies. However, the molecular basis of rhodopsin accumulation remains elusive. We found that the novel norpAP24 suppressor, diehard4, is responsible for the inability of endo-lysosomal rhodopsin trafficking and retinal degeneration in Drosophila models of retinal dystrophies. We found that diehard4 encodes Osiris 21. Loss of its function suppresses retinal degeneration in norpAP24, rdgC306, and trp1, but not in rdgB2, suggesting a common cause of photoreceptor death. In addition, the loss of Osiris 21 function shifts the membrane balance between late endosomes and lysosomes as evidenced by smaller late endosomes and the proliferation of lysosomal compartments, thus facilitating the degradation of endocytosed rhodopsin. Our results demonstrate the existence of negative regulation in vesicular traffic between endosomes and lysosomes. We anticipate that the identification of additional components and an in-depth description of this specific molecular machinery will aid in therapeutic interventions of various retinal dystrophies and GPCR-related human diseases.
| Malfunctioning of phototransduction is the major cause of human blindness. Without functional phototransduction, rhodopsin-1, the major visual pigment, is rapidly endocytosed and accumulated in late endosomes. Impaired lysosomal delivery of endocytosed rhodopsin and its degradation has been reported to trigger progressive and light-dependent retinal degeneration in Drosophila models. It is intriguing why endocytosed rhodopsin accumulates in late endosomes instead of being delivered to lysosomes for degradation. Is this attributable to a saturation of rhodopsin endocytosis, which impedes the delivery capacity of the cell? To investigate the underlying mechanisms of rhodopsin accumulation in late endosomes, we used a suppressor of phototransduction mutants, which was identified previously from our unbiased genetic screen. This suppressor, called diehard4, shifts the membrane balance between late endosomes and lysosomes, resulting in the facilitated degradation of endocytosed rhodopsin. Our results clearly demonstrate that a previously unknown mechanism of negative regulation is actively engaged in vesicular traffic between endosomes and lysosomes in fly photoreceptors. We showed that eliminating such blockage alone was enough to rescue retinal degeneration in phototransduction mutants. From these results, we anticipate that the identification of additional components and an in-depth description of this molecular machinery will aid in therapeutic interventions of various retinal dystrophies and neurodegenerative disorders.
| Retinitis pigmentosa is the most common form of retinal degeneration and the major cause of human blindness [1]. Most degenerative retinal dystrophies are caused by various genetic mutations. Malfunctioning of phototransduction is the predominant cause of retinal dystrophies, in that most genes involved in the functioning and regulation of the phototransduction cascade are directly or indirectly related to retinal degeneration syndromes [2], [3]. Therefore, it is not surprising that rhodopsin-1, the major visual pigment, was the first molecule identified as a target for such mutations [4], [5].
Drosophila norpA (phospholipase C, PLC) acts as a central effector molecule in phototransduction [6]. It has been used as an invertebrate model for studying molecular mechanisms of retinal degeneration caused by malfunctioning of the phototransduction cascade [7]. Interestingly, cGMP phosphodiesterase, which relays the signal from G-proteins in vertebrate phototransduction, is also known to trigger retinal degeneration in mouse models [8]–[10]. The loss of norpA function essentially shuts down the phototransduction cascade, resulting in a failure to raise intracellular Ca2+ levels through light-sensitive channels. Thus, Ca2+-dependent enzymes required for rhodopsin recycling cannot be activated, resulting in the formation of the stable rhodopsin-arrestin complex [11]–[14]. It has been reported that excessive endocytosis followed by the formation of stable rhodopsin-arrestin complexes and accumulation of internalized rhodopsin in late endosomes trigger apoptosis in norpA mutant photoreceptor cells [12].
The “granule group” genes in Drosophila have been known for their vital role in lysosomal biogenesis and functioning [15], [16]. A previous study found that the functional loss of the “granule group” genes resulted in rhodopsin accumulation in the Rab7-positive late endosomes and triggered retinal degeneration in norpA mutant photoreceptor cells [12], [17]. Therefore, the accumulation of internalized rhodopsin in late endosomes and impaired endo-lysosomal trafficking clearly causes retinal degeneration in both the norpA and the “granule group” mutant photoreceptors. However, the molecular basis of this pathologic accumulation remains unknown.
The role of excessive endocytosis of light-activated rhodopsin on saturating the capacity of the trafficking machinery for the endo-lysosomal progression, resulting in the accumulation of endocytosed rhodopsin in the late endosomes remains controversial. Alternatively, previously unknown regulatory mechanisms prevent endocytosed rhodopsin from further movement toward lysosome. A growing number of evidences support the fact that the eukaryotic cell utilizes active regulatory mechanisms in monitoring and maintaining the intracellular membrane balance of the endo-lysosomal system [18]–[21]. Therefore, it is imperative to identify genetic components underlying rhodopsin accumulation and present epistatic evidences that possibly override the endo-lysosomal blockage in phototransduction mutants.
Triplo-lethal (Tpl) locus, cytologically defined as the 83D4-E2 region in chromosome 3 in Drosophila, was identified as a sole locus responsible for both triplo-lethality and haplo-lethalith in segmental aneuploids [22]. Point mutations responsible for the Tpl phenotype remain unidentified [23], although the Ell product, a transcription elongation factor, was found to be a suppressor of the Tpl phenotype [24]. Therefore, it is proposed that this phenotype is caused by a gene cluster that shows at least partial redundancy and its dosage is critical to its function [25].
The Osiris gene family was identified in an effort to explain the Tpl phenotype as an effect of a gene cluster. This is a large conserved family, with most genes (20 of 23) located within the Tpl locus [26]. Although the cellular function of the Osiris family of proteins is currently unknown, all members share characteristic features, including endoplasmic reticulum signal sequences, a pair of cysteine residues near the amino terminus, a putative transmembrane domain, an AQXLAY motif, and a number of endocytic signaling motifs such as YXXØ [26], [27].
Previously, we screened for norpAP24 suppressors by random mutagenesis. The screening had the advantage of the yeast site-specific recombination FLP-FRT system and could identify both essential and nonessential genes [28]. Here we report that the novel norpAP24 suppressor, diehard4 (die4), is responsible for the inability of endo-lysosomal rhodopsin trafficking and retinal degeneration in norpAP24 mutants. We found that die4 encodes Osiris 21 (Osi21). A loss of function of Osi21 suppresses retinal degeneration in various phototransduction mutants. In addition, the loss of function shifts the membrane balance between endosomes and lysosomes, resulting in the facilitated degradation of endocytosed rhodopsin. Our results demonstrate that the existence of negative regulation in vesicular traffic between endosomes and lysosomes. This mechanism may trigger retinal degeneration in phototransduction mutants.
Drosophila norpA encodes eye-specific phospholipase C and acts as a central effector in phototransduction [6]. The norpA photoreceptor has been used as a model system for studying progressive retinal dystrophies in humans because the loss of its function leads to rapid light-dependent retinal degeneration as a result of excessive endocytosis of stable rhodopsin-arrestin complexes and accumulation of internalized rhodopsin in late endosomes [11], [12], [14]. Previous studies [29] have shown that the norpAP24 (a strong hypomorphic allele of norpA [30]) photoreceptor showed progressive retinal degeneration (Figure 1A–B). Its degenerative phenotype appeared within three days and was obvious within four days upon constant light exposure. We found that wild-type (Canton-S) flies showed no sign of retinal degeneration even after seven days of constant light exposure (Figure 1F), indicating that norpAP24 degeneration was strongly dependent on light.
The die4 mutant was previously identified as a norpAP24 suppressor from a genetic screen by using eye-specific FLP-FRT mosaic flies [31], delaying degeneration several days (Figure 1G). The die4 mutant was generated using ethyl methanesulfonate (EMS) mutagenesis, possibly bearing multiple mutations. In addition, mosaic screening enables identification of both lethal and non-lethal mutations. Because of these complexities, we used multiple mapping methods to identify the exact mutation responsible for the suppressive phenotype of the die4 mutant. We previously reported that the mutation in the cytological region of 32D5 to E4 of the die4 chromosome, is responsible for the suppressive phenotype [31]. Although the die4 chromosome is homozygous lethal, this mutation is viable, in that the genomic deficiency, Exel6028, failed to complement the suppressive phenotype of die4, and was still viable (Table S1).
In this context, we performed a complementation test of die4 with 11 genes deleted in Exel6028 to identify EMS-induced mutations responsible for the suppressive phenotype (Figure S1A, Table S2). We identified that the loss of Osi21 (CG14925) is responsible for the suppressive phenotype of die4, in that the Minos transposon-inserted allele of Osi21, Mi{ET1}Osi21 [32], failed to complement die4 and the introduction of the genomic fragment encompassing Osi21 reversed the suppressive effect of die4/Mi{ET1}Osi21 at the deep pseudopupil (DPP) level (Figure 1G, Table S2). Sequence analysis of the die4 chromosome revealed significant amino acid changes (G149S, M181T, and F229L) in Osi21 (Figure 2, Figure S1). These results were confirmed by targeted knock-down of Osi21 using RNAi method (Figure 1C). Therefore, the suppressive effect of die4 on norpAP24–triggered retinal degeneration is due to the loss of Osi21 function. We therefore conclude that die4 is a loss-of-function allele of Osi21.
Osi21 was identified as an Osiris family protein, without known cellular functions, on the basis of sequence homology [26]. Computational analysis, as described by Shah et al. [27], was performed using its amino acid sequence, which revealed that OSI21 includes (1) an endosome/lysosome sorting signal, (2) a two-Cys region, (3) duf1676 (Pfam family: PF07898), and (4) a YXXØ motif (Figure 2) as predicted by previous studies [26], [27], [33]. Interestingly, Osi21 is located on the 2L chromosome. Thus, Osi21 is not linked to the cluster of 20 Osiris family genes that are located in the Triplo-lethal region (Tpl) of the 3R chromosome and is responsible for the Tpl phenotype, suggesting that its cellular function differs from that of the other Osiris family proteins.
Drosophila rdgC encodes rhodopsin-specific phosphatase and requires rhodopsin recycling [34], [35]. The loss of rdgC function leads to light- and age-dependent retinal degeneration as a result of excessive endocytosis of stable rhodopsin-arrestin complexes [14]. Because the rdgC mutant photoreceptor cells share the cause of retinal cell death with norpAP24, we used rdgC306, the loss-of-function rdgC mutant, to test the effect of die4 on retinal cell death. DPP analysis and histological analysis using electron microscopy showed that die4 protects retinal degeneration due to rdgC306 (Figure 3A–C, Figure S2), suggesting that Osi21 is not a specific regulator in norpA-triggered retinal degeneration but plays an essential role in retinal degeneration caused by intracellular accumulation of cytotoxic rhodopsin.
We also examined the effect of die4 on retinal degeneration due to the loss of trp function, a light-sensitive Ca2+ channel [36]. Drosophila trp1 was recovered as a spontaneously occurring temperature-sensitive loss-of-function mutant at a temperature of 24°C [36], [37] and is known to show light-enhanced retinal degeneration [38]. Although a dysfunction in Ca2+ fluctuation was suggested as a cause of its retinal degeneration phenotype, its mechanism of degeneration remains unknown. Interestingly, we found that the loss of die4 function protects retinal degeneration caused by trp1 at the DPP and ultrastructural levels (Figure 3D–F, Figure S2). This result suggests that intracellular accumulation of cytotoxic rhodopsin also causes retinal degeneration in trp1 mutant photoreceptor cells.
We used rdgB2 photoreceptors as a negative control to assume the functions of die4 because cytoplasmic rhodopsin aggregation is not involved in retinal degeneration in rdgB2 photoreceptors [39]. As expected, die4 was unable to suppress rdgB2-triggered retinal degeneration (Figure 3G). Combined together, these double mutant analyses suggest that intracellular rhodopsin aggregation triggers unregulated cell death in norpAP24, rdgC306, and trp1 photoreceptors, and that Osi21 is a key regulator in the formation of rhodopsin aggregation, wherein the loss of Osi21 function suppresses retinal degeneration in these mutant photoreceptor cells.
Previous studies found that norpA and rdgC mutant photoreceptor cells die due to excessive endocytosis of rhodopsin-arrestin complexes and accumulation of endocytosed rhodopsin in late endosomes [11], [12], [14]. These findings indicate that the inability of rhodopsin transport and degradation through the endo-lysosomal system cause unregulated cell death in norpA and rdgC mutant photoreceptors. In this context, we tested the possibility that Osi21 acts as a regulator that maintains membrane homeostasis between endosomes and lysosomes in which the functional loss of Osi21 shifts the membrane balance of the endo-lysosomal system.
To test our hypothesis, we examined whole mounts of Drosophila retinas by confocal microscopy. We found that the loss of Osi21 function minimally affected the Rab5-positive vesicles (early endosomes) (Figure 4A–B) and didn't affect the Rab6-positive vesicles (Golgi complexes) (Figure 4C–D). However, the loss of Osi21 function significantly affected the Rab7-positive vesicles (late endosomes). Compared to the wild-type photoreceptor cells (Figure 4E–F), both size and number of Rab7-positive vesicles were greatly reduced in Osi21 knock-down photoreceptor cells (Figure 4G). Accordingly, lysosomal compartments proliferated in Osi21 knock-down photoreceptor cells (Figure 4I–J), suggesting that the membrane balance of endo-lysosomal trafficking shifted toward lysosomes. Interestingly, the loss of Osi21 function only affected the number, but not size, of the lysosomal compartments in Osi21 knock-down photoreceptors, thus reflecting limited lysosomal rhodopsin flow in newly eclosed flies.
The reduced Rab7-positive vesicles in Osi21 knock-down photoreceptor cells may be attributed to Gal4 titration due to the existence of a second UAS promoter of in the Osi21 knock-down construct. Thus, we used the w; Rhi1::Gal4, UAS::YFP-Rab7/+; UAS::RFP-arf72A/+ as a titration control and showed that the second UAS promoter did not affect YFP-Rab7 expression (Figure 4F). Quantification of each vesicle clearly showed that among the Rab5-, the Rab6- and the Rab7-positive vesicles, the loss of Osi21 function only affected the Rab7-positive vesicles (Figure 4K). Minimal increase of the Rab5-positive area in Osi21 knock-down photoreceptors may be the secondary effect caused by a reduction in Rab7-positive vesicles. Accordingly, biochemical analysis of isolated endo-lysosomal vesicles using Iodixanol density gradients (See Text S1) showed that immunoreactivities of Rab7, which was colocalized with endocytosed rhodopsin, were shifted toward the lower density fractions by the loss of Osi21 function (Figure S3), suggesting a reduced fraction of late endosomes in Rab7-positive vesicles [40], [41]. These results suggest that the loss of Osi21 function specifically affects the membrane balance between late endosomes and lysosomes.
Because the specific shift of membrane balance between late endosomes and lysosomes raised a strong possibility of direct regulation of Osi21 on membrane homeostasis of the endo-lysosomal system, we examined the subcellular localization of the OSI21 protein. Newly eclosed flies were reared in a light/dark cycled incubator and were exposed to bright light (2900 lux) for 90 min to induce massive rhodopsin endocytosis and its accumulation in late endosomes. The subcellular localization of OSI21-GFP from whole mount ommatidia was examined using confocal microscopy. We assumed that the OSI21-GFP is functional because its expression counterbalanced the suppressive effect of Osi21-RNAi in the DPP level (Data not shown). We found the OSI21-GFP localization partially overlapped with Lysotracker staining (Figure 5A and E, Pearson's correlation coefficient: 0.617). In addition, majority of Osi21-GFP colocalized with endocytosed Rh1-RFP (Figure 5B and E, Pearson's correlation coefficient: 0.635), suggesting that Osi21 functions directly on the endo-lysosomal membrane system in a way that Osi21 negatively regulates late endosomal membrane traffic toward lysosomes, resulting in rhodopsin accumulation in late endosomal compartments.
Changes in membrane balance between late endosomes and lysosomes may also affect the dynamics of vesicular traffic and the rate of rhodopsin degradation, in which the loss of Osi21 function facilitates rhodopsin traffic toward lysosomes and its lysosomal degradation, resulting in a delay of retinal degeneration in norpAP24 photoreceptor cells. In fact, reduced rhodopsin content due to vitamin A deprivation or mutation in the rhodopsin gene rescued norpA-triggered retinal degeneration [42]. Our analysis by confocal microscopy showed that, compared to the norpAP24 photoreceptor cells (Figure 5C), norpAP24 mutant photoreceptor cells with the Osi21-RNAi transgene showed greatly proliferated lysosomes (Figure 5D–E). These lysosomes were colocalized with endocytosed rhodopsin (Pearson's correlation coefficient: 0.604). Although we often found that small amounts of endocytosed rhodopsin escaped Osi21 blockage (Figure 5C, arrowhead) and was colocalized with the lysosome (Pearson's correlation coefficient: 0.342), there were less lysosomes in the control norpAP24 photoreceptor cells. Moreover, majority of lysosomes did not colocalized with endocytosed rhodopsin, indicating that such colocalization reflected regular lysosomal turnover. These results raise the possibility that the loss of Osi21 function facilitates the rhodopsin degradation in lysosomes.
In this context, first, we examined the rate of rhodopsin endocytosis and degradation by time course measurements of pulse-chased photoreceptors by confocal microscopy. For the measurements, we expressed RFP-tagged rhodopsin under the control of the hs::Gal4 driver. Newly eclosed norpAP24 and norpAP24; Osi21-RNAi flies were kept in complete darkness for 24 h and then subjected to heat-shock. These flies were then kept in the darkness for another day to allow synthesis and transport of Rh1-RFP to the rhabdomere, following which they were exposed to bright light to chase Rh1-RFP. Whole mount photoreceptors were examined by confocal microscopy at 24 h interval for 96 h by typing the photoreceptor based on the progression of rhodopsin endocytosis and degradation (Figure 6A). We found that the initial movement of endocytosed rhodopsin toward the endosomal system was not different between the norpAP24 and norpAP24; Osi21-RNAi photoreceptor (Figure 6B, see the percentage of Type I and Type II photoreceptors). However, norpAP24 mutant photoreceptor cells with the Osi21-RNAi transgene showed facilitated clearance of endocytosed rhodopsin (Figure 6B, see the percentage of Type III and Type IV photoreceptors), indicating facilitated degradation of endocytosed rhodopsin through the loss of Osi21 function.
The effect of the Osi21 loss-of-function on the rhodopsin contents in norpAP24 photoreceptors was also examined by western blot analysis by using flies eclosed within 12 h. Flies were reared either in the dark to prevent rhodopsin endocytosis or in 18 h light/8 h dark cycles to stimulate rhodopsin endocytosis. No significant differences in rhodopsin content were observed in the dark-reared norpAP24 mutant photoreceptor cell with the Osi21-RNAi transgene, compared to dark-reared norpAP24 photoreceptors (Figure 6C, lanes 2–3). However, the rhodopsin content was greatly reduced in the norpAP24 mutant photoreceptor cells with the Osi21-RNAi transgene by the bright light stimulation (Figure 6C, lanes 4–5). These results suggest that more endocytosed rhodopsin was transported into and degraded by lysosomes because of Osi21 loss-of-function.
Massive influx of Rh1 into the endosomal system may saturate endosomal trafficking machinery, resulting in the late endosomal accumulation of endocytosed rhodopsin. Therefore, we tested the effect of the Osi21 loss-of-function on rhodopsin content when the endosomal trafficking machinery was activated by overexpressing Rab5 or Rab7 (Figure S4). To induce the maximal rhodopsin endocytosis, newly eclosed flies were exposed to bright light (2900 lux) for 48 h. No significant influence of Rab5 overexpression on the degradation of endocytosed rhodopsin was observed (Figure 6D, left). However, overexpression of Rab7 synergistically accelerated rhodopsin degradation with the loss of function of Osi21 (Figure 6D, right). Considering no significant decrease in rhodopsin content was observed with Rab7 overexpression alone, our results suggest that Osi21 negatively regulated rhodopsin transport between late endosomes and lysosomes by interacting with the Rab7-positive trafficking machinery. Therefore, we conclude that Osi21 is a critical negative regulator of vesicular traffic between endosomes and lysosomes. Its functional loss suppresses retinal degeneration in phototransduction mutants by changing the membrane dynamics between late endosomes and lysosomes and by facilitating the degradation of endocytosed rhodopsin.
In both vertebrates and invertebrates, malfunctioning of phototransduction may stimulate the cell death machinery, resulting in retinal degeneration [43]. Without functional phototransduction, rhodopsin-1, the major visual pigment, is rapidly endocytosed and accumulated in the late endosomes [12]. Impaired lysosomal delivery of endocytosed rhodopsin and its degradation trigger progressive and light-dependent retinal degeneration in phototransduction mutants [12], [17]. However, the mechanism underlying the accumulation of endocytosed rhodopsin in late endosomes, instead of delivering to lysosomes for degradation, remains to be elucidated.
In the current study, we used die4, the norpAP24 suppressor, to investigate the molecular basis of the accumulation of rhodopsin in late endosomes in phototransduction mutants. We found that the loss of die4 function delays retinal degeneration in norpAP24, rdgC306 and trp1, but not in rdgB2. These results suggest that, at least, norpAP24, rdgC306, and trp1 photoreceptor cells die through a shared route. Previous research suggested that the generation of stable rhodopsin-arrestin complexes is the major cause of cell death in norpAEE5 [11] and rdgC306 [14]. The formation of stable rhodopsin-arrestin complexes in the norpA mutant photoreceptor is attributable to its inability to activate the calcium-dependent phosphatase, RDGC, which dephosphorylates rhodopsin (Figure 7). The calcium-dependent phosphatase also remains inactive in the trp1 photoreceptor upon light exposure since the cation specific calcium channel is lost in trp1 [36]. Therefore, all three phototransduction mutants share a common feature; the formation of stable rhodopsin-arrestin complexes.
On the other hand, norpAP24, rdgC306 and trp1 require light activation of rhodopsin but not subsequent phototransduction for retinal degeneration [44]. In contrast, rdgB2 requires both, whereby (1) rdgB2 flies fail to degenerate in complete darkness [44], (2) the rdgB2 retinal degeneration is rescued by norpAP24 [44], and (3) the rdgBKS222 retinal degeneration is rescued by trp1 [38]. These findings are used to infer that rdgB2 photoreceptor cells die via a different route.
We found that the loss of die4 function delays retinal degeneration in norpAP24 longer than those in rdgC306 and trp1. These results suggest that the blockage of endo-lysosomal trafficking by Osi21 is not the sole cause of retinal degeneration in rdgC306 and trp1 mutants. Recently, Sengupta et al. [45] proposed that PI(4,5)P2 depletion by NORPA underlies retinal degeneration in trpCM and trp343 mutants. Interestingly, both mutants exhibit faster light-dependent retinal degeneration than trp1 mutants. Preventing the formation of stable Rh1-Arr2 complexes by red light slows down the retinal degeneration in trpCM and trp343 mutants comparable to the trp1 degeneration in white light, suggesting that the endocytosis of Rh1-Arr2 complexes contributes retinal degeneration in trp mutants and different results are attributable in part to the allelic differences. In addition, PI(4,5)P2 depletion affects arrestin-mediated endocytosis [45], so that Rh1 internalization might be reduced in trp mutants. However, their results raise a strong possibility that prolonged activation of NORPA possibly contributes to degenerative syndromes in both trp and rdgC mutants.
Double mutant photoreceptor cells are eventually degenerated; they lost their DPP with extended exposure to bright light. Although DPP analysis does not provide a measure of the retinal degeneration process, it faithfully measures a complete loss of the ommatidial structure as it reaches the end of the degenerative process. DPP analysis in the current study suggests that the loss of Osi21 function delays the onset of retinal degeneration in norpAP24, rdgC306 and trp1 mutants. However, the loss of Osi21 function slows down the retinal degeneration in norpAP24, but not in rdgC306 and trp1: the slope of DPP loss was similar to that of the control soon after DPP loss occurred in double mutants. These results also suggest that the activity of Osi21 is not the sole cause of the rdgC and the trp degeneration.
We conclude that Osi21 acts as a negative regulator of endo-lysosomal membrane traffic between late endosomes and lysosomes. This conclusion is based on the following observations: (1) Both the size and number of the late endosomes are significantly reduced in Osi21 knock-down photoreceptor cells, (2) the lysosomal compartments are greatly proliferated in Osi21 knock-down photoreceptor cells, (3) the OSI21 protein is localized in the endo-lysosomal compartments, (4) the loss of Osi21 function in the norpAP24 mutant photoreceptor facilitates the degradation of endocytosed rhodopsin, and (5) Rab7 overexpression alone fails to affect the rhodopsin content of the norpAP24 photoreceptor. However, overexpression of Rab7 synergistically accelerates rhodopsin degradation with the loss of Osi21 function, suggesting that Osi21 directly interacts with the Rab7-positive trafficking machinery. These results clearly demonstrate that the existence of negative blockage regulating the membrane balance of the endo-lysosomal system, and not the capacity of endo-lysosomal trafficking machinery, causes retinal degeneration in phototransduction mutants.
Heptahelical G protein-coupled receptors (GPCRs) are considered the most diverse and therapeutically important family of receptors [46], [47]. Like many vertebrate GPCRs, light-activated rhodopsin-1 in Drosophila is rapidly phosphorylated by a specific kinase, called rhodopsin kinase. Phosphorylated rhodopsin-1 is desensitized by arrestins and is endocytosed to terminate further signaling activity (Figure 7). Because of this similarity, Drosophila rhodopsin-1 has been used as a prototype to study agonist-induced desensitization and internalization of vertebrate GPCRs [48].
In Drosophila photoreceptors, Arr2 promotes rhodopsin endocytosis and degradation when stable Rh1-Arr2 complexes are generated by loss of norpA or rdgC while Arr1 promotes rhodopsin endocytosis and recycling in the normal condition [14], [49]. Although Arr1 was previously reported to localize in endosomes [49], we found that Arr2 was absent in the endosomal system (data not shown), indicating Arr2 dissociates from Rh1 near the rhabdomeric membrane. This is reminiscent of functional classification of vertebrate GPCRs: Class A and Class B [50]–[52]. Thus, it can be clearly surmised that Drosophila photoreceptors operate two separate mechanisms of Rh1 endocytosis: (1) Arr1 for quenching and recycling, and (2) Arr2 for quenching and degradation. Since Arr2 is several folds more abundant than Arr1 in Drosophila photoreceptor cells to ensure rapid quenching of rhodopsin signaling for visual sensitivity [53], [54], the negative blockage by Osi21 may be evolved to counterbalance excessive Rh1 degradation as a result of Arr2 binding to activated Rh1. Therefore, it should be further investigated whether arrestins play roles in the decision between recycling and degradation for endosomal Rh1, and in the activation of cell death machinery.
Post-endocytic trafficking of GPCRs implicates in many human diseases [55]. Especially, stable rhodopsin-arrestin complexes in vertebrates also result in photoreceptor degeneration [56], [57]. In addition, cytoplasmic accumulation of proteins often implicates various neurodegenerative disorders, including the accumulation of rhodopsin in retinitis pigmentosa [56] and the accumulation of polyQ-expanded huntingtin in Huntington's disease [58]. Our results suggest that Osi21 regulation may underlie accumulation of disease-causing proteins in the endosomal compartment and that the elimination of Osi21 regulation may clean up this pathologic accumulation. Therefore, the identification and characterization of this specific cellular machinery may provide a therapeutic target for many GPCR-related human diseases and neurodegenerative disorders.
Drosophila was grown on standard food in a 25°C incubator. Standard genetic schemes were used to generate flies with the genotypes described. The Canton-S and w1118 fly were used as a wild-type strain, norpAP24, rdgC306, trp1 and rdgB2 as loss-of-function strains of phototransduction. Genomic deficiencies listed in Table S1. Loss-of-function mutants used for complementation test listed in Table S2. A w norpAp24 eyFLP chromosome was made using meiotic recombination to subsequently generate die4 mosaic flies in the norpAp24 background. P[ry+; w+]30C, P[ry+; hs-neo; FRT]40A, and P[w +]70C FLP recombinase target (FRT), die4 chromosomes and FRT40A GMR-hid were used in combination with the w norpAp24 eyFLP chromosome to make photoreceptor cells exclusively homozygous for the die4 FRT chromosome [59]. The second and third chromosome inserts of the Rh1::GAL4 driver were derived from the Rh1::GAL4 line constructed by Tabuchi et al. [60] and used for driving expression of various UAS targets including the die4 knock-down construct, UAS::Osi21-RNAi and fluorescently subcellular markers, UAS::YFP-Rab5, UAS::YFP-Rab7 and UAS::GFP-Rab6. UAS::Rh1-GFP was constructed using the prh1::eGFP construct from Pichaud and Desplan [61]. All Drosophila stocks except Osi21 knockdown strain were obtained from the Bloomington Stock Center at Indiana University. The UAS::Osi21-RNAi strain was obtained from Vienna Drosophila RNAi Center (VDRC, Vienna).
Primers for Osi21 or ninaE (Rh1) were specifically designed for use in the Gateway system (Invitrogen, Inc., Carlsbad, CA). Exact primer sequences for the rescue experiment, the expression of GFP-tagged Osi21 or RFP-tagged Rh1 were listed in Table S3. Directional cloning into the pENTR TOPO vector and the destination vectors (Carnegie Institution of Washington) followed manufacturer's instruction. The pTW, pTWG, and pTWRvector were used as destination vectors for Rescue constructs, UAS::Osi21-GFP, and UAS::Rh1-RFP, respectively. Plasmid isolation was performed from positive clones using the Qiagen Midi Kit (Valencia, CA). After injection, G0 flies were crossed with w; SM1/Sco; TM2/Sb balancer flies. The progeny from the cross were sorted for mini-w+ eyes. Mini-w+ was used as a marker to determine the presence of the transgene. Flies with the mini-w+ eye (G1 generation) were subsequently crossed with w1118 flies to map the location of the transgene. After mating with w1118 flies, mini-w+ flies were crossed to the driver stock. In all cases, mini-w+ flies were crossed to w; Rh1::GAL4 to drive expression of the transgene.
The deep pseudopupil (DPP) was visualized in red-eyed flies which shows a bright trapezoidal structure when a white light illuminates the retina from the back of the head [62]. The flies of each genotype were collected daily and raised under the appropriate light condition. The flies were also scored daily for the presence of the deep pseudopupil. During degeneration, deep pseudopupils become increasingly diffused before being completely lost. The deep pseudopupil was scored as negative as soon as its trapezoidal shape became indistinct. The percentage of flies that retained their deep pseudopupils for a given day was calculated. In figure 1, total 20 flies were analyzed for each genotype tested. In figure 2, three replicates with a total of 100 flies were analyzed for rdgC306, trp1 and rdgB2 with or without die4 to determine the average percentage of deep pseudopupil -positive flies and the standard error for each day.
Detailed procedures were also described previously [63]. Flies eclosed within 6 h were sacrificed with or without light treatment. For the whole-mount ommatidia isolation, fly heads were removed from bodies. A sagittal cut was made on the fly head creating two halves. The brain and proboscis were then removed. The eye was placed on a microscope slide containing 1× PBS. For Lysotracker (Invitrogen, Inc., Carlsbad, CA) staining, fly eyes were preincubated in ∼1 µM Lysotracker for 90 min and washed 3 times in 1× PBS for 30 min. Then, 2% paraformaldehyde in 1× PBS was used as a fixative and treated for 30 min followed by twice wash in 1× PBS. Residual pigments in fly retina were eliminated in 0.1% Triton X-100 (Sigma, MO) for 4 h, then washed twice for 10 min. Each ommatidia were removed using a sharp platinum needle from fly retina. Usually, a large piece of retina was then teased apart and mounted using mounting medium (Vector Labratories, Inc., Burlingame, CA). The FV500 confocal laser scanning microscope (Olympus Optical, Japan) was used for examining individual ommatidium. Optical images were acquired with an ×100 objective. Confocal images were analyzed with ImageJ software (NIH, MD) for quantitative analysis. For quantitative measurement of endosome/lysosome, mean size, number and total area of each vesicle in a photoreceptor cell were calculated with the Analyze Particle function. Measured values were normalized with the known distance option of imageJ. Statistical significances were calculated with two-tailed t tests using Prism 5.01 software. For colocalization analysis, Pearson's correlation coefficient (Rr) was calculated with Intensity colocalization analysis function of imageJ. The values for Rr range from 1 (perfect correlation) to −1 (perfect exclusion). Thus, a value close to 1 indicates reliable colocalization.
Retinal degeneration was examined with electron microscopy using retinal tissue sections. Fly eyes were prepared for electron microscopy using procedures described by Washburrn and O'Tousa (1992). Electron microscopy sections were ∼80–100 nm thick, stained first in 5% uranyl acetate in 50% EtOH and then in Reynold's lead citrate. The Hitachi H600 electron microscope was used to take electron micrographs. The micrographs shown in all figures are taken from ommatidia cross-sectioned at a depth of R1-R6 photoreceptor nuclei to present a similar view of various genotypes.
Usually, 2–5 fly heads were homogenized in buffer A (20 mM Tris-HCl(pH 7.5), 100 mM NaCl, 5 mM MgCl2, 10% sucrose, 1% glycerol, 1 mM EDTA, 1% CHAPS and Complete Protease Inhibitor Cocktail) with pellet pestle. The homogenate was centrifuged at 4°C and 14,000 rpm for 2 min. The supernatant was separated on 12% SDS polyacrylamide gel and then transferred to PVDF membrane at 100 V for 1 hour. The membrane was blocked by 5% nonfat milk in TBST (TBS with 0.5% Tween 20) for 60 min. After blocking, the membrane was incubated with the mouse anti-rhodopsin antibody 4C5 (diluted 1∶5000, Developmental Studies Hybridoma Bank) for 60 min at room temperature or overnight at 4°C. The polyclonal rabbit anti-arrestin2 antibody (Genscript, NY) is diluted 1∶30,000. The anti-mouse or rabbit IgG HRP conjugated antibody was diluted 1∶5000 in TBST containing 5% skin milk and the membrane was washed with TBST for 30 min. The blotted membrane was detected with a homemade ECL solution for 1 min, and then exposed to X-ray film. The detected bands were quantified using the Quantity One (Bio-Rad).
Newly eclosed norpAP24 and norpAP24; Osi21-RNAi flies with UAS::Rh1-RFP under control of hs::Gal4 were kept in complete darkness for 24 h and then subjected to heat-shock for one hour in the 37°C incubator three times at six-hour interval. These flies were kept in the 25°C incubator for another day and then moved under 2900 lux light. Whole mount photoreceptors were examined by confocal microscopy at 24 h intervals for 96 h. At each time point, approximately 30 photoreceptor cells from at least five individuals were scored and categorized as follows: Type I (most Rh1-RFP localizes in the rhabdomere), Type II (Rh1-RFP localizes equally in the rhabdomere and the cytoplasm), Type III (most Rh1-RFP localizes in the cytoplasm) and Type IV (most Rh1-RFP disappears due to degradation). The Kolmogorov-Smirnov test for equality distribution was performed using the STATA software package.
|
10.1371/journal.pcbi.0030184 | Bistability and Oscillations in the Huang-Ferrell Model of MAPK Signaling | Physicochemical models of signaling pathways are characterized by high levels of structural and parametric uncertainty, reflecting both incomplete knowledge about signal transduction and the intrinsic variability of cellular processes. As a result, these models try to predict the dynamics of systems with tens or even hundreds of free parameters. At this level of uncertainty, model analysis should emphasize statistics of systems-level properties, rather than the detailed structure of solutions or boundaries separating different dynamic regimes. Based on the combination of random parameter search and continuation algorithms, we developed a methodology for the statistical analysis of mechanistic signaling models. In applying it to the well-studied MAPK cascade model, we discovered a large region of oscillations and explained their emergence from single-stage bistability. The surprising abundance of strongly nonlinear (oscillatory and bistable) input/output maps revealed by our analysis may be one of the reasons why the MAPK cascade in vivo is embedded in more complex regulatory structures. We argue that this type of analysis should accompany nonlinear multiparameter studies of stationary as well as transient features in network dynamics.
| Molecular studies of cell communication systems lead to models with multiple free parameters. Analysis of dynamical behavior of these models presents considerable challenge. We have developed a computational approach for the efficient exploration of dynamic behavior in such models and applied this method to the model of the Mitogen Activated Protein Kinase cascade, a signaling network conserved in all eukaryotes. Previous analysis of this model suggested that it works as a reversible switch. We have shown that it can also function as an irreversible switch and as a clock.
| Physicochemical models of signaling pathways are characterized by high levels of structural and parametric uncertainty [1–7], reflecting both incomplete knowledge about signal transduction and the intrinsic variability of cellular processes. As a result, these models try to predict the dynamics of systems with tens or even hundreds of free parameters [8–10]. At this level of uncertainty, model analysis should emphasize statistics of systems-level properties, rather than the detailed structure of solutions or boundaries separating different dynamic regimes [11–18]. Chemical network theory and monotone systems approaches can characterize dynamics of biochemical networks based only on their structure, independently of a particular choice of parameters [19–21]. Under certain conditions, these methods can rule out whole classes of behaviors, such as bistability or oscillations, but they do not provide information about the relative prevalence of coexisting dynamic patterns. At the other extreme of model analysis techniques are continuation algorithms, which track steady states or limit cycles as a function of just one or two model parameters at a time [9,22].
While the information provided by continuation methods is only local, they can be efficiently combined with random parameter sampling algorithms, enabling the statistical exploration of systems-level properties, such as stability and robustness [23,24]. Here, we use this approach to characterize the statistics of steady-state input/output maps in the model of the Mitogen Activated Protein Kinase (MAPK) cascade, a network present in all eukaryotic cells and one of the most extensively modeled signaling systems [25]. The first model of the MAPK cascade was developed by Huang and Ferrell, and used as a basis for connecting the structure of the cascade and its dynamics (Figure 1A). Based on mass-action kinetics, the model described the dynamics of 22 species participating in ten reactions [26]. Each of the 37 model parameters, which have been either estimated or extracted from cellular and biochemical experiments, was specified within a reasonably broad interval. Huang and Ferrell hypothesized that the three-tiered structure of the MAPK cascade controls its steady-state input–output behavior. Based on simulations with hundreds of randomly generated parameter sets, they found that the input–output map is ultrasensitive. Importantly, this prediction was supported by biochemical experiments in Xenopus oocyte extracts [26].
In a later sequence of papers, Ferrell and co-workers demonstrated that ultrasensitivity can lead to bistability in positive feedback networks, in which the activated MAPK positively regulates the input to the cascade [27–29]. Recently, however, Kholodenko and co-workers have established that bistability is possible at the level of a single stage of the MAPK cascade [30]. Specifically, when the same phosphatase (e.g., MAPK'Pase) dephosphorylates both the monophosphorylated and double-phosphorylated forms of the substrate (e.g., MAPK), the double-phosphorylated form competitively inhibits the second dephosphorylation. In combination with the conservation of the total amount of substrate, this generates an equivalent of a direct positive feedback and can lead to bistability [30,31]. The extent to which this single-stage phenomenon influences the dynamics of the entire MAPK cascade has been unclear. Here, we demonstrate that a significant fraction of the multidimensional parameter space in the Huang-Ferrell model exhibits bistability and oscillations. Furthermore, our computational results strongly suggest that single-stage bistability is a necessary condition for the oscillatory behavior at the cascade level.
We used a combination of parameter sampling and continuation algorithms to characterize the statistics of input–output (I/O) maps in the Ferrell-Huang model [26]. Just as in the original publication, the I/O map describes the system response, taken to be the fraction of MAPK in the double-phosphorylated state, as a function of a distinguished model parameter, the input to the first stage of the cascade (Figure 1A). Specifically, the 36-dimensional vector of the remaining model parameters was repeatedly generated by Monte Carlo sampling from the hypercube defined by Huang and Ferrell (Table S1). For each of the generated parameter sets, a pseudoarclength-continuation algorithm was used to compute the steady-state I/O map [32]. This approach can both locate steady states and characterize their stability as a function of the input to the cascade. We developed a classification procedure for assigning the I/O maps to one of the three categories: “single-valued,” “oscillatory,” and “hysteretic” (Figure 1B; see Protocol S1 for details of the sampling, continuation, numerical stability analysis, and classification protocols).
The summary of the classification results, based on 20,000 parameter sets, is presented in Figure 2. We found that ∼80% of the generated models led to single-valued I/O maps (Figure 2A). Surprisingly, the rest of the generated models corresponded to strongly nonlinear I/O maps. Specifically, ∼10% of models had I/O maps with regions of oscillations (Figure 2B), while ∼10% of models were bistable (Figure 2C; see Table S2 for examples). While the existence of bistable I/O maps could have been expected on the basis of the single-stage results by Kholodenko et al., our results provide the first evidence of oscillatory behavior in the MAPK cascade in the absence of explicit negative feedback [30,33]. The large sample size in our calculations ensured tight confidence intervals for these estimates of the frequencies of the three different classes of I/O diagrams (see also Figure S2). All of the bistable I/O maps had their left-most turning point for positive values of the input. Thus, we did not observe bistability at zero values of the input; such diagrams were proposed to mediate irreversible cell-fate transitions in Xenopus oocyte maturation [29].
Based on the results of our sampling/continuation approach, we characterized the statistical properties of the I/O maps. By fitting the single-valued I/O maps to Hill functions, we found that, with high probability, they are ultrasensitive, i.e., are characterized by high Hill constants (nH > 1), Figure S1). In particular, with probability ∼74%, single-valued I/O map is characterized by a Hill coefficient greater than 2: P(nH > 2) ≈ 0.74. Focusing on the hysteretic and oscillatory maps, we established that they involve concentration ranges that can be adequately described by a deterministic approach, i.e., they are characterized by reasonably large molecular copy numbers for all of the model components (assuming the volume of an oocyte cell is ∼1 μL, a concentration even as low as 10−9 μM still corresponds to approximately 600 molecules). The oscillatory solutions in the model were of the relaxation type, their amplitudes spanned the entire dynamic range of the outputs (from unphosphorylated to fully phosphorylated MAPK, Figure S3E), and their periods were quite long (typically ½ hour, Figure S4). See Figure S3 for a summary of the statistical properties of oscillatory and bistable regimes.
The upper and lower boundaries of the suggested range for each of the parameters in the original Huang-Ferrell paper were given by one-fifth and five times the mean parameter value, respectively [26]. Using our sampling/continuation approach, we found that oscillatory and bistable I/O maps occur for much smaller ranges of parametric uncertainty (Figure 3). Thus, the existence of deterministic oscillations and bistability is a robust property of the Huang-Ferrell model.
In the next set of computational studies, we explored the origin of oscillatory and bistable regimes. To simplify the notation, we label the different stages of the full MAPK cascade, i.e., the activation of MAPKKK, double-phosphorylation of MAPKK, and MAPK, with the numbers 1, 2, and 3, respectively, and use terms like “system 2”, “system 2+3” or “system 1+2+3” to indicate different reaction networks consisting of a single stage, two consecutive stages, or all stages of the full MAPK cascade, respectively. As a first step towards the analysis of the full model, we used our sampling/continuation approach to characterize the statistics of I/O maps in all possible single-stage and two-stage subsets of the full model (Table 1).
As expected on the basis of previous analytical and computational results [30,34–36], we observed that the first stage is always monostable, while the second and third stages, each of which is formed by two consecutive double phosphorylation–dephosphorylation cycles, supports bistability. While bistability exists already at a single-stage level, our results strongly suggest that the emergence of oscillations requires at least two stages, one of which should be based on double phosphorylation (Table 1). Based on this, we hypothesized that the existence of single-stage bistability is a necessary condition for oscillations in multistage networks.
To test this hypothesis, we checked whether multistage networks with oscillatory I/O maps contain bistable single stages as their building blocks. Remarkably, for all possible multistage networks, i.e., system 1+2+3, 1+2, 2+3, we observed that oscillatory behavior requires at least one bistable single-stage module, e.g., stage 2 or 3 being bistable for the 1+2+3 system (Table 2). Note that there are no qualitative differences between the two-stage and three-stage cascade networks, with respect to their ability to support bistability and oscillations. Interestingly, this correlation between single-stage and multistage dynamics does not necessarily hold for bistable I/O maps (Table 2): multistage bistability can emerge from coupling of monostable stages.
We subsequently analyzed the connection between multistage limit cycles and single-stage bistability. As expected from the established correlation between single-stage bistability and multistage oscillations, we found that, in all cases, multistage limit cycles are “built” around hysteresis loops of bistable single stages (Figure 4B shows an example of such a correlation). This might explain the predominantly relaxation character of the oscillations in the MAPK cascade (see above); this strongly suggests the relation between the modularity of the network structure and modularity of network dynamics.
By analyzing the rates of individual reactions along the limit cycle, we established that multistage oscillations rely on the backwards coupling between a bistable stage and the preceding stage in the cascade (e.g., Figure 4A). Specifically, when the bistable stage is in the “off” state (point “a” in Figure 4B), the kinase which carries out both of the phosphorylations within this stage is complexed with its substrates. As a consequence, it is protected from dephosphorylation by the phosphatase in the preceding stage, and the total concentration of the kinase gradually increases (r1 > r2 in Figure 4C). However, when the bistable stage switches to the “on” state (point “b” in Figure 4B), at a high total concentration of the kinase, this kinase runs out of substrates and itself becomes a substrate for the upstream phosphatase. As a result, the total concentration of this kinase decreases (r1 < r2 in Figure 4C). At low levels of kinase activity, the substrates of this kinase within the bistable stage quickly become dephosphorylated, and, eventually, the stage quickly undergoes the transition back to the “off” state. We have established that this simple sequence of events accounts for oscillations in all observed multistage systems within the MAPK cascade (Table 1). Thus, the oscillatory solutions, which were identified on the basis of a brute force computational approach, turned out to have a transparent mechanistic origin.
Finally, we assessed the possibility of synthesizing the multistage oscillations from individual components. For this, we estimated the probability that a single, randomly generated bistable stage would lead to oscillations when embedded within the MAPK cascade (Table 3). The results of this analysis strongly suggest that single-stage bistability is a necessary but not a sufficient condition for multistage oscillations. The same results also show that single-stage bistability is also not sufficient for generating the bistable multistage I/O maps. At the same time, the odds of observing cascade-level oscillations are greatly increased (more than 3-fold) by the presence of single-stage bistability (based on the data in Table 3).
Based on the combination of random parameter search and continuation algorithms, we developed a methodology for the statistical analysis of mechanistic signaling models. In applying it to the well-studied MAPK cascade model, we discovered a large region of oscillations and explained their emergence from single-stage bistability. At this time, it is unclear whether such oscillations and bistability exist within the isolated MAPK cascade. However, our results suggest that oscillations and bistability do not necessarily imply the presence of explicit feedback loops.
The surprising abundance of strongly nonlinear (oscillatory and bistable) input/output maps revealed by our analysis may be one of the reasons why the MAPK cascade in vivo is embedded in more complex regulatory structures [9]. Numerous feedbacks targeting the MAPK circuit may either enhance the nonlinear behavior, e.g., by extending the range of inputs supporting bistability and oscillations, or eliminate it altogether, converting the switch-like behavior into a graded I/O response. In addition to feedbacks, synthesis and degradation of pathway components or their nucleocytoplasmic shuttling can affect the MAPK cascade dynamics [37–40]. The effects of these processes on the cascade dynamics can be systematically explored within our continuation/sampling approach.
Our objective has been to characterize the relative abundance of qualitatively different types of I/O maps. The rapid convergence of these estimates is an intrinsic feature of the Monte Carlo integration algorithms, which have been used in computational statistical physics for more than half a century. Hence, these kinds of approaches to statistical exploration of network dynamics will be effective whenever the outcomes of computations can be assigned to a finite number of classes. In our case, the outcomes of continuation runs were classified as “single-valued,” “oscillatory,” and “hysteretic” (see Protocol S1). In a different context, it may be important to characterize the statistics of transients induced by changes in the network inputs [41–43]. Given an appropriate classifier for transient solution features, one can identify the regions of the parameter space that lead to either adapting or sustained responses [40,42,44].
Recent single-cell measurements of protein levels show that they are characterized by high levels of variability. For example, measurements with GFP-labeled proteins in yeast and mammalian cells reported coefficients of variation around 20% [45,46]. Within this context, one can ask how robustly it is possible to guarantee a given type of network function. A computational approach to addressing this question can rely on the combination of a simple probability model for protein levels with a deterministic continuation algorithm. In this way, one can estimate the probability that a given I/O map will change its class, e.g., become oscillatory instead of hysteretic, when the model parameters are sampled from the multivariable distribution localized in parameter space.
Figure 5A presents an illustrative example of this type of calculation. Here we took the single-valued I/O map and perturbed it by sampling the parameters from the multivariable normal distribution, with means equal to the base values of parameters in the Huang-Ferrell model and coefficients of variation equal to 0.2. For this particular choice of the base model parameters and probability model, the I/O map remains single-valued (see Table 4), i.e., the classification of the I/O map as single-valued is robust. This is not, however, true in general, since in other regions of the parameter space one can easily find single-valued I/O maps that become either oscillatory or hysteretic upon localized variations of model parameters (unpublished data). Given the fact that these types of calculations are quite inexpensive at this time, we argue that this type of analysis should accompany multiparameter nonlinear studies of network dynamics.
Another motivation for a more detailed analysis of the distribution of different types of I/O maps in the multidimensional parameter space is provided by problems related to the evolutionary dynamics of signaling networks [47,48]. Mutations in the genes which encode components of signaling networks can affect both the protein levels and the rate-constants for protein/protein interactions. One can think that mutations in the regulatory sequence may translate into protein abundance, while mutations in the coding sequence may affect the protein activity and, hence, the rate constants in the model [43]. Depending on their location within the gene sequence, these changes can lead to either small or large shifts in the space of model parameters. Given a model of a mutational process and a biochemical and biophysical understanding of the connection between the gene sequence and protein abundance, one can systematically explore the connection between the dynamics of the genotype and network dynamics. For example, one can ask how easily a given mutational process can lead to a qualitative change of the I/O map. As an example, we computed the class change probabilities of the three different I/O maps in the Huang-Ferrell model upon simulated gene deletions and duplications (Figure 5B, Table 4). A similar type of approach may prove useful for interpreting the population level data on sequence variations in genes within the MAPK and other signaling pathways [49].
The mathematical model of the MAPK cascade, described in Text S1, can be reduced to an equivalent Ordinary Differential Equation (ODE) system (Text S2). The procedure of Monte-Carlo sampling, pseudoarclength continuation, and categorization of the steady-state I/O maps for the reduced ODE system is described in Protocol S1. Numerical integration, used in obtaining initial guesses for steady states and for approximating oscillatory solutions, was performed using the stiff solver ODE15S in MATLAB, a commercial software package available at http://www.mathworks.com/. Numerical computations of steady-state solutions and stability/bifurcation analysis were performed in MATLAB code. The statistical frequencies in Figures 2 and 3 and Tables 1–4 are reported with 95% confidence intervals. |
10.1371/journal.pmed.1002088 | Availability and Use of HIV Monitoring and Early Infant Diagnosis Technologies in WHO Member States in 2011–2013: Analysis of Annual Surveys at the Facility Level | The Joint United Nations Programme on HIV and AIDS (UNAIDS) 90-90-90 targets have reinforced the importance of functioning laboratory services to ensure prompt diagnosis and to assess treatment efficacy. We surveyed the availability and utilization of technologies for HIV treatment monitoring and early infant diagnosis (EID) in World Health Organization (WHO) Member States.
The survey questionnaire included 14 structured questions focusing on HIV testing, cluster of differentiation 4 (CD4) testing, HIV viral load (VL) testing, and EID and was administered annually from 2012 to 2014 through WHO country offices, with each survey covering the previous 12-mo period. Across 127 targeted countries, survey response rates were 60% in 2012, 67% in 2013, and 78% in 2014. There were encouraging trends towards increased procurement of CD4 and VL/EID instruments in reporting countries. Globally, the capacity of available CD4 instruments was sufficient to meet the demand of all people living with HIV/AIDS (PLWHA), irrespective of treatment status (4.62 theoretical tests per PLWHA in 2013 [median 7.33; interquartile range (IQR) 3.44–17.75; median absolute deviation (MAD) 4.35]). The capacity of VL instruments was inadequate to cover all PLWHA in many reporting countries (0.44 tests per PLWHA in 2013 [median 0.90; IQR 0.30–2.40; MAD 0.74]). Of concern, only 13.7% of existing CD4 capacity (median 4.3%; IQR 1.1%–12.1%; MAD 3.8%) and only 36.5% of existing VL capacity (median 9.4%; IQR 2.3%–28.9%; MAD 8.2%) was being utilized across reporting countries in 2013. By the end of 2013, 7.4% of all CD4 instruments (5.8% CD4 conventional instruments and 11.0% of CD4 point of care [POC]) and 10% of VL/EID instruments were reportedly not in use because of lack of reagents, the equipment not being installed or deployed, maintenance, and staff training requirements. Major limitations of this survey included under-reporting and/or incomplete reporting in some national programmes and noncoverage of the private sector.
This is the first attempt to comprehensively gather information on HIV testing technology coverage in WHO Member States. The survey results suggest that major operational changes will need to be implemented, particularly in low- and middle-income countries, if the 90-90-90 targets are to be met.
| Global 90-90-90 targets have reinforced the importance of functioning laboratory services to ensure prompt diagnosis and to assess treatment efficacy.
World Health Organization (WHO) guidelines for the provision of antiretroviral therapy (ART) for people living with HIV emphasize the need to increase access to HIV testing, treatment, and viral load monitoring.
In order to understand gaps in laboratory capacity to support further ART scale-up, WHO assesses country capacity to provide HIV diagnosis and treatment monitoring tests, how efficiently HIV diagnosis and treatment monitoring technologies are utilized, and the need for scale-up to meet predicted demand.
We obtained data on the availability and utilization of technologies for HIV treatment monitoring and early infant diagnosis (EID) across 60% of 127 countries surveyed from 2012–2014.
Globally, the capacity of available CD4 instruments was found to be sufficient to meet the demand of all people living with HIV/AIDS (PLWHA), irrespective of treatment status; however, capacity to measure viral load was inadequate to cover needs in most reporting countries.
Of concern, only 13.7% of existing CD4 capacity and only 36.5% of existing viral load (VL) capacity was being utilized across reporting countries in 2013. Lack of reagents, equipment not being installed or deployed, maintenance, and staff training requirements were among the reported reasons for underutilization.
Overall, among the countries surveyed, increasing numbers of CD4 and VL/EID instruments are being procured.
However, capacity to measure viral load is insufficient to meet current needs, and underutilization of CD4 and VL technology is widespread.
Major improvements in laboratory capacity utilization are needed in low- and middle-income countries if the 90-90-90 targets are to be met.
| The ambitious Joint United Nations Programme on HIV and AIDS (UNAIDS) 90-90-90 targets require coordinated action to ensure that by 2020, 90% of all people living with HIV know their HIV status, 90% of all people diagnosed with HIV infection receive antiretroviral therapy (ART), and 90% of those receiving ART achieve durable viral suppression [1]. Currently, an estimated 16 million people are receiving ART. Rapid ART scale-up has reinforced the importance of strengthening laboratory services now considered as a critical component of a health system to increase access to ART and to improve the quality of treatment and care for people living with HIV/AIDS (PLWHA) [2–4]. The 2015 World Health Organization (WHO) guidelines [2] recommending that ART is prescribed to all people as soon as possible after a HIV-positive diagnosis regardless of CD4 cell count imply a significant rise in the number of people who need to be started and maintained on treatment. Progress towards the 90-90-90 targets will require a significant expansion of HIV testing to diagnose HIV infection and to monitor treatment efficacy in a robust tiered laboratory network [5,6].
Expanding access to treatment requires high-quality HIV testing technologies, including CD4 testing to assess risk of disease progression, viral load (VL) testing to monitor treatment efficacy, early infant diagnosis (EID) to determine HIV-infection status in HIV-exposed children, and other monitoring capabilities within a tiered laboratory network. Technologies that can be used at point of care (POC) provide an important opportunity to expand access to HIV-related testing [7,8]. The availability and utilization of HIV EID and treatment monitoring technologies in many HIV/AIDS endemic countries have not been formally assessed. A detailed analysis is needed if we are to effectively tackle future challenges to ART scale-up. To this end, in 2012, WHO started to conduct annual surveys to assess the availability and the utilization of CD4, VL, and EID testing technologies in WHO Member States. Full and detailed datasets supporting our findings have been provided [9]. We present 3-y survey data and an assessment of trends of instruments available and tests performed, as well as an analysis of potential theoretical capacity versus the demand.
An English-language electronic questionnaire survey tool used annually by WHO to assess ART use in WHO Member States was revised to include 14 structured questions on the availability, functionality, and utilization of CD4, VL, and EID laboratory technologies and the market share for different branded technologies. The questionnaire survey is included as S1 Survey Questionnaire.
The 2012 survey covers 1 January 2011 to 31 December 2011, the 2013 survey covers 1 January 2012 to 31 December 2012, and the 2014 survey covers 1 January 2013 to 31 December 2013. Throughout this article, data from the 2012 survey will be referred to as 2011 data, data from the 2013 survey as 2012 data, and data from the 2014 survey as 2013 data.
Questionnaires were distributed via WHO regional offices to WHO country HIV officers supporting Ministry of Health HIV programme managers, who used data from their national annual reports to complete the survey within 4 mo (April–July of each survey year). Data collection was focused on the public health sector, which includes ministry of health and various nongovernmental organization (NGO) HIV programmes. The study did not cover the private sector. Unlike NGOs, the private sector does not report data to the national HIV programme. Questionnaires were returned by email to WHO’s Department of HIV/AIDS for data cleaning, verification, and analysis.
Surveys were sent to 127 countries, distributed as follows by WHO region:
The completed questionnaires were jointly collected by WHO country offices, the six WHO regional offices (WHO Regional Office for Africa, Brazzaville; WHO Regional Office for the Americas, Washington; WHO Regional Office for the Eastern Mediterranean, Cairo; WHO Regional Office for Europe, Copenhagen; WHO Regional Office for South-East Asia, New Delhi; and WHO Regional Office for the Western Pacific, Manila) and WHO headquarters, Geneva.
For the analysis of deployment of existing HIV EID and monitoring technologies, additional data were used. Specifically, numbers of PLWHA were retrieved from the UNAIDS AIDSinfo Online Database [10]. The midestimate of the number of PLWHA was used for the analysis. When data on PLWHA were unavailable from the above source, information was extracted from the latest country report available on the UNAIDS website [11]. If the country report provided the number of PLWHA only for 1 y of interest, that number was used for all 3 reporting years. If the country report provided only the number of diagnosed HIV patients, that number was rounded up to the next 100. Data on the number of PLWHA on ART were provided by responding countries. Testing technologies were categorized as conventional or POC according to definitions applied by the manufacturer.
The lowest published theoretical number of tests that an instrument can perform per technician per day was used to calculate theoretical capacity of CD4 and VL instruments in all countries responding to at least one survey [12–15]. Capacity analyses were based on the assumption that personnel work an average of 8 h per d, 250 d per y. The internationally accepted number of work days (260) was reduced in consideration of national holidays. The average number of national holidays was calculated for a subset of 18 countries and applied to all countries participating in these surveys [16].
Statistical analysis was performed in XLSTAT statistical application for Microsoft Excel (Addinsoft). Data visualization was done using Tableau software (Tableau Software). The p-values for the analysis of trends in instrument market share, nonutilization, and maintenance contracting and servicing were calculated using a chi-square test. In the case of rejection of the null hypothesis of multiple proportions equality, the chi-square test was followed by the Marascuilo procedure employed to simultaneously test all possible pairs of proportions and to identify the proportion(s) responsible for the rejection of the null hypothesis. A p-value of <0.05 was considered significant.
Survey response rates, based on 127 targeted countries, were 60% (76 countries) in 2012, 67% (85 countries) in 2013, and 78% (99 countries) in 2014 (countries that responded to at least one diagnostic survey question). Over the 3 survey years, 55 (43%) countries responded to all three surveys, 35 (28%) countries to two surveys, 25 (20%) countries to one survey, and 9 (7%) responded to none of the three surveys. As each survey covers the previous 12-mo period, the results refer successively to year 2011, 2012, and 2013.
Reporting countries have accumulated CD4-testing capacity sufficient to meet WHO recommendation to perform CD4 tests upon HIV diagnosis and every 6 to 12 mo thereafter. CD4 testing is not necessary for a stabilized patient on ART with suppressed VL, if VL testing is available (Table 1) [5,6].
Across 40 countries reporting CD4 instrument data for 3 consecutive y, the number of CD4 instruments has risen every year between 2011 and 2013, mostly due to the procurement of CD4 instruments for use at POC, which rose from the previous year by 36% in 2012 and by 53% in 2013. Growth for conventional CD4 instruments, however, was only 11% in 2012 and 4% in 2013 in countries responding to all three surveys (Fig 1). As a result, the theoretical capacity of CD4 conventional and POC instruments increased from 66.3 million tests in 2011 to 90.6 million tests in 2013, which equates to a theoretical CD4 testing capacity of 4.59 (median 5.21; IQR 3.07–9.57; median absolute deviation [MAD] 2.56) tests per PLWHA in 2011, 5.17 (median 6.71; IQR 4.10–11.64; MAD 3.31) tests in 2012, and 6.25 (median 7.27; IQR 3.51–11.32; MAD 3.83) tests in 2013 (Fig 2).
Three conventional CD4 instruments, BD FACSCount, CyFlow Counter, and BD FACSCalibur, and only one POC instrument, Alere Pima Analyser, have dominated the CD4 technology market for 3 consecutive y (Table 2). In 2011, 2012, and 2013, 74%, 77%, and 72% of reporting countries utilized WHO prequalified (WHO PQ) CD4 technologies (X2 [2; 214] = 0.5, p = 0.771, effect size [ES] Cramer’s V = 0.03).
All countries with CD4 instruments and with information on number of PLWHA were analysed and the theoretical capacity per PLWHA was measured: there were 66, 71, and 68 countries with data available for number of CD4 instruments and number of PLWHA, respectively in 2011, 2012, and 2013. CD4 capacity was sufficient to cover CD4 testing demand in responding countries, with theoretical CD4 capacity across 68 reporting countries in 2013 of 4.62 tests per PLWHA (median 7.33; IQR 3.44–17.75; MAD 4.35) (Fig 3).
All countries that reported the number of CD4 tests performed and the number of CD4 instruments were analysed to calculate the utilization rate of the available theoretical CD4 capacity. CD4 instruments were considerably underutilized, with only 7.1%, 7.7%, and 13.7% of existing CD4 capacity utilized across 51, 45, and 50 countries reporting data on the number of tests done and the number of CD4 instruments in 2011, 2012, and 2013, respectively (X2 [2; 186,171,250] = 2,062,586, p < 0.001, ES Cramer’s V = 0.07) (median utilization 3.6% [IQR 0.8%–8.0%; MAD 3.1%] in 2011; 5.9% (IQR 2.6%–12.9%; MAD 4.6%) in 2012; and 4.3% [IQR 1.1%–12.1%; MAD 3.8%] in 2013) (Fig 4). The instrument utilization rate was calculated by dividing the total number of CD4 tests performed by the theoretical CD4 capacity (theoretical number of CD4 tests) for all reporting countries. Suboptimal numbers of CD4 tests were performed per patient on ART per year: on average, 1.44 across 44 countries in 2011 (median 1.25; IQR 0.65–2.07; MAD 0.62), 1.32 across 50 countries in 2012 (median 1.70; IQR 1.13–2.47; MAD 0.72); and 1.41 across 46 countries in 2013 (median 1.54; IQR 0.89–2.26; MAD 0.71). In 2013, 14 (30.4%) countries performed <1 CD4 test per PLWHA on ART, which falls short of current recommendations.
The surveys highlighted major issues concerning the functionality and use of CD4 instruments. Thirty-seven (53.6%), 33 (44.6%), and 35 (49.3%) countries reported CD4 conventional and/or POC instruments not in use in 2011, 2012, 2013, respectively (X2 [2; 214] = 1.2, p = 0.558, ES Cramer’s V = 0.05). In 2013, 455 (7.4%) of all CD4 instruments reported by responding countries were not utilized. The percentage of reported instruments that were currently not being used was significantly higher for CD4 POC than for conventional systems in 2011 (18.2% versus 8.3%; X2 [1; 3,785] = 62.3, p < 0.001, ES phi = 0.13) and 2013 (11.0% versus 5.8%; X2 [1; 6,123] = 51.1, p < 0.001, ES phi = 0.09) but lower in 2012 (2.7% versus 5.1%; X2 [1; 4,428] = 10.2, p = 0.001, ES phi = 0.05), although the ES was negligible in 2012 and 2013 (Fig 5). Key reasons for nonutilization included breakdown for CD4 conventional instruments and lack of reagents for POC CD4 instruments (Fig 6). A number of instruments were reported to be in country but not yet installed, possibly because of lack of technical support, training, and/or reagents, as well as shortages of personnel and lack of deployment planning for newly procured instruments.
While nonutilization of CD4 conventional instruments due to breakdown slightly declined from 2011 to 2013 (X2 [2; 10,672] = 11.1, p = 0.004, ES Cramer’s V = 0.02), and nonutilization due to lack of reagents remained mostly unchanged during the same period (X2 [2; 10,672] = 1.9, p = 0.386, ES Cramer’s V = 0.01), the situation for POC CD4 instruments worsened from 2011, with breakdown and lack of reagents being the key reasons reported for nonutilization in 2013 (breakdown, X2 [2; 3,664] = 50.7, p < 0.001, ES Cramer’s V = 0.08; lack of reagents, X2 [2; 3,664] = 41.8, p < 0.001, ES Cramer’s V = 0.07).
Less than 50% of conventional CD4 instruments were covered with maintenance contracts across all 3 survey years; however, there was an increase in the percentage of instruments reported to be under contract in 2012 and 2013 from 2011 (X2 [2; 10,672] = 71.4, p < 0.001, ES Cramer’s V = 0.06) (Fig 7). The proportion of POC CD4 instruments with maintenance contracts dropped significantly from 55% in 2011 to 5% coverage in 2013 (X2 [2; 3,664] = 953.3, p < 0.001, ES Cramer’s V = 0.36). The underlying reasons were not investigated; however, the two potential explanations are previously existing maintenance contracts not renewed after expiry and purchasing of more instruments without maintenance contracts. Servicing levels were also low and declined significantly between 2012 and 2013 (conventional CD4, X2 [1; 7,597] = 174.6, p < 0.001, ES phi = 0.15; POC CD4, X2 [1; 2,954] = 629.3, p < 0.001, ES phi = 0.46): only 26% of conventional CD4 instruments and only 1% of POC CD4 instruments were serviced in 2013 (Fig 7).
WHO recommends, in line with the 90-90-90 targets, to perform VL testing 6 mo and 12 mo after ARV treatment initiation and annually thereafter if the patient is stable on ART and to perform at least two EID tests per HIV-exposed infant, including tests at birth, at 6 wk, confirmatory tests, and tests at 9 mo in some countries [5,6]. Data show that countries continue to expand quantitative/qualitative (VL/EID) nucleic acid testing (NAT) capacity to meet WHO recommendations (Table 3).
Based on our analysis of 38 countries that reported across all survey years on this survey question, NAT capacity increased by 24.4% between 2011 and 2013 (to 9.7 million tests in 2013 from 7.8 million tests in 2011). Abbott m2000, COBAS AmpliPrep/COBAS TaqMan (48 or 96), COBAS AMPLICOR Analyzer, and NucliSENS EasyQ were the most common instruments on the market, with Abbott m2000 and COBAS AmpliPrep/COBAS TaqMan (48 or 96) expanding their market presence between 2011 and 2013 (Table 4). The survey reveals that technologies that can be used at POC for nucleic acid (VL/EID) testing are not yet being used in reporting countries.
VL instrument capacity is sufficient to cover PLWHA currently on ART but not adequate to cover all PLWHA (Fig 8). The theoretical VL capacity per patient on ART per year was on average 2.10 tests in 2011 (median capacity per patient on ART 5.14; IQR 1.27–17.31; MAD 4.16); 2.20 tests in 2012 (median 4.63; IQR 1.45–11.47; MAD 3.83); and 1.23 tests in 2013 (median 3.59; IQR 0.81–9.12; MAD 2.96) (Fig 8), suggesting that current theoretical VL capacity covers the current needs of PLWHA on ART. Yet, country specific analysis of theoretical VL capacity available revealed that in 2013, 28% of reporting countries did not have adequate VL capacity to perform ≥1 VL test per patient on ART (the theoretical need if all HIV-positive individuals were put on treatment, consistent with the latest WHO recommendations). Theoretical VL capacity per PLWHA (including those who are not on ART) was even lower, reaching only 0.52 tests per y in 2011 (median 0.84; IQR 0.27–2.53; MAD 0.71); 0.65 tests per y in 2012 (median 1.28; IQR 0.35–3.50; MAD 0.99); and 0.44 tests in 2013 (median 0.90; IQR 0.30–2.40; MAD 0.74) (Fig 8). Fifty-four percent of reporting countries did not have the capacity to perform ≥1 VL test per PLWHA per y.
In addition, utilization of instruments was low, although increasing, throughout the reporting years. Only 13.7% of theoretical capacity was reportedly being used in 2011 (median 7.5%; IQR 2.6%–23.9%; MAD 6.0%); 16.9% (median 11.5%; IQR 2.2%–32.5%; MAD 9.8%) in 2012; and 36.5% (median 9.4%; IQR 2.3%–28.9%; MAD 8.2%) in 2013 (X2 [2; 29,633,750] = 1,722,577, p < 0.001, ES Cramer’s V = 0.17) (Fig 9). Only 25% of 33 reporting countries performed ≥1 VL test per patient on ART in 2013.
Encouragingly, our data show that across 33, 44, and 51 reporting countries, the number of EID tests performed per infant born to an HIV-positive mother was 1.17 (median 1.10; IQR 1.00–1.91; MAD 0.14) in 2011, 1.43 (median 1.32; IQR 1.01–2.14; MAD 0.32) in 2012, and 1.15 (median 1.20; IQR 1.00–1.94; MAD 0.20) in 2013, suggesting high coverage of EID testing, including confirmatory testing (Fig 10). In 2013, only 3 countries reported less than 1 EID test per infant born to an HIV-positive mother.
As with survey findings for CD4 instrumentation, the surveys highlighted major issues concerning the functionality and use of VL instruments. Twenty-one (35.6%), 22 (34.4%), and 23 (38.3%) countries reported 12%, 14%, and 10% of all VL/EID instruments nonutilized in 2011, 2012, and 2013, respectively (reporting countries, X2 [2; 183] = 0.2, p = 0.896, ES Cramer’s V = 0.02; nonutilized instruments, X2 [2; 2,119] = 7.7, p = 0.022; the difference between 2011 and 2012 and between 2011 and 2013 was not statistically significant, ES Cramer’s V = 0.04). Nonutilization was mostly due to noninstallation and a lack of reagents, with no improvements noted over the 3 survey years (noninstallation, X2 [2; 2,119] = 0.02, p = 0.993, ES Cramer’s V = 0.002; lack of reagents, X2 [2; 2,119] = 5.5, p = 0.065, ES Cramer’s V = 0.04; Fig 11). The number of instruments with maintenance contracts was below 50% across all 3 survey years—increasing from 21% in 2011 to 49% in 2012 and then declining to 38% in 2013 (X2 [2; 2,119] = 105.8, p < 0.001, ES Cramer’s V = 0.16). Servicing levels were also low, with 47% of all instruments reportedly serviced in 2012 and 31% in 2013 (X2 [1; 1,557] = 44.7, p < 0.001, ES phi = 0.17). No instruments were reported to be serviced in 2011, which may be attributed to a lack of corresponding data at the country level.
Survey results across 3 y demonstrate a trend of increasing numbers of CD4 and VL/EID instruments procured in responding countries. Increasing instrument numbers suggest the expansion of services, particularly at the lower levels of the testing network.
Our data show that for CD4, the number of instruments present in countries is sufficient to meet current demand for not only patients on ART but all PLWHA in the countries surveyed. For VL, data show that while responding countries can theoretically meet demand for testing of PLWHA on ART, current instrument capacity is not yet sufficient to cover all PLWHA: over half (54%) of the responding countries did not have the theoretical VL capacity to perform ≥1 VL test per PLWHA in 2013. These data highlight a lack of capacity in many reporting countries to implement new WHO recommendations [2] on VL testing for treatment monitoring and the need for continued support if new goals are to be achieved.
Despite the fact that many countries have instruments in place and the theoretical capacity to respond to testing needs, our analysis indicates widespread underutilization of CD4 and VL technology. Only 13.7% of existing CD4 capacity was utilized in 2013, with 30.4% of countries performing less than 1 CD4 test per patient on ART per y, which falls short of current recommendations. For VL technology, only 36.5% of theoretical capacity was utilized across reporting countries (median 9.4%; IQR 2.3%–28.9%; MAD 9.8%) in 2013. The stark contrast between available instrument capacity and instrument utilization can be partly explained by the substantial number of machines reported as nonutilized each survey year because of breakdowns, stock outs of reagents, lack of installation, or other reasons (e.g., 7.4% of all CD4 instruments and 10% of VL/EID were not in use by the end of 2013). Another contributing factor, not assessed in these surveys, may be inadequate geographic distribution of CD4 and VL instruments, which may be deployed in low-volume sites, resulting in underutilization of technology. The theoretical capacity should help programme managers and funding agencies to plan and deploy instruments based on their capacity and the volume of tests expected to be done in the laboratory facility. Programme managers and funding partners should not expect 100% capacity utilization: nonetheless, efficient use of available equipment and those to be procured should remain a goal for the national programme managers. Of additional concern are data showing extremely low coverage of instruments with maintenance contracts and infrequent or total absence of servicing for in situ instruments (below 50% for all machines reported across responding countries).
There is a need to assess and address the root causes of instrument underutilization. Lack of reagents, uninstalled and underutilized equipment, maintenance requirements, and staff training are issues that national programme managers and policy makers can and must address. Many of the commonly observed equipment challenges can be addressed and prevented from recurring with site selection and deployment planning that includes timely training and retraining, staff proficiency testing, supply chain strengthening, and enforceable maintenance and servicing contracts. Countries could also explore using equipment lease or reagent rental contracts in which the company provides the machine and the country pays for the reagents. With this scheme, diagnostic companies replace or timely repair a nonfunctioning machine and timely supply reagents as agreed in the contract to ensure that the laboratory activities are not interrupted.
Our findings support the need to strengthen diagnostic capacity in reporting countries [17,18]. Recent data from Médecins Sans Frontières have shown financial constraints as a key reason for incomplete or slow implementation of VL testing [19,20]. Many countries still face numerous implementation and funding shortfalls that make it difficult to put the new WHO guidelines into practice [19,20].
Decentralization of HIV care, in some of the hardest hit countries, will remain difficult unless POC technologies meeting WHO prequalification requirements become available and can be deployed for use in peripheral, low-volume treatment centres [21]. Several POC CD4, POC VL, and POC EID technologies are expected to be available in early 2016 [22,23]. However, POC will not solve all the shortfalls raised by our survey, although it may circumvent the need for trained laboratory specialists at the lower levels of the tiered health system. Regardless of the need for POC, it is clear that laboratory-based monitoring will remain a key component of HIV programmes now and in the future [19,20,24]. While we await new POC VL and POC EID technologies, strengthening of the transport systems that facilitate specimen collection from remote ART facilities (e.g., dried blood spot specimens for EID) must remain a focus.
This is the first attempt to comprehensively gather information on HIV testing and monitoring technology in WHO Member States, with the goal to inform programme managers and funding partners in their effort to increase access to HIV monitoring technologies. Caution needs to be taken when using the results of these surveys because of under-reporting in some national programmes and the potential for reported information to be incomplete. Nevertheless, the response rate had increased by 2013, mitigating responder bias and suggesting a welcome increase in interest in HIV diagnostics. Another important limitation to note is that the survey findings were limited to the public sector. In some countries, the private sector makes an important contribution to HIV treatment and care, and associated diagnostic testing and monitoring will not be reflected in these survey findings. The next step will be to provide guidance and institute data quality control procedures to ensure that comprehensive, validated datasets are available on a yearly basis in order to reduce some of the limitations and inconsistencies that arise when comparing large datasets across multiple countries. WHO will continue to evaluate the market share and performance of laboratory technologies through these annual surveys. Finally, future analyses should consider other potential explanations for variability in laboratory capacity, including level of economic development, donor assistance, and prevalence of HIV and other infections requiring similar laboratory diagnostic approaches, such as viral hepatitis.
Despite significant progress in ART scale-up, which has enabled 16 million people to receive treatment, meeting the UNAIDS 90-90-90 targets depends heavily on the commitment and capacity of governments and international partners to improve access to high-quality testing for EID and treatment monitoring. With laboratory systems in reporting countries expanding, a national laboratory strategic plan to strengthen services must be developed, implemented, and monitored by governments and their national and international partners. The focus of international community, to ensure optimal use of laboratory technologies, should be on those countries where interventions for scaling up access to HIV diagnostic technologies are most needed.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.